The Role of Remote Sensing-based in Crop Yield Prediction: A Systematic Literature Review of Approaches, Data Sources, and Challenges

Soka Zimba 1

Aaron Zimba 1

Bob Jere 1

1 School of Computing and Technology with applied science ZCAS University Lusaka Zambia

Soka Zimba^a, Aaron Zimba^a, Bob Jere^a

^a School of Computing and Technology with applied science, ZCAS University, Lusaka, Zambia

ABSTRACT

Crop yields is crucial to food security, agricultural management, and policy planning with the growing climate variability and resource limitations. Remote sensing with machine learning and deep learning has become an effective tool of yield estimation that can be performed at scale and in an objective manner. The current paper reports a systematic literature review of remote-sensing-based crop yield prediction including 106 peer-reviewed articles published in 2015–2025, which is conducted in a PRISMA-compliant manner. The review covers the important methodological strategies, sources of data, types of crops, geographic coverage, and performance measures, challenges, and research trends. Sentinel-2 is the most popular satellite platform with its best balance of spatial resolution, revisit rate, spectral content, and free access which is usually complemented by SAR, Landsat, MODIS, UAVs and ancillary data by multi-modal sensor fusion. In crops like wheat, maize, rice, and soybean, higher order Deep Learning and fusion-based methods are normally associated with coefficients of determination (R² ) between 0.75 and 0.90, which is higher than other single-source and pure statistical methods. Nevertheless, some of these issues have not been fully addressed such as the unavailability of ground truth data, cloud pollution, trade-off in spatial resolution, lack of model transferability and uneven evaluation procedures. The new trends emphasize the increased significance of attention procedures, transfer learning, explainable Artificial Intelligence, data assimilation with crop growth models, and cloud-based systems of operations. Overall, this review offers a systematic review of the existing knowledge, unveils the key gaps, and represents evidence-based recommendations on the direction of future research and functional implementation in the field of precision agriculture and global food security. This review contributes to the literature in that it is a systematic synthesis of methods of modelling, data, and evaluation practices and where research gaps and methodological biases are identified that would influence future remote sensing-based crop yield prediction.

Keywords:

Remote sensing

crop yield prediction

machine learning

deep learning

PRISMA

systematic review

precision agriculture

satellite imagery

sensor fusion

INTRODUCTION

The challenges that global food security is undergoing due to the variability of climate, growth of population and limited resources are unprecedented [1]. Precise and updated crop production forecasts are important to various stakeholders such as farmers to manage their operations, policymakers to plan food security, commodity merchants to predict the market, and international agencies to have early warning measures [2]. Conventional yield estimation techniques are very dependent on ground surveys, crop cutting trial, and statistical reporting system, which are labour intensive, time consuming, expensive, and will in most instances give information at a time when decisions cannot be taken proactively [3]. The remote sensing technologies have become the revolutionary instrument in the agricultural monitoring by providing the synoptic, repeated and objective monitoring of the crop status in various spatial and temporal scales[4]. The ability to monitor on agricultural systems in space has grown enormously since the early Earth observation satellites were launched in the 1970s [5]. The latest satellite constellations are much more spatially resolved (sub-metre to decametre), time-frequency (daily to weekly revisits), and spectral (multispectral to hyperspectral) rich [6]. Complementing satellite platforms, Unmanned Aerial Vehicles (UAVs) can be used to provide an ultra-high-resolution field and sub-field-scale monitoring [7]. Combining remote sensing with new computational algorithms, especially ML and DL, has led to a paradigm shift in the prediction of crop yield [1][2]. These data-driven models are able to fit non-linear and complex relationships between spectral signature, environmental variables, and end yield results. Moreover, the multi-modal sensor fusion plans that implement optical, microwave SAR, thermal, and ancillary data, including weather, soil, and topography have been demonstrated to have a better predictive performance than the unicast schemes [8][9][10]. Even though there has been a milestone, there still exist major knowledge gaps and issues of operation. The diversity of the methods and data used, geographical setting, and evaluation techniques makes the synthesis and comparison of the results of the research more complicated [11]]. Problems like pollution of clouds, trade-offs in spatial resolution, lack of ground truth data, and poor model application across regions are barriers to operational applications [12][13][14].

Related works

2.1

Overview of Remote Sensing in Crop Yield Prediction

2.2.1

Concept of Remote Sensing in Agriculture

Remote sensing (RS) can be defined as the process of obtaining data on an object or a phenomenon remotely, that is, at a distance, most often using sensors on a platform (satellite, aircraft, unmanned aerial vehicle (UAV)). These are electromagnetic radiation sensors that are reflected off or emitted by the vegetation and the soil. Spaceborne satellites such as Landsat and Sentinel series, manned and unmanned aircraft and UAVs [17] are common RS platforms. They vary dramatically in terms of their spatial, spectral, and temporal coverage, such as the Sentinel-2 satellites provide images in multiple spectral bands with 10–30 meter resolutions and revisit of 5–16 days, whereas the UAVs can take high-resolution images on command with the resolutions of centimetres to meters[16]. It is available to the general population to access many medium resolutions (10–100 m) satellite photos at no cost [18]. Crops reflect or emit sunlight, which is sensed by passive sensors (e.g., optical and thermal cameras) and emitted by active sensors (e.g., radar and LiDAR), which reflect the backscatter [19]. Active radar sensors such as Sentinel-1 C-band SAR can work at any weather conditions, which is capable of providing measurements of the day and night providing insights into the structure of crops and moisture level[16]. RS has been critical in the context of modern farming as a tool of precision or digital farming, with hash tag providing high-throughput, non-destructive data on crop health, enabling farmers and other stakeholders to measure and visualize soil and crop health across different stages of growth economically [16]. RS is a useful early warning of nutrient stress and pest problems to implement agricultural measures on time [20].

2.2.2

Remote Sensing Data Sources for Crop Yield Prediction

Multispectral optical satellite data has been significantly used in crop yield models, with the most important sensors being MODIS, Landsat 8 and 9, and Sentinel-2 that have different spatial resolution and revisit frequency [21][22]. MODIS provides essential broad-scale vegetation indices and land-surface temperatures data, whereas Landsat and Sentinel-2 provide finer details (10–30 m) to present the field-level evaluation on a comprehensive level [23]. Satellite radar images, such as Sentinel-1, complement this information by providing information about crop structure and crop moisture, especially in cloudy marriages, due to the decreased atmospheric interference of the microwave signal [23]. Thermal and hyperspectral sensors also play an important role, thermal cameras can measure canopy temperature, which is linked to water stress and pest pressure, and hyperspectral instruments can make a detailed biochemical evaluation of crops [24]. The choice of the data sources must be careful depending on trade-offs between spatial, spectral and temporal resolution [23]. UAVs can be used to deliver high spatial resolution to particular tasks such as seedling detection, but yield estimation is often based on coarser imagery [16][23][25]. New developments in satellite constellations, including Sentinel-2 and Landsat 9 have been able to overcome past shortcomings by providing high revisit frequencies with moderate resolution potentials [23].

2.2.3

Vegetation Indices and Biophysical Parameters

The vegetation indices (VIs) are critical instruments that convert the crude reflectances of remote sensing into the indices of the vegetation properties [26]. The most common is the Normalized Difference Vegetation Index (NDVI), which makes use of red and near-infrared (NIR) bands to measure canopy greenness and relate to leaf area and biomass [27]. Others of significant indexes are the Enhanced Vegetation Index (EVI), corrects on saturation and soil background, the Soil-Adjusted Vegetation Index (SAVI), and the Green NDVI (GNDVI) which increases sensitivity of chlorophyll and relies on the green band rather than the red [28][25]. The indices are essential in the modeling of yield; researchers, such as those of [25], note that NDVI and GNDVI are popular in predicting crop yields, with GNDVI being better than NDVI in assessing the yield of maize [16][29][30]. Choosing a good index is very crucial to the accuracy of the model. Such biophysical parameters which can be obtained with the help of remote sensing data are leaf area index (LAI), fraction of absorbed photosynthetically active radiation (FAPAR), canopy cover, and biomass [31]. The estimation of these parameters can be done using VIs or retrieval algorithms and it has been found that LAI works especially well in yield models using a range of regression techniques. Combination of LAI and many VIs has been demonstrated to improve the accuracy of estimates of biomass. All in all, VIs and related biophysical variables are important proxies to measure crop health and productivity and are the core of numerous yield prediction models [23][32].

Limits of VIs however include saturation under high vegetation density and reduces their sensitivity to changes in biomass, the influence of atmospheric conditions, such as clouds and aerosols, on optical measurements. As researchers, it is stated that these aspects restrict the usefulness of RS-based estimates [16]. Multi-index techniques can be used to increase the dynamic range to overcome the problem of saturation, and atmospheric corrections and combination with active sensors such as radar can be used to fill weather-induced data gaps, as was the case with the useful Sentinel-1 SAR data application in cloudy regions[16].

2.2

Machine Learning Principles in Crop Yield Prediction

2.2.1

Role of Machine Learning in Yield Estimation

ML is a major improvement in crop yield estimation relative to his traditional models that were based on linear models or empirical regression of only few factors[33][34]. ML methods are able to identify complex, non-linear relationships directly out of large datasets, such as RS and weather data[35]. According to [36][37], ML has acquired popularity in large scale yield prediction and phenotyping because there is a lot of sensor data[38][39]. These models perform well in consuming multi-source data that is complex and combining the RS data with climatic, soil and management data, improving the yield predictions. As an example, one of the studies showed better corn yield predictions in the USA through the Landsat-based indices coupled with climate variables[39]. The flexibility in ML models provides the opportunity of modeling without specified functional form and allowing some complex interaction among crops and their environment[23]. The high quality of advanced ML models including ensemble trees and neural networks extends the capabilities of linear models, and thus, leads to higher accuracy in the diverse environmental settings[40]. However, the models need careful tuning to avoid overfitting, and it can be controlled with the help of cross-validation and regularization when the models are used in data-rich agricultural tasks.

2.2.2

Common Machine Learning Algorithms

There are a number of common ML algorithms that are used in prediction of crop yield[41][42]. Linear and multiple regression are models that are based on the relationship between yield and predictors such as NDVI and weather variables; these models are useful in their interpretability, but not much can be done concerning non-linear effects[41]. The Support Vector Machines (SVM) is the most suitable against non-linear patterns and most of the data dimensionality is moderate[43]. According to[42][44], the ensemble technique, Random Forest (RF), is considered to be robust against overfitting and flexible with mixed data, and this makes it one of the most used ML algorithms in yield estimation. XGBoost and LightGBM are also the gradient boosting machines that are characterized by their iterative nature in prediction accuracy improvement and also are significantly popular in terms of popularity. The k-Nearest Neighbors (k-NN) algorithm is a prediction based on a weighted average of the nearest samples, but it may be misunderstanding to the parameter used and data scaling[42]. Studies have shown that the most common ML options in modern research include linear regression, Random Forest and gradient boosting, whereas SVM and k-NN do not receive much attention as compared to ensemble-based applications in addressing complex problems[45]. The next part deals with deep learning models including CNNs and LSTMs.

2.2.3

Feature Engineering and Data Preprocessing

The construction of ML models to predict yields entails carefully obtaining features of the raw data, which are divided into remote sensing features, environmental and soil features, and normalization and dimensionality reduction techniques[46]. The Remote sensing characteristics are based on the satellite data, especially the vegetation indices such as NDVI, GNDVI and EVI which are derived using certain satellite bands. Time-series imagery is frequently used by researchers to determine the important predictors of yield (based on time). As an example, [47] focused on the time when NDVI reaches its highest value, often known as the peak of season, and accumulation of greeness, measured by the area of the NDVI curve, which are the key predictors of maize and soybean yield models[47].

When studying the variability of yield outside the remote sensing, it is important to include environmental variables which includes meteorological variables, rainfall, temperature, sunlight, and soil properties such as texture and moisture[48]. Other studies focus on the relevance of integrating climate indices and cumulative precipitation with remote sensing at the expense of improving ML inputs[46]. Normalization and dimensionality reduction are required prior to modeling in order to make the features comparable, especially with high dimensional dataset which contain hyperspectral bands[49]. Principal Component Analysis (PCA) or variable importance filtering are some of the techniques that are used to narrow down input variables, which may reduce the number of vegetation indices utilized in optimizing the model performance[23].

The problem of missing or noisy data has been the key to developing trustworthy ML models, especially in the case of satellite time-series data that is vulnerable to atmospheric interferences or problems with sensors[50][51]. Some of the common techniques used are temporal interpolation and gap-filling with nearby clear-sky observations or Synthetic Aperture Radar (SAR) data[52][53]. Also, in case of ground-truth yields, noise can be minimized by smoothing, whereas in the absence of such, regularization techniques or imputation can be applied. Proper feature engineering entails the usage of various types of data like the research proved, which highlights the significance of combinations of features in models to improve the models and recommend against undue redundancy of features[54], [55]. Post-hoc testing allows the researcher to rank vegetation indices in comparison to yields and this approach is used to identify important variables in predicting yields[23].

2.2.4

Model Training and Evaluation Metrics

The training of machine learning models is done on historical data, which is further subdivided into training, validation, and testing subsets[41][56]. Strategies that are frequently used in data partitioning are k-fold cross-validation and hold-out validation with 70–80 percent of the data training and the remaining testing[57]. Cross-validation, e.g. 5-fold or leave-one-year-out, can be used to determine the stability of the model and variate hyper-parameters, including tree depth and learning rate[58]. Standard regression measures are used to assess model performance. Root Mean Square error (RMSE) of the prediction error magnitude, Mean absolute error (MAE) of the average absolute error and Coefficient of determination (R2) which shows the proportion of variance which the model accounts for, and Mean absolute percentage error (MAPE) which reveals the relative errors are considered to be the key measures used in the regression analysis evaluation [59]. According to a study by Shawon et al., RMSE, R2 and MAE were the most popular metrics that were used in the literature on the topic, as each provided a unique assessment of the degree of accuracy. R2 is used to define overall fit, whereas RMSE and MAE are used to give information about the magnitude of actual errors[60]. Also, the metrics used in agriculture, such as bias or the share of correctly predicted fields, are used as well[45]. Generally, results of cross-validation show average RMSE between folds, which is an indication of the generalizability of the model, and has also been applied to deep learning solar racking[61]. Various measures were employed in most of the research, and the equations are presented below:

2.2.4.1

Root Mean Square Error (RMSE) is a measure of the average magnitude of the prediction error between the observed values (y_i) and the predicted values (

$\:{\widehat{y}}_{i}$

). It represents the square root of the mean of the squared differences between predictions and actual observations.

$\:RMSE\:=\:\sqrt{\frac{1}{n}}\sum\:_{i=1}^{n}({y}_{i}-{\widehat{y}}_{i}{)}^{2}\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:(1)$

2.2.4.2

Mean Square Error (MSE) is average squared difference between the observed and predicted values. MSE is always non-negative, with smaller values indicating better model performance because the errors are squared; MSE penalizes large prediction errors more heavily than small ones. The unit of MSE is the square of the target variable’s unit, which can make interpretation less intuitive compared to RMSE.

$\:MSE\:=\:\frac{1}{n}\sum\:_{i=1}^{n}({y}_{i}-{\widehat{y}}_{i}{)}^{2}\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:(2)$

2.2.4.3

The coefficient of determination (

$\:{R}^{2}$

) is the measure of the proportion of variance in the observed data that is explained by the model.

$\:{y}_{i}$

is the observed values while

$\:{\widehat{y}}_{i}$

is the predicted values.

$\:\stackrel{-}{{y}_{i}\:}$

mean of observed values. The key properties are that

$\:{R}^{2}=1$

the perfect prediction.

$\:{R}^{2}=0,$

this means model performs no better than predicting the mean.

$\:{R}^{2}<0\:$

is when the model performs worse than the mean predictor. In crop yield modeling, a higher

$\:{R}^{2}$

indicates that the model effectively captures the variability in yield caused by environmental and management factors.

$\:{R}^{2}=1-\frac{{\sum\:}_{i=1}^{N}({y}_{i}-{\widehat{y}}_{i}{)}^{2}}{{\sum\:}_{i=1}^{N}({y}_{i}-\stackrel{-}{{y}_{i}}{)}^{2}}\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\left(3\right)$

2.2.4.4

Mean Absolute Error (MAE) is the measure of the average magnitude of the prediction errors, without considering their direction.

$\:{y}_{i}$

= actual (observed) value while

$\:{y}_{i}$

= predicted value, and n = number of observations.

$\:\text{M}\text{A}\text{E}\:=\:\frac{1}{n}\sum\:_{i=1}^{n}|{y}_{i}-\widehat{{y}_{i}}\left|\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\right(4)$

2.2.4.5

Normalized Root Mean Square Error (NRMSE) an expression of the RMSE relative to the range of the observed data, making it dimensionless.

$\:{y}_{\text{m}\text{a}\text{x}}$

$\:{y}_{\text{m}\text{i}\text{n}}$

= maximum and minimum observed values. This allows comparison of model performance across different datasets or regions, even if they have different scales. This allows comparison of model performance across different datasets or regions, even if they have different scales.

$\:\text{NRMSE}=\frac{\text{RMSE}}{{y}_{\text{m}\text{a}\text{x}}-{y}_{\text{m}\text{i}\text{n}}}\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\left(5\right)$

2.2.4.6

Mean Percentage Error (MPE) is for evaluating the average percentage deviation of predicted values from observed values. B_i = observed value, E_i = estimated (predicted) value,

$\:\stackrel{\prime }{B}$

= mean of observed values. It describes the behavior of the model concerning the observed values, indicating that positive values suggest an underestimation of these values, whereas negative values imply an overestimation.

$\:\text{MPE}=\frac{1}{n}\sum\:_{i=1}^{n}\left(\frac{{B}_{i}-{E}_{i}}{\stackrel{\prime }{B}}\right)\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\left(6\right)$

2.2.4.7

MAPE is an expression of the average prediction error as a percentage of the actual values. It provides an intuitive measure of error which allows easy comparison between models, for instance, MAPE = 10%, predictions deviate by 10% on average.

$\:\text{MAPE}=\frac{1}{n}\sum\:_{i=1}^{n}\mid\:\frac{{B}_{i}-{E}_{i}}{{B}_{i}}\mid\:\times\:100\:\:\:\:\:\:\:\:\:\left(7\right)$

2.3

Deep Learning Principles in Crop Yield Prediction

2.3.1 Introduction to Deep Learning in Remote Sensing

DL uses the neural network architectures that have more than one layer that is intended to learn hierarchical features in data by default, as opposed to the more traditional ML, which frequently requires manual feature engineering[62]. The main benefits of DL to yield prediction are: feature learning, which is automatic, through convolutional and recurrent networks which learn complicated spatial or temporal structures without accounting for any pre-computed indices; scalability to large data volumes, so that large volumes of satellite images and weather data can be analysed using convolutional neural networks (CNNs); and predicting complex data types, so that convolutional neural networks (CNNs) can be used to process images, and recurrent neural networks (RNNs) or long short-term memory networks (LSTMs) can be used to process time The potential of DL in prediction based yield is its capacity to capture multispectral lines of time-series and spatial contexts[64]. Nevertheless, it usually requires more data and computing power and is not interpretable. Although these issues exist, a variety of studies show improved performance in yield predictions with the use of DL approaches, whereby CNNs on UAV images have shown to greatly minimize uncertainty in the predictions when compared to the more basic models[63].

2.3.2 Convolutional Neural Networks (CNNs)

CNNs are the most efficient to recognize objects and can even interpret deteriorated images due to noise. Therefore, there have been applied to crop yield work to acquire spatial features such as texture and canopy structure on input patches of images[64]. Weight sharing and pooling give them the ability to generalize in various fields. Patch-based learning is a practical method, which splits a bigger satellite image into smaller patches, and a CNN predicts local yields. This method supplements small datasets by adding the number of samples[62][65]. CNNs have been effectively used to produce yield mapping and crop classification and it has been found out that when regularization is done appropriately and depth is considered, CNNs can learn the features relevant to yield when presented with an image[63].

2.3.3 Recurrent Neural Networks (RNNs) and LSTM Models

Recurrent Neural Networks (RNNs) and Long Short-Term Memory (LSTM) models have been gaining popularity in crop yield prediction recently as they are capable of capturing time-series-based data[66][67]. They record temporal responses, which have impacts on yields, e.g. how early season water stress affects end of season yields. The researchers note that LSTMs are important in the modeling of relationships that have a very long time difference and that this is essential in the analysis of multi-date imagery or long-term climate conditions[39]. A number of studies have managed to apply LSTMs to phenology-based predictions and showed better predictions of yields when time-series data were combined across multi-years. BiLSTMs have demonstrated superior performance to traditional RNNs in yield prediction of rice[25].

2.3.4 Hybrid Deep Learning Models

In hybrid deep learning models, the authors combine Convolutional Neural Networks (CNNs) with RNNs to use the advantages of both spatial and temporal feature detection. The CNN-LSTM model takes the images and removes the spatial features before sending the features through time using the LSTMs[68]. This strategy has shown better performance on predicting yields as was shown by [25], who discovered that CNN-LSTM fusion performed better than single convolutional methods in predicting the yield of wheat. Moreover, the combination of remote sensing information with other auxiliary data (canopy surface temperature, water stress indices etc.) has been useful. The research observed that when optical and thermal data are combined in a CNN-LSTM model to predict the yield of corn, prediction accuracy is enhanced relative to the situation when vegetative indices are used to do the same[69][25]. In general, hybrid models based on the combination of different sources of data and analysis methods are always more effective than the single-source methodologies used in traditional forecasts of yields due to the use of complementary information between different sources.

2.3.5 Data fusion and MultiSource integration

Information integration and data fusion increase the accuracy of yield prediction by exploiting multiple forms of data, such as remote sensing (RS) data, available via optical data and synthetic aperture radar (SAR) systems[70][71]. An outstanding one is a combination of Sentinel-2 NDVI and Sentinel-1 SAR backscatter; thus, the models are able to surmount cloud cover and efficiently obtain an assessment of the level of chlorophyll[72]s[73][74]. Multi-temporal fusion, or image stacking, can be used to augment revisit rates and overcome data gaps in the image by stacking both images of different satellites[23]. Further clarification of the variability in yield through the combination of climate and soil data with RS characteristics can also be given, such as the ground-based temperature data may be supplemented by the MODIS land surface temperature to capture the impacts of heat stress.

In practice, fused datasets are incorporated into ML and DL models and in this way, the models are able to learn the associations between various data modalities of yield. This may include the establishment of one feature, derived by means of optical vegetation indices, SAR data, and weather variables[16]. Although there are some advanced methodologies, such as the application of attention mechanisms, lots of the studies do not come to a conclusion that complicated feature stacking is the only way to have a good performance of the model. Recent reviews attest to the fact that multi-source fusion is likely to be beneficial; data on multimodal RS produces more accurate and detailed analyses of crop conditions, which ultimately generates improved yield estimates[75][76][77]. The error rate of built in models that consider optical, radar and climate data is generally seen to be lower than their single-source counterparts[78].

Widely used data fusion strategies are feature-level fusion, which involves the combination of features of all the data sources before modeling them, and decision-level fusion, in which the results of different models are fused[71]. It has been discovered by multiple researchers that feature stacking is already sufficient to achieve good performance, in particular, when the learning algorithms are powerful. One of the strongest points of this strategy is that it helps to solve the personal weaknesses of each information source; the low sensitivity of SAR to the chlorophyll can be covered by the abilities of optical data, enhancing the overall predictability of harvests[79].

Multi-source data is added to the model, which enhances the quantity of information available. As an example, radar data can give information about the soil moisture and structure, which NDVI cannot detect, thermal data can be used to indicate the crop stress, and climate records can give a background of rainfall-irrigation effects. Experiments have always shown that models, which make use of fused data, perform better than models, which use single data types[23][16]. Altogether, multi-source data fusion currently becomes a popular strategy in yield prediction, caused by the presence of large open RS and climate data.

3 Methodology

It is a systematic literature review that follows the PRISMA guidelines in investigating studies in which remote sensing data is used to predict crop yields[80]. It involves a wide range of techniques, such as classical statistical and sophisticated deep learning on a wide range of crops and geographic locations. The synthesis of the review offers the main points of understanding methodological diversity and performance trade-offs. It also provides evidence-based recommendations to the stakeholders and also provides the research gaps that are important in subsequent research in remote-sensing yield prediction.

4 Research Objectives

The specific objectives of this systematic review are to:

To systematically identify, screen, and synthesize peer-reviewed literature on remote sensing for crop yield prediction

Categorize and characterize methodological approaches, including statistical models, machine learning algorithms, deep learning architectures, and hybrid methods

Document the range of remote sensing data sources, sensors, spatial/temporal resolutions, and derived features

To analyse reported performance metrics and accuracy benchmarks across different methods, scales, and crops

To identify persistent challenges and limitations that constrain operational implementation

To highlight emerging innovations and recommend priority areas for future research

5 Research Questions

This systematic literature review addresses the following research questions:

What are the primary methodological approaches employed for remote sensing-based crop yield prediction?

Which remote sensing data sources, sensors, and spectral features are most utilised?

What crop types and geographics regions have been studied?

What performance metrics are reported, and what accuracy ranges are achieved?

What are the key challenges, limitations, and knowledge gaps in current research?

What emerging trends and future research directions can be identified?

Fig. 1

systematic literature review selection flow process[107].

5.1 Eligibility Criteria

Table 1

Selection Criteria
S/N	Inclusion	Exclusion
1	Empirical research articles, systematic reviews, or meta-analyses	Did not utilize remote sensing data
2	Utilized satellite, airborne, or UAV-based remote sensing data as primary or supplementary input	Focused exclusively on crop classification, disease detection, or other non-yield outcomes
3	Focused on crop yield prediction, estimation, or forecasting, and other deep learning metrics	Lacked quantitative performance evaluation
4	Reported quantitative performance metrics such as ( R², RMSE, MAE)	Opinion pieces, editorials, or non-peer-reviewed grey literature without methodological rigor
5	Peer-reviewed journal articles, conference proceedings, or preprints
6	Published in English
7	Date restrictions 2015 to 2025 applied and two articles added to capture historical development and recent innovations	Articles older than 2014 not included except historical articles.

5.2 Information Sources and Search Strategy

2.3.1 Databases Searched

A comprehensive search was conducted across five major academic databases on November 12, 2025: Google Scholar, IEEE Xplore, ScienceDirect, Elsevier and PubMed.

2.3.2 Search Terms and Queries

Search strategies were conducted by utilising Boolean Operators query to each database’s capabilities:

Boolean query: (remote sensing OR satellite imagery OR earth observation) AND (crop yield prediction OR agricultural forecasting OR yield estimation) AND (machine learning OR deep learning OR modelling approaches)

Boolean query: (remote sensing [Title/Abstract] OR satellite imagery[Title/Abstract]) AND (crop yield[Title/Abstract] OR agricultural production[Title/Abstract] OR yield prediction[Title/Abstract]) AND (machine learning[Title/Abstract] OR deep learning[Title/Abstract] OR modelling[Title/Abstract])

2.3.3 Search Results

The search yielded a total of 18,314 initial records from all the databases recorded. After screening and inclusion, we were left with 106 articles from different databases, as shown in Fig. 2 below.

Fig. 2

Distribution of the articles as per the database.

An analysis of the reference sources shows that MDPI journals constitute the largest share of the reviewed literature (31 articles), followed by IEEE publications (26 articles) and Elsevier journals (24 articles). A clear increase in MDPI and IEEE publications is observed after 2020, reflecting the growing interest in machine learning and deep learning applications for crop yield prediction using remote sensing data.

5.3 Study Selection Process

4.3.1 Deduplication

All the records retrieved were successfully exported and merged into one and unified database. Thereafter, an automated deduplication process was run, using criteria like DOI (Digital Object Identifier) and matching on titles and authors. This careful procedure led to a streamlined dataset available as 106 unique papers.

4.3.2 Screening and Eligibility Assessment

Following the PRISMA guidelines, the study selection process was conducted in three main stages. First, selection of the studies followed the PRISMA guidelines and was based on three major steps. To select eligible participants, initially, a title and abstract screening of the 106 unique records that were obtained was conducted to identify the eligibility according to predefined criteria. Then, the records which successfully passed this first screening were further reviewed on a full-text basis to complete their inclusion in the study. Lastly, the included papers were re-ranked.

4.3.3 Final Inclusion

Following the screening and eligibility screening, 106 articles were then left to undergo qualitative synthesis and analysis.

5.4 Data Extraction and Synthesis

The information about the reviewed articles has been extracted in the following six critical key areas: Study Characteristics that include the authors, year of publication, and journal, geographic region and crop types, Remote Sensing Data that involve sensor platforms, spatial and temporal resolution, spectral bands and derived indices, Methodological Approaches, which has the description of the variety of statistical models, ML algorithms, DL architectures, and hybrid approaches, Performance Metrics, which reported the values of R², RMSE, MAE, symmetric mean absolute percentage error (sMAPE) among other accuracy measures, Challenges and Limitations, The great heterogeneity of studies, their design, the scales of implementation, types of crops and approach to methodologies applied resulted in the fact that a narrative synthesis approach was considered more suitable than a quantitative meta-analysis, and more diverse data sources could be interpreted.

5.5 Quality Assessment

Such criteria as the clarity of methodological description, the validity of used validation strategies, including cross-validation and the utilisation of independent test sets, clarity of performance metrics reporting, and a critical discussion of limitations and generalisability of the results are considered to be the quality assessment criteria.

6 RESULTS

5.1 What are the primary methodological approaches employed for remote sensing-based crop yield prediction?

The systematic search was done to find 106 unique studies published between 2007 and 2025, and the data indicates that publications have increased significantly since 2022 as demonstrated in Figs. 4 and 5. This can be explained by the accessibility of Sentinel-2 data and by development of deep learning frameworks. These studies included different geographic areas, crops and scales between plot and on a global scale and also employed different methodological options. The article of 2007 was included in one as it gives background and forms of vegetation indices that is regularly used in predicting crops. The reason is that it gave specific data regarding the vegetation index. The 2014 article was chosen as it has been significant in terms of learning supervised and unsupervised by the ML methods that provide the basis of ML and DL in the modern world. Over the last several years, the methodology of crop prediction has evolved and methods of leveraging ML and DL timeline have developed, as illustrated by the Fig. 3 below:

2000–2010: Statistical Era

├── Linear regression + NDVI

├── Multiple regression + weather

└── Empirical crop models

20210 − 2015: Machine Learning Emergence

├──Random Forest adoption

├── SVM for yield prediction

├── Feature engineering focus

└── Landsat + MODIS primary data

2015–2018: Deep Learning Introduction

├── Early CNN applications

├── LSTM for time series

├── Sentinel-2 Launch(2015)

└── Transfer learning exploration

2018–2022: Advanced Architectures

├── Attention mechanisms

├── Multi-modal fusion

├── Hybrid physical-statistical

├── SAR-optical integration

└──Gaussian Process spatial models

2022–2025: Operational Maturity

├── Transformer adaptations

├── Foundation models exploration

├── Operational systems deployment

├── Standardization efforts

├── Cross-regional transfer learning

└── Explainable AI focus

Fig. 3

Methodological Evolution Timeline

Figure 4: The articles' distribution utilised frequency in years

5.2 Which remote sensing data sources, sensors, and spectral features are most commonly utilized?

The reviewed literature identified six major methodological approaches, often combined in hybrid frameworks to leverage complementary strengths.

5.2.1 Statistical and Empirical Models

The classical statistical techniques, such as linear regression, multiple regression, and piecewise empirical equations, have been popularly employed as benchmark techniques and applicable to large-scale applications[1][81]. Such models usually correlate VIs or biophysical parameters and produce results. Although computationally efficient and interpretable, they tend to have a low ability to describe nonlinear relationships and interactions among a number of variables. The usefulness of remote sensing in empirical modelling has been proven in the fact that maize yields can be successfully predicted by empirical models based on remote sensing, particularly Leaf Area Index (LAI) and Normalised Difference Vegetation Index (NDVI) data [82]. Besides, the use of remote sensing data has resulted in significant enhancement of statistical models that are traditionally characterized by a time-intensive nature, thus yielding reliable and efficient predictions at a regional scale[83]. [81]has shown the application of NDVI and meteorological parameters in wheat yield forecasting in India through the use of regression models, which have moderate accuracy in regional forecasting.

5.2.2 Machine Learning Algorithms

The reason is that machine learning methods are now the norm of the day owing to their capabilities to model nonlinear relationships and also work with high-dimensional feature spaces[2]. The most commonly used algorithms are:

RF is a tree-based algorithm that is well known because of its robustness and ability to rank the features as highly important, in addition to the fact that it is able to withstand overfitting, thereby becoming among the widely used algorithms in a variety of applications[2][3][84]. SVR is a method based on a kernel, and it is specifically effective when working with small to medium-sized datasets[3]. SVM, RF, and Extremely Randomized Trees (ERT) are some of the methods that have been demonstrated to be effective in predicting crop yields when applied to remote sensing data, as noted by [85]. Two other significant sequential ensemble techniques used are Gradient Boosting and XGBoost, which improve predictive accuracy during iteration, which, in most cases, leads to higher accuracy[2][3]. Also, Gaussian Processes are a type of probabilistic model that, in addition to making predictions, also generate is uncertainty, giving a more in-depth picture of predictive confidence[86]. Nevertheless, [3] compared several ML algorithms to estimate wheat yields based on the analysis of Sentinel-2 and hyperspectral data and identified that LSTM was the best-performing algorithm with RMSE = 0.201 t/ha against 0.260–0.489 t/ha of the other models.

5.2.3 Deep Learning Architectures

DL has become a revolutionary methodology, especially to process high-dimensional spatial-temporal remote sensing data [23][4][6]. CNNs are used to derive spatial information of imagery, which is an efficient way of extracting local patterns and textures[4][6]. Conversely, RNNs and especially LSTM and Bidirectional Gated Recurrent Unit (BiGRU) can think about time-based dynamics and phenological patterns based on time-series data[3][4][6]. The CNNs and LSTM networks are suitable to perform the spatiotemporal analysis of crop growth and boost prediction results[87][88]. Attention Mechanisms are added to the model to allow it to dynamically weight relevant spatial or temporal features, thereby making it easier to interpret and increasing its overall performance [6][89]. The non-stationarity of spatial data has also been significantly enhanced through the integration of remote sensing data with ensemble learning methods such as Geographically Weighted Random Forest Regression (GWRFR) to provide better accuracy in the prediction[90]. Multi-Branch or Dual-Stream Networks are also intended to provide separate processing of data modalities or T scales in separate pathways before the two are combined[4][[91][10]. An applicable case of these ideas can be exemplified by the work of [4], who built a dual-stream deep learning system to predict winter wheat yield in China on a county level. Their model modelled a coefficient

of determination (R2) of 0.79 and an RMSE of about 650.21 kg/ha when the sequential and the static features were processed individually. Also, a different deep attention model was developed to predict the yield of maize in Northeast China, and achieved an R² of approximately 0.896 and an RMSE of approximately 908 kg/ha. This shows how the attention mechanisms can be useful in emphasizing critical growth periods when making the predictions of the yield[6]. The comparison between the techniques of ML and DL used is presented in Table 2 below.

Table 2

Comparison of Machine Learning and Deep Learning Methods
Method Category	Algorithms	References	Advantages	Disadvantages	Typical R² Range	Best Use Cases
Statistical Models	Linear regression, MLR, PLSR	[92]	Simple, interpretable, low data requirements	Limited nonlinearity capture	0.50–0.70	Baseline, regional scale, limited data
Tree-Based ML	Random Forest, XGBoost, GBDT	[2][3][84]	Robust, feature importance, handles nonlinearity	Can overfit, limited spatial modelling	0.60–0.80	Tabular features, operational systems
Kernel ML	SVM, SVR, Gaussian Process	[3]	Strong theoretical foundation, uncertainty quantification	Computational cost, hyperparameter sensitivity	0.60–0.75	Small-medium datasets, uncertainty needed
Shallow Neural Networks	MLP, feedforward networks	[93]	Flexible, universal approximation	Requires careful tuning, prone to overfitting	0.65–0.80	Moderate complexity problems
CNNs	ResNet, VGG, U-Net	[4][6]	Spatial feature extraction, translation invariance	Requires large data, computationally intensive	0.70–0.85	Image-based inputs, spatial patterns
RNNs/LSTMs	LSTM, GRU, BiLSTM	[3][4][6][25] [66][67]	Temporal modelling, sequence learning	Gradient issues, sequential processing	0.70–0.85	Time-series VIs, phenology modelling
Attention Models	Transformers, self-attention	[6][89]	Dynamic weighting, interpretability	High computational cost, data hungry	0.75–0.90	Multi-temporal, multi-modal fusion
Hybrid DL	CNN-LSTM, dual-stream	[87][88]	Spatial-temporal learning, multi-scale	Complex architecture, difficult training	0.75–0.90	Multi-modal, spatial-temporal data

NOTE: Represented performance ranges, which have been synthesised after several different studies, depend on the crop type, the level of space, sensor resolution, and availability of data. The greater accuracies tend to be observed at the aggregated administrative scales and field-scale applications tend to have more variability.

Table 2 synthesises an evident evolution of methods of crop-yield prediction showing that the interpreting, yet constrained statistical models give way to more expressive ML and DL algorithms. The ML models represented as trees provide a viable trade-off between stability and computational efficiency, and DL and hybrid designs are always expected to be more accurate at prediction because they can represent more complex nonlinear and spatial-temporal relationships on which crop growth processes rely on. Nevertheless, these performance improvements are accompanied by higher data needs, computational, and interpretability, which limit their real-world use especially in areas that lack data. The comparative trends provided in Table 2 thus indicate not only the difference in methodological performances but also the trade-offs that are significant between the accuracy, scalability and applicability.

5.2.4 Hybrid Physical-Statistical Models

Hybrid approaches integrate data-driven methods with process-based crop growth models through Data Assimilation (DA) or model coupling [11]. These methods integrate mechanistic insights of crop physiology with the pattern recognition abilities of ML and DL, enhancing both generalization and interpretability. They involve incorporating physical constraints and agronomic knowledge, which leads to improved extrapolation beyond the conditions present in the training data and enhances interpretability for effective stakeholder communication. Reviews by [11] emphasize data assimilation frameworks that combine remote sensing observations with crop models like WOFOST and DSSAT, identifying them as promising strategies for scalable and interpretable agricultural forecasts[11]. However, challenges related to data quality, model generalizability, and interpretability persist, particularly due to the "black box" nature of deep learning models. Addressing these challenges, including incorporating diverse data sources and transfer learning techniques, can improve model robustness across various regions and cropping systems[87][1].

Fig. 5

Cross-Tabulation of Evaluation Metrics vs Model Type

The Counts are approximations because many studies report multiple metrics, and only studies explicitly reporting quantitative yield prediction performance were included.

5.2.5 Representation and Transfer Learning

In order to solve the problem of a lack of ground truth data, as well as to increase cross-regional transferability when using remote sensing applications, the different advanced representation learning methods have been considered. They belong to the group of techniques known as Autoencoders and Compositional Autoencoders, which are aimed at producing compressed and informative representations of high-dimensional data sets acquired by means of remote sensing [12][13]. Such a method assists in dealing with big data successfully by deriving valuable characteristics[14].

The other significant technique that is used is Transfer Learning, which involves pretraining models on areas where there is plenty of data and then fine-tuning them to use them on target areas that do not have such data [13]. Such a plan gives the opportunity to make the best use of available resources and enhances the performance of models in low-data settings.

Besides, Domain Adaptation is used to separate the effects of such factors as genotype, environment, and management practices[12]. In this manner, it increases the generalization capacities of the models and thus makes them more stable in various contexts. One of the most striking examples that can demonstrate how these techniques are used in practice is the research that [14] conducted. This study shows that deep transfer learning can be an effective way of using remote sensing data to forecast agricultural yields in regions where there is a lack of ground truth data. With these techniques, dependence on labelled data to train the model is greatly minimized, hence making it easier to utilize data in predicting yield.

5.2.6 Multi-Modal Sensor Fusion

Sensor fusion approaches can be used to augment data sources that come from different sources to overcome the sensor limitations of an individual sensor [8][94][10]. Important ones are Optical + SAR Fusion, which combines spectral data with microwave backscatter to overcome cloud cover and enhance sensitivity to structure and moisture; Multi-Resolution Fusion, which combines frequent coarse-scale data of MODIS with fine-scale data of Sentinel-2 and Landsat; and Multi-Modal Feature Fusion, which integrates remote sensing data with weather, soil, topography, and management information[8][94][95]. The effectiveness of these techniques has been demonstrated by researchers, and [8] has reported the high results in the improvement of estimation of yield under cloudy conditions owing to optical-Synthetic Aperture Radar (SAR) fusion. Also, [8] proposed an adaptive multi-view fusion-based approach to predict yield of sub-fields, where the R2 of the predictor values on the sub-field level ranges around 0.68, and the predictor values of field averages in multi-country tests around 0.80[94].

5.2.7 Remote Sensing Data Sources and Sensors

The studies reviewed had a wide range of remote sensing platforms that had different features in terms of spatial resolution, temporal frequency, spectral coverage, and cost.

5.2.7.1 Satellite Platforms

Sentinel-2 is operated by the European Space Agency and has spatial resolution of 10–20 meters with a 5-day revisit rate that has 13 spectral bands[3][91]. Its free availability, high resolution, and ability to visit the same location frequently allow it to be commonly used in field to regional research, especially in multi-temporal vegetation index research, crop and plant phenology research, as well as sub-field variation mapping. In contrast, the Landsat NASA/USGS program provides 16-day revisit cycle with a 30-meter resolution, which provides a helpful long history beginning in the 1970s[88][96]. This enables trend analysis and long term climate yield relationships and thus is appropriate in the regional scale of to national scale yield forecasting. The MODIS system is characterized by coarse resolution of 250–1000 meters and revising the system on a daily basis successfully supports large-scale yield forecasting and early warning systems. The SAR such as the Sentinel-1 and RADARSAT satellites have microwave sensors that go through clouds to supplement optical data, particularly during overcasts[88][96].Multispectral and hyperspectral satellite images are useful in measuring crop health and yield estimation. All these images record different spectral bands that could be utilized to calculate vegetation indices such as NDVI and EVI that are essential in measuring plant health, as well as projecting yield[97][98]. Based on these systems, the structure of crops and the estimation of soil moisture are effective and provide predictions at all weather conditions. Commercial high-resolution satellites, including PlanetScope, RapidEye, and WorldView, on the other hand, offer sub-meter to 3-meter resolutions with daily capability to revisit a field or sub-field scales with revisit rates up to once per week[96]. They are good at mapping within-field variability and are also good at predicting yields. An experiment by [99] showed that PlanetScope at 3-meter resolution found about 15% more within-field variability than Sentinel-2 at 10 meters to predict corn yield, which is what higher resolution can bring to precision agriculture.

5.2.7.2 Airborne and UAV Platforms

One of the most notable features of UAVs, also known as drones, include centimetre to decimetre resolution, as well as flexible scheduling with customisable sensors[100][7]. Their main application is in the field of research with scales between plot to field and are important in precision agriculture, where they are used to do ultra-high resolution yield mapping, phenotyping, and variable rate management.

Hyperspectral sensors which are characterized by the capability of measuring hundreds of narrow spectral bands provide highly detailed biochemical information which is essential to special crop surveillance and other research uses[101][102]. They are used in determining the extent of vegetation, detecting stress in crops as well as estimating the biochemical parameters.

The use of UAVs to predict the yield of grain crops is systematically reviewed by [7] and highlights the advantages associated with using ultra-high spatial and time resolution to monitor fine-scale variations within and between fields. This in-depth study highlights the enormous contribution of UAVs and hyperspectral sensors to development of agricultural monitoring and management systems[6].

5.2.8 Spectral Indices and Derived Features

Yield prediction using remote sensing methods is extremely dependent on derived spectral indices and biophysical parameters which are linked to crop health, biomass and crop productivity.

5.2.8.1 Vegetation Indices

The most commonly used vegetation index is NDVI, which is determined as the following formula NIR - Red/NIR + Red and is associated with green biomass and leaf area index[88][2][99][81]. Enhanced Vegetation Index (EVI) has the following advantages, 2.5 x(NIR-Red)/(NIR + 6xRed − 7.5xBlue + 1) has less atmospheric implications than NDVI[2]. The purpose of using the Soil-Adjusted Vegetation Index (SAVI) is to reduce the influence of soil brightness in thin canopies[3]. Studies[99] have demonstrated that the relationship of green chlorophyll vegetation index (GCVI) with corn yield has been the highest. Simple Ratio (SR) which is NIR/Red has been found to perform better than conventional indices in estimating wheat yield[3]. Also, water related indices such as NDWI and LSWI are employed to determine canopy water content and moisture stress[95]. The following Table 3 indicates the usual vegetation indices and their uses.

Table 3

Common Vegetation Indices and Their Applications
Index	Formula	References	Sensitivity	Optimal Range	Primary Applications	Limitations
NDVI	(NIR - Red) / (NIR + Red)	[88][2] [99][81]	Green biomass, LAI	0.2–0.8	General vegetation monitoring	Saturates at high biomass
EVI	2.5 × (NIR - Red) / (NIR + 6×Red − 7.5×Blue + 1)	[2]	Biomass, reduced saturation	0.2–0.9	Dense canopies, high biomass	Complex calculation
SAVI	1.5 × (NIR - Red) / (NIR + Red + 0.5)	[3]	Vegetation with soil background	0.1–0.7	Sparse canopies, early season	Requires soil adjustment factor
MSAVI	(2×NIR + 1 - √[(2×NIR + 1)² − 8(NIR-Red)]) / 2	[3]	Soil-adjusted vegetation	0.1–0.7	Variable soil backgrounds	Complex interpretation
SR	NIR / Red	[3]	Biomass, chlorophyll	1–10+	Simple ratio, wheat yield	Sensitive to atmospheric effects
GCVI	(NIR / Green) − 1	[99]	Green chlorophyll	Variable	Corn yield, chlorophyll content	Less studied than NDVI
NDWI	(NIR - SWIR) / (NIR + SWIR)	[95]	Water content	-1 to + 1	Moisture stress, irrigation	Sensitive to atmospheric water
LSWI	(NIR - SWIR) / (NIR + SWIR)	[95]	Liquid water, canopy water	-1 to + 1	Water stress detection	Requires SWIR bands
GNDVI	(NIR - Green) / (NIR + Green)	[95]	Chlorophyll content	0.2–0.8	Nitrogen status, photosynthesis	Similar to NDVI limitations

The table above shows the common vegetation indices utilised and their application. It shows the formulas, sensitivity, optimal range, and its limitations.

5.2.8.2 Biophysical Parameters

LAI is defined as the total one-sided leaf area per unit of ground area and serves as a strong predictor of biomass and yield[95][7]. Biomass estimates can be derived from multispectral or hyperspectral data, which are utilized as intermediate predictors[7][95]. Additionally, canopy height models are extracted from technologies such as LiDAR or photogrammetric UAV data[7].

5.2.8.3 Advanced Features

The paper addresses several strategies of learning features in a data analysis scenario. It puts much emphasis on the use of PCA, randomized-PCA, autoencoders, and deep representations learning as an alternative to, or augmentation of, traditional hand-crafted indices[89][12][86]. Another aspect covered in the text is the time characteristics, which are phenological measures of time of maximum VI and the total VI at the time of the season, or seasonal tendencies[94][10]. Further, it discusses the importance of texture and spatial context including local neighbourhood data and spatial autocorrelation characteristics as very important aspects of the analysis[10].

5.3 What crop types and geographic regions have been studied?

5.3.1 Crop Types

The analysed literature includes a broad range of crops, with a focus mostly on staple cereals and oilseeds. The following are the important discoveries: Wheat has been highly studied in different areas such as China, India, Pakistan, Europe, and North America[3][4][81][103]. Maize (or corn) has been a central target in high stakes agricultural regions like the U.S. Corn Belt, China, and Brazil[6][86][6]. Brazil, Argentina, the U.S., and China have high research on soybean[86][10][104]. The studies by Rice focus on the nations of Asia, especially China, Taiwan, and India[104][84]. Other crops that were also studied in the literature include rapeseed / canola, sorghum, potato, tobacco and many types of tree crops[95][105][105].

Table 4

Performance Summary by Crop Type and Method
Crop	Method	Spatial Scale	R² Range	RMSE Range	Representative Study	Key Finding
Wheat	LSTM	Field	0.85–0.90	0.20–0.25 t/ha	[3]	LSTM outperformed RF, SVR, and GBDT
Wheat	Dual-stream DL	County	0.79	650 kg/ha	[3][4][81][103]	Sequential + static features are effective
Wheat	Deep learning	District	0.75–0.85	Variable	[3]	Multi-source data improved accuracy
Maize/Corn	Attention DL	County	0.896	908 kg/ha	[3][4][6]	The attention mechanism highlighted critical stages
Maize/Corn	PlanetScope + Sentinel-2	Field	0.81	Variable	[106]	High resolution captured within-field variability
Maize/Corn	Deep GP	County	0.75–0.85	Variable	[86]	Spatial GP captured the spatial correlation
Soybean	Feature fusion	Field	0.86	Variable	[10]	Multi-modal fusion superior
Soybean	ML + weather	Regional	0.70–0.80	Variable	[86][10][104]	Weather integration critical
Rice	ML time series	Regional	0.70–0.80	Variable	[84]	Temporal patterns important
Sorghum	ML + RS	Regional	0.65–0.75	Variable	[94][10][105]	Useful in data-scarce regions
Multi-view	Adaptive fusion	Sub-field	0.68	Variable	[94]	Adaptive weighting improved performance
Multi-view	Adaptive fusion	Field average	0.80	Variable	[94]	Aggregation improved accuracy

5.3.2 Geographic Distribution

Studies examined major agricultural regions worldwide, detailing key crops cultivated in the following countries: In Asia, China is noted for wheat, maize, and rice[3][4][6][84][81][103]India primarily grows wheat and multiple other crops; Pakistan primarily grows wheat; and Taiwan focuses on rice. North America features the USA with corn and soybeans, and Canada with wheat and canola[86][96][6]. In South America, both Brazil and Argentina are significant producers of soybeans, while Uruguay also contributes to this sector[10][104]. Europe includes Germany and several multi-country studies, while the agricultural landscape in Africa highlights South Sudan's production of sorghum[94][10][105]. Table 4 shows the performance by crop type and method utilised.

Table 5

Geographic Distribution of Studies
Related works	Region	Number of Studies	Primary Crops	Key Characteristics	Data Availability
[86]; [101]	East Asia (China)	25	Wheat, maize, rice	Large-scale studies, county-level focus, advanced DL	Moderate–High
[97][98][96]	South Asia (India, Pakistan)	8	Wheat, multiple crops	District-level, operational systems	Moderate
[96][96]	North America (USA, Canada)	15	Corn, soybean, wheat	Corn Belt focus, high-quality ground truth	High
[9]	South America (Brazil, Argentina)	8	Soybean, maize	Field-level precision agriculture	Moderate
[88]	Europe (Germany, multi-country)	10	Wheat, rapeseed, multiple	Multi-country comparisons, precision agriculture	High
[9][98]	Africa (South Sudan, others)	3	Sorghum, multiple	Data-scarce, operational challenges	Low
[101] [89]	Southeast Asia (Taiwan, others)	4	Rice, multiple	Regional forecasting	Moderate
[65]	Global / multi-region	12	Multiple	Large-scale comparisons, transferability	Variable

5.3.3 Spatial Scales

Research has been done on different spatial scale in agricultural research and forecasting. At the plot or field level, studies based on the usage of UAVs and high-resolution satellite images have been conspicuous as indicated in references[7][102][89]. On a smaller scale of sub-field, precision agriculture has been greatly utilized and the related literature may be found in reference [91][99]. On a greater scale, namely on the administrative unit level, that is, the county or the district level, the focus is on forecasting within administrative units as explained in reference [4][86][81]. Also, the regional and national levels are taken care of by the large-area operational forecasting, as mentioned in the sources [88][96]. Lastly at a global level, there has been a drive to observe the agricultural outputs at the continental level, all the way to the global level with the findings in reference [88].

5.3.4 Temporal Aspects and Lead Time

5.3.4.1 Temporal Resolution and Phenology

As demonstrated in the results of different studies, temporal information is important in predicting yield due to the thoroughness of certain aspects of growth and climate factors. The main areas of focus are the application of time series observations of growth season where sequential growth observations are made to monitor the phenological development[4][11][89]. Flowering, grain filling and maturity are the critical development phases that are stated as the most predictive areas when it comes to yield results[6][99]. Also, it has been suggested to use multi-year training data to improve the robustness of the model and in a manner that it will capture the variation of climate over time[11][96]. A major conclusion made in[32] shows that the model agreement (with an R2 of 0.81) is best when 86 days after seeding, which is essential in the realization of the usefulness of growth stage observation in the prediction of yield[99].

5.3.4.2 In-Season Forecasting and Lead Time

The lead time of the predictions in agricultural forecasting is important in their operational value. Predictions made early in the season are less accurate and have great operational value in terms of planning[4]. The mid-season predictions are balanced as they provide the precision at 1–2 month prior to the harvest hence enabling decision-making. Conversely, pre-harvest estimates have the highest accuracy levels but they have a smaller lead time with regard to any intervention that might be required[4][6]. One interesting conclusion of the study is that deep learning algorithms have been able to make useful predictions of the most important crops on a county basis like winter wheat, maize, some 1–2 months before harvesting season[4][6][96].

5.4 What performance metrics are reported, and what accuracy ranges are achieved?

5.4.1 Common Evaluation Metrics

Regression metrics studies make use of a number of standard measures that determine the performance of a model. The most commonly reported one is the R2, which is the fraction of variance that the model has been able to explain[3][4][6][91][10][99]. RMSE represents absolute error in units of yield, e.g. tons/hectare (t/ha), kilograms/hectare (kg/ha), and bushels/acre (bu/ac)[3][4]. Deep Learning Model with an attention mechanism and remote sensing to estimate maize in the black soil area of Northeast China[6][95]. MAE is determined to obtain the mean relative difference between the actual values[4][95]. Also, Percentage Errors, such as the sMAPE, MAPE and MASE, have been used to evaluate the relative performance of the models[4]. The frequency of used metrics in the various studies below is indicated in Fig. 6.

Table 6

Frequency of Evaluation Metrics Used in Crop Yield Prediction Studies
Evaluation Metric	Number of Articles	Relative Frequency	Typical Use Case	Implication for model comparison
R² (Coefficient of Determination)	38	Very High	Model explanatory power and goodness-of-fit	Tends to overstate model performance, particularly in regional or national studies, making cross-scale and cross-crop comparisons unreliable
RMSE	34	Very High	Absolute prediction error magnitude	Favours models optimized for extreme-error reduction, limiting comparability when yield ranges differ
MAE	21	Moderate	Robust error estimation	Provides fairer comparison of typical performance, but limited adoption reduces benchmarking consistency
MAPE / sMAPE	9	Low	Relative error comparison	Improves comparability across crops and regions, but inconsistent normalization undermines reproducibility
NRMSE	7	Low	Cross-region normalization	Biases comparisons toward higher-yield systems, limiting applicability in smallholder or low-yield contexts
MSE	6	Low	Intermediate loss metric	Rarely used for comparative synthesis, mainly relevant for model optimization rather than evaluation
Bias / Mean Bias Error	4	Rare	Systematic bias analysis	Can misrepresent predictive skill, unsuitable as a standalone comparison metric

Fig. 6

The frequency utilized in the evaluation metrics distribution in articles

5.4.2 Performance Ranges by Method and Scale

Regional applications have baseline statistical models with a typical R 2 of between 0.50 and 0.70[81]. Conversely, the R2 values of RF, SVR, and XGBoost are between 0.60 and 0.80 when used between field and county scale, with RMSE between 0.260 and 0.489 tons per hectare in the example of wheat[2][3][84]. DL methods, such as CNN, LSTM networks, and Attention mechanisms, are more successful with higher performance, as they optimize their applications reached a range of R2 between 0.75 and 0.90[3][4][6]. The results are significant, with an R2 of 0.79 and a RMSE of some 650 kg/ha of a winter wheat county model, and an R2 of 0.896 and an RMSE of some 908 kg/ha of a maize model with attention.[4]The LSTM model of wheat is 0.201 tons per hectare[3]. Multi-modal fusion methods will give an R2 of approximately 0.68 in sub-field predictors, and an impressive 0.80 in field means, and an impressive R2 of 0.86 in soybean when using feature fusion[10]. The R2 of high-resolution satellite images of PlanetScope and Sentinel-2 indicates 0.81 in the favourable periods of corn growth[106]. Lastly, the use of UAVs has very different results with a range of R2 values that range between 0.50 and 0.99, depending on the crop, sensor and technique used[7][106].

5.4.3 Factors Affecting Performance

The results of the performance differ considerably based on a range of factors: Spatial Scale demonstrates better performance at aggregated scales because of averaging effects[94]; Crop Type indicates variability associated with canopy architecture and phenology[106]; Data Fusion indicates that multi-source approaches are better than single-source strategies[8][94]; Temporal Sampling implies that denser time series improve performance, as is the case with deep learning applications[4][89]. Ground Truth Quality explains the importance of accurate reference data in order to have effective training and validation[12][14].

5.5 What are the key challenges, limitations, and knowledge gaps in current research?

The literature which has been reviewed has always revealed a number of pertinent challenges which limit generalization of models and operational implementation.

5.5.1 Data Availability and Ground Truth Scarcity

The scarcity of training data of agricultural yield prediction is a known issue particularly in the developing world where the yield labels (plot- and county-level) are prone to be limited[14]. In addition, remote sensing pixel resolution does not correspond to units of ground truth aggregation[106]. Moreover, there are temporal gaps owing to the records of incomplete historical yields, which restrict the effectiveness of training multi-year models[86]. To solve these problems, a number of mitigation measures have been suggested such as transfer learning, domain adaptation and representation learning, which will minimize the use of labelled data[12][14].

5.5.2 Cloud Cover and Optical Data Gaps

The use of cloud pollution is one of the major issues that affect the effectiveness of optical sensors particularly in humid tropical regions and during the monsoon season [8]. To resolve this problem, a number of mitigation measures have been offered: all-weather operational capability can be enhanced by fully integrating 2SAR and optical data, the continuity of data can be achieved by using the algorithms of temporal compositing and gap-filling, and the use of dense time series provided by high-revisit satellite constellations such as Sentinel-2 and PlanetScope can help to improve the reliability of the data[4][8][106].

5.5.3 Spatial Resolution Trade-offs

The reading talks about the various solutions to remote sensing technologies and their uses. The details of coarse resolution data like the one found in MODIS have been observed to fail to capture the within-field variability, making it only appropriate in large-scale implementations[88][96]. Conversely, fine-scale patterns can be observed in high-resolution data provided by UAVs or commercial sources, however, it is less scalable and can be expensive per large scale[7]. In response, Sentinel-2 of 10 meters resolution and Landsat of 30 meters are mentioned as offering an optimal trade-off between the different applications[3][106].

5.5.4 Model Generalization and Transferability

Black-box machine learning models and deep learning models are prone to location-specific overfitting, learning patterns that are unique to a single location, and which do not translate similarly to other locations[12][14]. The quality of data and the capacity to generalize the model results from remote sensing to predict yield is one of the major challenges associated with remote sensing to predict yield because the performance of machine learning models depends on the quality of the input data and is dependent on whether the model can generalize to different regions and crop types[87]. The interactions between genotype, the environment and the management are also complex and it is thus challenging to separate the three without directly modelling them[12]. Moreover, models that are trained in a particular region perform poorly once implemented in other regions without any retraining[14]. Researchers suggest a number of approaches to alleviate such concerns: representation learning should be used to decouple the effects of both genotype and environment, transfer learning and domain adaptation methods should be used, hybrid physical-statistical models should be developed, and cautious cross-validation approaches should be employed with attention to spatial and temporal factors[11][14][1][2][12].

5.5.5 Multi-Modal Fusion Complexity

The heterogeneous data is difficult to handle because it has diverse spatial, temporal, and spectral features of different sensors[94][10]. Negative transfer can occur in which poor fusion strategies can result in poor performance[94]. Also, processing and fusing of processing of several data streams is very expensive computationally[10]. The development of adaptive fusion architectures and the application of attention mechanisms to weight the contribution of various modalities in a proper manner are suggested as solutions to the problems[94][10].

5.5.6 Evaluation and Reproducibility Gaps

The fact that the performance measures are not reported consistently across various studies poses problems with effective cross-study comparison[2][11]. This is also made not easy by the fact that it uses different scales e.g. plot, field, county, and regional predictions which cannot be readily compared[94]. Also, a lack of standardized datasets and evaluation guidelines is quite significant, which impedes sound benchmarking in this direction. A number of studies present insufficient methodological information and do not share code and data, which are associated with the problem of insufficient reproducibility[2].

Table 7

Key Research Gaps and Future Directions
Related Works	References	Research Gap	Current Status	Priority Level	Expected Timeline	Required Resources
Statistical & empirical models using NDVI, LAI	[82][81]	Limited capacity to model nonlinear relationships	Mature; mainly baseline methods	Low	Short-term	MODIS/Landsat data, basic computing infrastructure
Machine learning models: RF, SVR, XGBoost	[2][3]	Risk of overfitting; limited spatial-temporal learning	Widely adopted; operational in many regions	Medium	1–3 years	Sentinel-2 imagery, yield statistics, ML toolkits
Deep learning models: CNN, LSTM	[63][4]	High data volume and computational cost; low interpretability	Rapidly advancing research area	High	2–4 years	Large labeled datasets, GPUs, cloud platforms
Hybrid CNN–LSTM and attention-based DL models	[25] [6]	Limited transferability across regions and crops	Emerging; limited operational deployment	Very High	3–5 years	Multi-source RS data, advanced DL frameworks
Optical–SAR fusion using Sentinel-2 & Sentinel-1	[8][72]	Complex fusion and preprocessing workflows	Increasing adoption, especially in cloudy regions	High	2–4 years	Optical & SAR data, preprocessing pipelines
Transfer learning & Explainable AI	[13] [14]	Lack of standardized evaluation and stakeholder trust	Early-stage research trend	Very High	5 + years	Pretrained models, XAI tools, and high-performance computing

5.6 What emerging trends and future research directions can be identified?

5.6.1 Advanced Deep Learning Architectures

Attention mechanisms are concerned with dynamically locating relevant spatial places and time steps, which facilitate a process of data processing[6][89]. Transformer models have been observed to have the ability of modelling long-range dependencies of spatial-temporal data, which is an up-and-coming trend in the field. The self-supervised learning is increasingly being popular because it can pretrain on unlabelled remote sensing data, thus reducing the requirement of large, supervised datasets.

5.6.2 Explainable AI and Interpretability

The issue with black-box models in machine learning is that it undermines trust and agronomic knowledge by the stakeholders[11]. To manage this, one can use a few methods, such as feature importance analysis, attention visualization and SHAP values and saliency maps[6]. Also, there is an alternative promising solution - the hybrid models between interpretable process models and machine learning and deep learning methods[11].

5.6.3 Integration with Crop Growth Models

The process of combining remote sensing measurements with mechanistic crop models, like WOFOST, DSSAT and APSIM, is called data assimilation[11]. This method has a number of benefits such as compliance with physical limits, better extrapolating behaviour and better explanation of the findings[11]. Nonetheless, it is also associated with its challenges, in particular, linked with the computational complexity, the need to measure the parameters, and the uncertainties concerning the model structure[11].

5.6.4 Multi-Source and Multi-Scale Fusion

Adaptive Fusion is a form of learning to give the best weighting to a combination of different sources of information enabling more accurate incorporation of information[94]. The cross-Scale Integration deals with coarse and frequent observation combined with the fine and periodic observation to improve the utility of the data[94]. The use of distinct data bearing processing pipes in the processing of varying data types, such as optical, SAR, weather and soil data is known as Modality-Specific Encoders and will guarantee that each modality is well-processed[10].

5.6.5 Transfer Learning and Domain Adaptation

The article identifies three major strategies of enhancing the agricultural practices. The former is Cross-Regional Transfer that is based on training models on data-rich areas and then fine-tuning it to specific target areas[14]. The second plan is aimed at the Disentangled Representations in order to differentiate the genotype, environment and management effects[12]. The third method is Few-Shot Learning, which allows making fast adaptations to new crops or regions with few labelled data, which explains the topicality of this study as a new direction in the field.

5.6.6 Operational and Real-Time Systems

Google Earth Engine and Microsoft Planetary Computer are also cloud computing platforms that are in an emerging trend to enable scaled processing of geospatial data. Another significant trend is automated pipelines, which are end-to-end systems that simplify the chain of processes of data acquisition to yield forecasts. More so, policy tools and farm management are getting increasingly integrated with decision support systems, which improves decision-making in agricultural practices.

6 DISCUSSIONS

6.1 Synthesis of the key finding

This is a systematic literature review that has been based on 106 studies on remote sensing-based crop yield prediction. According to the synthesising of the findings, the field of remote sensing in crop yield prediction with the utilisation of ML and DL is advancing at a very high pace, with the wide range of different methodologies, the expansion of the range of data sources, and the tendency to become more and more sophisticated in its operations.

The trend in methodological developments in agronomy has changed in both directions. The primitive empirical relationships between vegetation indices and crop yield have given way to sophisticated DL systems capable of operating on multi-modal and spatial-temporal data. Conventional ML methods, such as RF, SVR, and XGBoost are solid baselines that have moderate computational requirements. DL techniques, such as CNN, LSTM, and other attention mechanisms, on the contrary, perform well with complex, high-dimensional inputs, particularly in cases where substantial volumes of training data exist[88][2][3][4][6]. CNN + LSTM which is the hybrid performs better because it utilises both spatial and temporal data. Moreover, there is also the approach of hybrid methodologies involving the integration of physical crop models and data-driven strategies that is also becoming an attractive way to combine the mechanistic insights with complex pattern recognition[11].

The current yield forecast systems have deployed a wide range of remote sensing technologies to increase the odds of the data sources. Sentinel-2 is one of them and is considered to be one of the crucial sensors due to its desirable combination of the 10 m spatial resolution, frequency of revisit 5 days, wide spectral coverage, and the fact that it is free of charge[3][94][99]. Also, SAR sensors can provide valuable complementary information especially where clouds are common, where a combination of optical and SAR methods have proven to be more effective than when only one source is used[8]. The UAVs enable ultra-high-resolution imaging which can be used in precision agriculture and plant phenotyping, but they have limited scalability[7]. The introduction of new inventive products like the Sun-Induced Fluorescence (SIF) indicates the need to incorporate new biophysical indicators in order to increase the predictability of the final yields.

State-of-the-art techniques in performance achievements usually exhibit R² values of approximately 0.75 to 0.90 and this is a high predictive power as long as enough training data is available[3][4][6][94][10]. Spatial scale also affects performance with best results at high levels of aggregation. Dense temporal sampling improves the training of deep neural networks and data merging techniques, with the results of multi-modal ones consistently being better. However, the direct comparisons of performance among different studies are made difficult due to differences in crops, scales, areas and assessment regimes.

Although there are major accomplishments, critical obstacles interfere with the deployment of operations in the field of yield monitoring[12][14]. The absence of ground truth data is especially a problem with model training in developing regions. Optical sensors still have problems with cloud pollution, and some of the mitigation is provided by SAR fusion[8]. One of the most important issues is that very little generalization of models is made across regions, years and crop types to create overfitted and lower transferability[12][14]. The challenges form the basis of ongoing research in transfer learning, domain adaptation as well as hybrid physical-statistical approaches.

Standardized benchmark datasets, evaluation protocols and performance reporting guidelines are required to promote the field and enable stringent cross-study comparisons[2][11]. Future studies must put more focus on model transferability, as shown by domain adaptation, disentangled representation, and introducing agronomic knowledge[12][14]. A promising direction will be hybrid methods that combine machine learning and deep learning data with process-based crop models by means of data assimilation, which builds generalizable and comprehensible systems[11]. In addition, open science through sharing code, data, and pretrained models will accelerate the pace and improve reproducibility[2].

The level of maturity of remote sensing based yield prediction is such that it can be operationalized especially in in-season prediction at the county to regional scales[4][96]. The choice of sensors: Sentinel-2 is a great free base that can be used in the majority of applications, SAR fusion is to be applied to the regions with a high probability of cloud cover, and UAVs are to be used in accuracy agriculture in the field scales[8][94][99][7]. Regarding the methodology, the Random Forest algorithm is cited as a powerful and readable option in the field of operational systems, but deep learning might be better in case of sufficient training data and computing capabilities[2][3][4]. Lastly, it is pointed out that remote sensing can be considered as a supplement to, but not a substitute of the ground-based monitoring systems, which are already in place, and can be combined successfully with the decision support tools[11].

The systematic review addresses the gaps of previous syntheses in a number of ways: it uses a multi-database search strategy to ensure comprehensive coverage of the literature and it is systematic and reproducible (PRISMA guidelines) [88][2][11][7][106]. It encompasses recent articles published between 2024 and 2025 with the latest developments in attention mechanisms, transfer learning, and adaptive fusion[6][10][12][14][94][10]. It is also in consideration of spatial scale effects between plot level to global scale[94][99][7]. Lastly, it provides a practical and actionable advice in the selection of sensors, methods, and approaches that can be used in particular application settings.

There are a number of notable research limitations that should be taken into consideration:- Firstly, publication bias could create a bias of positive outcome which in turn can bias the performance measurements. Secondly, there is language bias that can be caused by a preponderance of publications in English which may exclude useful regional research. Third, the difference in methodologies used by studies has affected the possibility of quantitative meta-analysis. Also, the fast development of the field implies that the results can be soon forgotten as the approaches and sensors are being improved continually. Finally, the omission of grey literature, including technical reports and documentation of the operational systems, may lead to the loss of practice-based information.

6.1 Future Research Directions

Upon a literature review, the following topical research priorities have been identified: building of multi-crop, multi-region, and multi-year benchmark datasets with standard train/test partitions[2]; setting up of reporting standards on performance measures, validation procedures, and methodology[2]; and development of open repositories of code, pretrained models, and evaluation schemes[2].

The most recent transfer learning methods are suggested to help with generalization to various regions, crops, and years[14]. Further studies are also aimed at coming up with ways of isolating the influences of genotype, environment, and management practices[12]. In addition, there are also priorities to investigate the methods of few-shot and zero-shot learning in order to adapt to new agricultural environments in the shortest possible time, responding to the new demand of the field.

The combination of ML/DL methods with mechanistic crop models can be used to improve agricultural modelling with data assimilation and model coupling[11]. An interest in the creation of differentiable crop models, which could be trained with the help of a gradient-based optimizer, is a new direction in the field. Moreover, the integrative paradigm of physical constraints and flexible representation learning is regarded as a potentially promising strategy to enhance the given models[11].

It is important to develop interpretable architectures in order to deliver agronomic details in addition to predictions that allow a better understanding and decision-making [11]. To gain an understanding of model decisions, more improvements in visualization methods, including attention maps and feature importance, are required[6]. The stakeholders involved in the co-design of these interpretable systems will ensure that the needs and views of the affected will be correctly reflected and incorporated into the design process[11].

In order to improve agricultural surveillance and decision making, there is a need to have adaptive fusion architectures that effectively navigate the heterogeneous spatial-temporal resolutions[94]. This includes exploring the best combinations of sensors depending on the types of crops, sizes, and the size of the region[8][94][10]. In addition to this, there is a need to develop the methodologies of integrating various types of data such as satellite imagery, UAV data, weather patterns, soil, and practices of management[10].

The document presents three major emerging requirements in agricultural technology; First, it is necessary to convert research prototypes into operational, automated systems to improve efficiency and productivity. Second, it focuses on intersecting yield predictions with the farm management and policy decision support systems, which would enhance the improved decision-making process by farmers and policymakers. Finally, the document recommends the creation of real-time, cloud-based solutions of monitoring crops across the globe to efficiently monitor and manage agricultural products in a global perspective.

In order to improve the accuracy of agricultural data, the ground truth yield data collection must be increased, in particular in developing areas[12][14]. This need should be met by developing new techniques to leverage such alternative sources of data as crowdsourcing and mobile applications. Moreover, the need to investigate the methods of synthetic data creation and enhancement is increasing to enhance data gathering and processing.

The document presents three of the emerging needs in the field of climate and agricultural yield management. To begin with, it focuses on the creation of models that are resilient to new climate realities and extreme weather impacts, where there is a need to adjust to prevailing environmental conditions. Second, it highlights the significance of exploring early warning systems, which can detect climate-based yield shocks and ensure that the effects of unfavourable weather on the crop yield can be alleviated. Finally, it recommends the evaluation of model performance when conditions of non-stationary climate is involved because there is a need to have predictive tools that are reliable in an ever-changing climate.

7 CONCLUSIONS

The proposed systematic literature review that follows PRISMA guidelines synthesizes the results of 106 papers that focus on the use of remote sensing in predicting crop yields. It shows a substantial shift of simple models to sophisticated methods of deep learning with R2 values of 0.75 to 0.90 in in-season forecasting. The review cites such key data sources as Sentinel-2, SAR, and UAV technologies, with the fusion of sensors being superior to non-interfering data sources. The challenges that were still present even with the advances include lack of data and resolution problems. Potentially useful, promising approaches such as transfer learning and adaptive fusion are also mentioned. The practical implications have shown that yield predictions based on remote sensing are especially useful in in-season predictions at a county or regional level. It is recommended that Sentinel-2 should be taken as a baseline data source base to be supplemented by SAR in cloud-prone regions and UAVs in precision agriculture. Furthermore, the choice of methodologies ought to be done with regard to performance, interpretability, of the computation requirements, and training data availability with integration into existing ground-based monitoring systems with regard to operational efficacy. Research priorities are focused on setting standardized metrics of model evaluation, improving the generalization of models through transfer learning, developing explainable AI, moving research prototypes to operational systems and addressing data gaps in developing regions to assure equitable gains of AI. Conclusively, remote sensing and high-order machine learning culture creates a strong avenue towards predicting crop yields, and it has a favourable impact on food security and agricultural policy, with a direction towards actual implementation and sustainability and stakeholder involvement in the global agricultural field.

8 Contribution to the knowledge

The systematic literature review contributes to the body of knowledge on remote sensing-based crop yield prediction in a distinct and original way:

8.1 Complete modelling paradigms integration.

The review provides a synthesis that is PRISMA compliant, which incorporates statistical, machine learning, deep learning and hybrid modelling paradigms in one comparative frame. The study clarifies the trend of predictive performance and feasibility of operation with the complexity of models by systematically benchmarking the performances of these methods on various crops, spatial scales, and data sources.

8.2 Sensor-model characterization based on evidence.

The presence of the study contributes to the knowledge of understanding the relationship between different sources of remote-sensing data: optical, SAR, UAV, multisource fusion with the modelling strategies to influence the accuracy, the robustness and the scalability in yield prediction. The review has shown recurrent trends that show when sensor fusion and high resolution data produces any meaningful gains and when simpler data configurations are adequate.

8.3 Critical synthesis of evaluation practices and methodology bias.

One of the contributions is analytical analysis of the metrics of evaluation used in research on yield prediction. The review discloses the structural biases underlying variance-based measures of performance, like R 2 and RMSE, which inflate reported performance and reduce cross-study comparability and thus provides best-practice advice on standardized model evaluation.

8.4 Cross cutting limitations in generalization and deployment identification.

The critique is systematic in integrating enduring challenges such as the lack of data, model transferability, cloud pollution, and lack of interpretability that limits the operational use of sophisticated yield-prediction models. These lessons go beyond case studies to point out barriers in the field to scalable implementation.

8.5 Development of a research agenda with a vision.

Based on the evidence synthesized, the paper presents priority research directions, which include data-efficient learning, explainable and physics-informed models, standardized benchmarking, and regionally transferable models. This is the agenda that provides strategic direction to the development of crop-yield prediction research to a reliable, real world decision support.

ACKNOWLEDGMENTS

This systematic literature review was conducted in accordance with PRISMA guidelines and recognizes the contributions of authors from all reviewed studies in furthering the advancement of remote sensing-based crop yield prediction.

Abbreviations

AI Artificial Intelligence

ANN Artificial Neural Network

CNN Convolutional Neural Network

CRediT Contributor Roles Taxonomy

DL Deep Learning

EVI Enhanced Vegetation Index

GBDT Gradient Boosting Decision Tree

GEE Google Earth Engine

GIS Geographic Information System

GNDVI Green Normalized Difference Vegetation Index

LAI Leaf Area Index

LSTM Long Short-Term Memory

MAE Mean Absolute Error

MAPE Mean Absolute Percentage Error

ML Machine Learning

MODIS Moderate Resolution Imaging Spectroradiometer

MSE Mean Squared Error

NDVI Normalized Difference Vegetation Index

NDWI Normalized Difference Water Index

NRMSE Normalized Root Mean Square Error

PRISMA Preferred Reporting Items for Systematic Reviews and Meta-Analyses

R² Coefficient of Determination

RF Random Forest

RMSE Root Mean Square Error

RS Remote Sensing

SAR Synthetic Aperture Radar

SAVI Soil-Adjusted Vegetation Index

SLR Systematic Literature Review

SVM Support Vector Machine

UAV Unmanned Aerial Vehicle

XGBoost Extreme Gradient Boosting

Funding

No financial support was offered to the current inquiry.

Author Contribution

S.Z. was engaged in the conceptualization and design of the study, and in preparing the materials, data analysis and writing the manuscript. A.Z. also gave advice, helped in design, data analysis and even helped in the conception to meet the standards of the published work. B.J. provided advice in regard to the data analysis. All the authors revised and approved the final manuscript.

Competing Interests

The authors do not have any conflicts of interest.

Ethics Approval

Not applicable.

Consent to Participate

Not applicable.

Consent to Publish

Not applicable.

REFERENCES

Muruganantham P, Wibowo S, Grandhi S, Samrat NH, Islam N. A Systematic Literature Review on Crop Yield Prediction with Deep Learning and Remote Sensing. Remote Sens. 2022;14(9). 10.3390/rs14091990.

Chlingaryan A, Sukkarieh S, Whelan B. Machine learning approaches for crop yield prediction and nitrogen status estimation in precision agriculture: A review. Comput Electron Agric. 2018;151:61–9. https://doi.org/10.1016/j.compag.2018.05.012.

Cheng E, et al. Wheat yield estimation using remote sensing data based on machine learning approaches. Front Plant Sci. 2022;13:1–16. 10.3389/fpls.2022.1090970.

Huang H, et al. Developing a Dual-Stream Deep-Learning Neural Network Model for Improving County-Level Winter Wheat Yield Estimates in China. Remote Sens. 2022;14(20). 10.3390/rs14205280.

Hazarika J, Roy M. Chapter-6 Satellite Farming. Chief Ed Dr RK Naresh, p. 95, 2023.

Li X, Lyu Y, Zhu B, Liu L, Song K. Maize yield estimation in Northeast China ’ s black soil region using a deep learning model with attention mechanism and remote sensing, pp. 1–23, 2025.

Yuan J, Zhang Y, Zheng Z, Yao W, Wang W, Guo L. Grain Crop Yield Prediction Using Machine Learning Based on UAV Remote Sensing: A Systematic Literature Review. Drones. 2024;8(10). 10.3390/drones8100559.

Mateo-Sanchis A, Piles M, Muñoz-Marí J, Adsuara JE, Pérez-Suay A, Camps-Valls G. Synergistic integration of optical and microwave satellite data for crop yield estimation. Remote Sens Environ. 2019;234:111460. https://doi.org/10.1016/j.rse.2019.111460.

Mena F et al. December., Adaptive fusion of multi-modal remote sensing data for optimal sub-field crop yield prediction, Remote Sens. Environ., vol. 318, no. 2024, p. 114547, 2025. 10.1016/j.rse.2024.114547

10.

Miranda M, Pathak D, Nuske M, Dengel A. Multi-Modal Fusion Methods with Local Neighborhood Information for Crop Yield Prediction at Field and Subfield Levels, in IGARSS 2024–2024 IEEE International Geoscience and Remote Sensing Symposium, 2024, pp. 4307–4311. 10.1109/IGARSS53475.2024.10640993

11.

Xiao G, et al. Progress and Perspectives of Crop Yield Forecasting With Remote Sensing: A review. IEEE Geosci Remote Sens Mag. 2025;13(3):338–68. 10.1109/MGRS.2025.3571906.

12.

Powadi AA et al. Enhancing yield prediction from plot-level satellite imagery through genotype and environment feature disentanglement, Front. Plant Sci., vol. 16, no. September, pp. 1–16, 2025, 10.3389/fpls.2025.1617831

13.

Pan SJ. Transfer learning. Data Classif Algorithms Appl. 2014;537–70. 10.1201/b17320.

14.

Huber F, Inderka A, Steinhage V. Leveraging Remote Sensing Data for Yield Prediction with Deep Transfer Learning. Sensors. 2024;24(3). 10.3390/s24030770.

15.

Di L, Yu E. Remote Sensing. in Remote Sensing Big Data. Cham: Springer International Publishing; 2023. pp. 17–43. 10.1007/978-3-031-33932-5_2.

16.

Khanal S, KC K, Fulton J, Shearer S, Ozkan E. Remote Sensing in Agriculture (Challenges and Opportunities). Remote Sens, 10, pp. 83–7, 2020.

17.

Mondal BP, Dutta S, Moursy ARA. Advances of Remote Sensing Technologies in Agriculture. Artif Intell Comput Vis Ecol Inf, 2025.

18.

Gao F. Remote Sensing for Agriculture. in Agro-geoinformatics: Theory and Practice. Springer; 2021. pp. 7–24.

19.

Panda SS, Rao MN, Thenkabail PS, Misra D, Fitzgerald JP. Remote sensing systems—Platforms and sensors: Aerial, satellite, UAV, optical, radar, and LiDAR. in Remote Sensing Handbook, Volume I. CRC; 2024. pp. 3–86.

20.

Weiss M, Jacob F, Duveiller G. Remote sensing for agricultural applications: A meta-review. Remote Sens Environ. 2020;236:111402. https://doi.org/10.1016/j.rse.2019.111402.

21.

Zhao Y, Potgieter AB, Zhang M, Wu B, Hammer GL. Predicting wheat yield at the field scale by combining high-resolution Sentinel-2 satellite imagery and crop modelling. Remote Sens. 2020;12(6). 10.3390/rs12061024.

22.

Dhillon MS, et al. Evaluation of MODIS, Landsat 8 and Sentinel-2 Data for Accurate Crop Yield Predictions: A Case Study Using STARFM NDVI in Bavaria, Germany. Remote Sens. 2023;15(7). 10.3390/rs15071830.

23.

Muruganantham P, Wibowo S, Grandhi S, Samrat NH, Islam N. A Systematic Literature Review on Crop Yield Prediction with Deep Learning and Remote Sensing. Remote Sens. 2022;14(9). 10.3390/rs14091990.

24.

Ram BG, Oduor P, Igathinathane C, Howatt K, Sun X. A systematic review of hyperspectral imaging in precision agriculture: Analysis of its current state and future prospects. Comput Electron Agric. 2024;222:109037. https://doi.org/10.1016/j.compag.2024.109037.

25.

Saha S, Kucher OD, Utkina AO, Rebouh NY. Precision agriculture for improving crop yield predictions: a literature review, Front. Agron., vol. 7, no. July, pp. 1–11, 2025, 10.3389/fagro.2025.1566201

26.

Giovos R, Tassopoulos D, Kalivas D, Lougkos N, Priovolou A. Remote sensing vegetation indices in viticulture: A critical review. Agric. 2021;11(5). 10.3390/agriculture11050457.

27.

Wang Q, Moreno-Martínez Á, Muñoz-Marí J, Campos-Taberner M, Camps-Valls G. Estimation of vegetation traits with kernel NDVI. ISPRS J Photogramm Remote Sens. 2023;195:408–17. https://doi.org/10.1016/j.isprsjprs.2022.12.019.

28.

Vera-Esmeraldas A, Pizarro-Oteíza S, Labbé M, Rojo F, Salazar F. UAV-Based Spectral and Thermal Indices in Precision Viticulture: A Review of NDVI, NDRE, SAVI, GNDVI, and CWSI, Agronomy, vol. 15, no. 11, pp. 1–31, 2025, 10.3390/agronomy15112569

29.

Fuentes I, et al. The normalised difference vegetation index as an analytic tool for wheat crop yield prediction: A review and meta-analysis. Precis Agric. 2025;26(4):55. 10.1007/s11119-025-10247-z.

30.

Sarvakar K, Thakkar M. Different Vegetation Indices Measurement Using Computer Vision. In: Chouhan SS, Singh UP, Jain S, editors. in Applications of Computer Vision and Drone Technology in Agriculture 4.0. Singapore: Springer Nature Singapore; 2024. pp. 133–63. 10.1007/978-981-99-8684-2_9.

31.

Peng D, et al. Assessing spectral indices to estimate the fraction of photosynthetically active radiation absorbed by the vegetation canopy. Int J Remote Sens. 2018;39(22):8022–40. 10.1080/01431161.2018.1479795.

32.

Barbosa BDS, e Silva Ferraz GA, Costa L, Ampatzidis Y, Vijayakumar V, Santos LM. UAV-based coffee yield prediction utilizing feature selection and deep learning, Smart Agric. Technol., vol. 1, p. 100010, 2021, https://doi.org/10.1016/j.atech.2021.100010

33.

Bali N, Singla A. Emerging Trends in Machine Learning to Predict Crop Yield and Study Its Influential Factors: A Survey. Arch Comput Methods Eng. 2022;29(1):95–112. 10.1007/s11831-021-09569-8.

34.

Panigrahi B, Kathala KCR, Sujatha M. A Machine Learning-Based Comparative Approach to Predict the Crop Yield Using Supervised Learning With Regression Models. Procedia Comput Sci. 2023;218:2684–93. https://doi.org/10.1016/j.procs.2023.01.241.

35.

Haseeb M, Tahir Z, Mahmood SA, Tariq A. Winter wheat yield prediction using linear and nonlinear machine learning algorithms based on climatological and remote sensing data. Inf Process Agric. 2025;12(4):431–44. https://doi.org/10.1016/j.inpa.2025.02.004.

36.

Dayil JL, Akande O, El A, Mahmoud D, Kimera R, Akande O. Department of Advanced Convergence, Handong Global University, 558 Handong-ro, Heunghae-eup, Buk-gu, Pohang 37554, Republic of Korea Department of Computer Science and Electrical Engineering, Handong Global University, 558 Handong-ro, Heunghae-eu, pp. 1–57.

37.

Bouguettaya A, Zarzour H, Kechida A, Taberkit AM. A survey on deep learning-based identification of plant and crop diseases from UAV-based aerial images, Cluster Comput., vol. 26, no. 2, pp. 1297–1317, Apr. 2023, 10.1007/S10586-022-03627-X

38.

Yoosefzadeh-Najafabadi M, Xavier A, Eskandari M, Hesami M. Machine learning after a decade: is it still a missing keystone in genomic-based plant breeding? Artif Intell Rev. 2025;58(9):260. 10.1007/s10462-025-11274-y.

39.

Zhang N, Zhou X, Kang M, Hu BG, Heuvelink E, Marcelis LFM. Machine learning versus crop growth models: an ally, not a rival. AoB Plants. 2023;15(2):1–7. 10.1093/aobpla/plac061.

40.

Binetti MS, Massarelli C, Uricchio VF. Machine Learning in Geosciences: A Review of Complex Environmental Monitoring Applications. Mach Learn Knowl Extr. 2024;6(2):1263–80. 10.3390/make6020059.

41.

Morales A, Villalobos FJ. Using machine learning for crop yield prediction in the past or the future, Front. Plant Sci., vol. 14, no. March, pp. 1–13, 2023, 10.3389/fpls.2023.1128388

42.

Maya Gopal PS, Bhargavi R. Performance Evaluation of Best Feature Subsets for Crop Yield Prediction Using Machine Learning Algorithms. Appl Artif Intell. 2019;33(7):621–42. 10.1080/08839514.2019.1592343.

43.

Imtiaz F. Integration of machine learning and remote sensing data for evaluating climate extremes in the agricultural production system of Prince Edward Island, no. March, 2025, [Online]. Available: https://islandscholar.ca/islandora/object/17755

44.

Reddy DJ, Kumar MR, Crop Yield Prediction using Machine Learning Algorithm, in. 2021 5th International Conference on Intelligent Computing and Control Systems (ICICCS), 2021, pp. 1466–1470. 10.1109/ICICCS51141.2021.9432236

45.

Shawon SM, Ema FB, Mahi AK, Niha FL, Zubair HT. Crop yield prediction using machine learning: An extensive and systematic literature review. Smart Agric Technol. 2025;10:100718. https://doi.org/10.1016/j.atech.2024.100718.

46.

Badshah A, Yousef Alkazemi B, Din F, Zamli KZ, Haris M. Crop Classification and Yield Prediction Using Robust Machine Learning Models for Agricultural Sustainability. IEEE Access. 2024;12:162799–813. 10.1109/ACCESS.2024.3486653.

47.

Radočaj D, Plaščak I, Jurišić M. Phenology-Based Maize and Soybean Yield Potential Prediction Using Machine Learning and Sentinel-2 Imagery Time-Series. Appl Sci. 2025;15(13). 10.3390/app15137216.

48.

Babaeian E, Sadeghi M, Jones SB, Montzka C, Vereecken H, Tuller M. Ground, Proximal, and Satellite Remote Sensing of Soil Moisture. Rev Geophys. 2019;57(2):530–616. 10.1029/2018RG000618.

49.

Rivera-Caicedo JP, Verrelst J, Muñoz-Marí J, Camps-Valls G, Moreno J. Hyperspectral dimensionality reduction for biophysical variable statistical retrieval. ISPRS J Photogramm Remote Sens. 2017;132:88–101. https://doi.org/10.1016/j.isprsjprs.2017.08.012.

50.

Miller L, Pelletier C, Webb GI. Deep Learning for Satellite Image Time-Series Analysis: A review. IEEE Geosci Remote Sens Mag. 2024;12(3):81–124. 10.1109/MGRS.2024.3393010.

51.

Mobarakeh ZM, Pourmanafi S, Ahmadi M. Employing sentinel-2 time-series and noisy data quality control enhance crop classification in arid environments: A comparison of machine learning and deep learning methods. Int J Appl Earth Obs Geoinf. 2025;142:104678. https://doi.org/10.1016/j.jag.2025.104678.

52.

Guo J, Wang S, Peng J, Liu J. Enhancing Land Surface Temperature Reconstruction: An Improved Interpolation of Mean Anomalies Based on the Digital Elevation Model (DEM-IMA). IEEE J Sel Top Appl Earth Obs Remote Sens. 2024;17:7371–85. 10.1109/JSTARS.2024.3378711.

53.

Czerkawski M, et al. Deep Internal Learning for Inpainting of Cloud-Affected Regions in Satellite Imagery. Remote Sens. 2022;14:1–18. 10.3390/rs14061342.

54.

Al-Jamimi HA. Synergistic Feature Engineering and Ensemble Learning for Early Chronic Disease Prediction. IEEE Access. 2024;12:62215–33. 10.1109/ACCESS.2024.3395512.

55.

Abimannan S, El-Alfy E-SM, Chang Y-S, Hussain S, Shukla S, Satheesh D. Ensemble Multifeatured Deep Learning Models and Applications: A Survey. IEEE Access. 2023;11:107194–217. 10.1109/ACCESS.2023.3320042.

56.

Sharma P, Dadheech P, Aneja N, Aneja S. Predicting Agriculture Yields Based on Machine Learning Using Regression and Deep Learning. IEEE Access. 2023;11:111255–64. 10.1109/ACCESS.2023.3321861.

57.

Lal A, Sharan A, Sharma K, Ram A, Roy DK, Datta B. Scrutinizing different predictive modeling validation methodologies and data-partitioning strategies: new insights using groundwater modeling case study. Environ Monit Assess. 2024;196(7). 10.1007/s10661-024-12794-w.

58.

Croci M, Ragazzi M, Grassi A, Impollonia G, Amaducci S. Assessing the temporal transferability of machine learning models for predicting processing pea yield and quality using Sentinel-2 and ERA5-land data. Smart Agric Technol. 2025;12:101207. https://doi.org/10.1016/j.atech.2025.101207.

59.

Chicco D, Warrens MJ, Jurman G. The coefficient of determination R-squared is more informative than SMAPE, MAE, MAPE, MSE and RMSE in regression analysis evaluation. PeerJ Comput Sci. 2021;7:1–24. 10.7717/PEERJ-CS.623.

60.

Hodson TO. Root-mean-square error (RMSE) or mean absolute error (MAE): when to use them or not. Geosci Model Dev. 2022;15:5481–7. 10.5194/gmd-15-5481-2022.

61.

Phiri M, Mulenga M, Zimba A, Eke CI. Deep learning techniques for solar tracking systems: A systematic literature review, research challenges, and open research directions. Sol Energy. 2023;262:111803. 10.1016/j.solener.2023.111803.

62.

Menon AS, Sankaran AJR, P K. Deep learning based farm-level crop yield prediction using multi-temporal satellite data for complex engineering application. Smart Agric Technol. 2025;12:101562. https://doi.org/10.1016/j.atech.2025.101562.

63.

Nevavuori P, Narra N, Lipping T. Crop yield prediction with deep convolutional neural networks. Comput Electron Agric. 2019;163:104859. https://doi.org/10.1016/j.compag.2019.104859.

64.

Bouanani NE, et al. Estimating Soil Attributes for Yield Gap Reduction in Africa Using Hyperspectral Remote Sensing Data with Artificial Intelligence Methods: An Extensive Review and Synthesis. Remote Sens. 2025;17(9):1–47. 10.3390/rs17091597.

65.

Khaki S, Wang L, Archontoulis SV. A CNN-RNN Framework for Crop Yield Prediction. Front Plant Sci. 2020;10:1–14. 10.3389/fpls.2019.01750.

66.

Park SH, et al. Development of a Soil Moisture Prediction Model Based on Recurrent Neural Network Long Short-Term Memory (RNN-LSTM) in Soybean Cultivation. Sensors. 2023;23(4). 10.3390/s23041976.

67.

Tian H, Wang P, Tansey K, Zhang J, Zhang S, Li H, China PR. Agric Meteorol. 2021;310:108629. https://doi.org/10.1016/j.agrformet.2021.108629.

68.

Nejad SMM, Abbasi-Moghadam D, Sharifi A, Farmonov N, Amankulova K, Lászlź M. Multispectral Crop Yield Prediction Using 3D-Convolutional Neural Networks and Attention Convolutional LSTM Approaches. IEEE J Sel Top Appl Earth Obs Remote Sens. 2023;16:254–66. 10.1109/JSTARS.2022.3223423.

69.

Gavahi K, Abbaszadeh P, Moradkhani H. DeepYield: A combined convolutional neural network with long short-term memory for crop yield forecasting. Expert Syst Appl. 2021;184:115511. https://doi.org/10.1016/j.eswa.2021.115511.

70.

Samadzadegan F, Toosi A, Dadrass Javan F. A critical review on multi-sensor and multi-platform remote sensing data fusion approaches: current status and prospects. Int J Remote Sens. 2025;46(3):1327–402. 10.1080/01431161.2024.2429784.

71.

Saki M, Keshavarz R, Franklin D, Abolhasan M, Lipman J, Shariati N. A Data-Driven Review of Remote Sensing-Based Data Fusion in Precision Agriculture From Foundational to Transformer-Based Techniques, IEEE Access, vol. 13, no. September, pp. 166188–166209, 2025, 10.1109/ACCESS.2025.3610649

72.

Jiao X, McNairn H, Yekkehkhany B, Dingle Robertson L, Ihuoma S. Integrating Sentinel-1 SAR and Sentinel-2 optical imagery with a crop structure dynamics model to track crop condition. Int J Remote Sens. 2022;43(17):6509–37. 10.1080/01431161.2022.2142077.

73.

Mesas-Carrascosa FJ, Arosemena-Jované JT, Cantón-Martínez S, Pérez-Porras F, Torres-Sánchez J. Enhancing Crop Yield Estimation in Spinach Crops Using Synthetic Aperture Radar-Derived Normalized Difference Vegetation Index: A Sentinel-1 and Sentinel-2 Fusion Approach. Remote Sens. 2025;17(8):1–16. 10.3390/rs17081412.

74.

Roßberg T, Schmitt M. A Globally Applicable Method for NDVI Estimation from Sentinel-1 SAR Backscatter Using a Deep Neural Network and the SEN12TP Dataset. PFG - J Photogramm Remote Sens Geoinf Sci. 2023;91(3):171–88. 10.1007/s41064-023-00238-y.

75.

Ramzan Z, Asif HMS, Yousuf I, Shahbaz M. A Multimodal Data Fusion and Deep Neural Networks Based Technique for Tea Yield Estimation in Pakistan Using Satellite Imagery. IEEE Access. 2023;11:42578–94. 10.1109/ACCESS.2023.3271410.

76.

Munaganuri RK, Rao YN. IEEE Access. 2024;12:52815–36. 10.1109/ACCESS.2024.3386552. PAMICRM: Improving Precision Agriculture Through Multimodal Image Analysis for Crop Water Requirement Estimation Using Multidomain Remote Sensing Data Samples.

77.

Mia MS, et al. Multimodal Deep Learning for Rice Yield Prediction Using UAV-Based Multispectral Imagery and Weather Data. Remote Sens. 2023;15(10). 10.3390/rs15102511.

78.

Li Y, Xiao X. Sensors. 2025;25(16):1–34. 10.3390/s25164991. Deep Learning-Based Fusion of Optical, Radar, and LiDAR Data for Advancing Land Monitoring.

79.

Chen J, Yu T, Cherney JH, Zhang Z. New Insights into Satellite-Based Forage Crop Monitoring. Remote Sens. 2024;16(5). 10.3390/rs16050734. Optimal Integration of Optical and SAR Data for Improving Alfalfa Yield and Quality Traits Prediction.

80.

Page MJ, et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ. 2021;372. 10.1136/bmj.n71.

81.

Prasad AK, Singh RP, Tare V, Kafatos M. Use of vegetation index and meteorological parameters for the prediction of crop yield in India. Int J Remote Sens. 2007;28(23):5207–35. 10.1080/01431160601105843.

82.

Fernandez-Ordonez YM, Soria-Ruiz J. Maize crop yield estimation with remote sensing and empirical models, Int. Geosci. Remote Sens. Symp., pp. 3035–3038, 2017, 10.1109/IGARSS.2017.8127638

83.

Kumari M, Suman S, Prasad D. Crop Yield Prediction using Remote Sensing: A Review, 2024, 10.1109/iccica60014.2024.10584591

84.

Son N-T, et al. Machine learning approaches for rice crop yield predictions using time-series satellite data in Taiwan. Int J Remote Sens. 2020;41(20):7868–88. 10.1080/01431161.2020.1766148.

85.

Kim J, Lee JK, Lee KM, Accurate Image Super-Resolution Using Very Deep Convolutional Networks, in. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 1646–1654. 10.1109/CVPR.2016.182

86.

You J, Li X, Low M, Lobell D, Ermon S. Deep Gaussian process for crop yield prediction based on remote sensing data, 31st AAAI Conf. Artif. Intell. AAAI 2017, pp. 4559–4565, 2017, 10.1609/aaai.v31i1.11172

87.

Khallikkunaisa S, Joshi S, Teja VK, S. R. E. S., and, Tharun R. Comprehensive Analysis of Crop Yield Prediction Using Deep Learning and Remote Sensing Techniques, 2025, 10.20944/preprints202506.0462.v1

88.

Capisonda-Porteza EJ. A Systematic Literature Review on Crop Yield Prediction with Deep Learning and Remote Sensing, Remote Sens., vol. 14, no. 9, p. 1990, 2022. 10.3390/rs14091990

89.

Meghraoui K, Sebari I, Bensiali S, El Kadi KA. Improving Yield Prediction at Field Scale by Exploring Temporal and Spectral Dependencies in High-Resolution Remotely Sensed Data using At-LSTM and R-PCA, in 2024 10th International Conference on Control, Decision and Information Technologies (CoDIT), 2024, pp. 1364–1368. 10.1109/CoDIT62066.2024.10708219

90.

Pei J. Leveraging Big Remote Sensing Data for Accurate Large-scale Crop Yield Predictions, 2025, 10.5194/egusphere-egu25-14490

91.

Mena F, et al. Adaptive fusion of multi-modal remote sensing data for optimal sub-field crop yield prediction. Remote Sens Environ. 2025;318. 10.1016/j.rse.2024.114547.

92.

Abrougui K, Gabsi K, Mercatoris B, Khemis C, Amami R, Chehaibi S. Prediction of organic potato yield using tillage systems and soil properties by artificial neural network (ANN) and multiple linear regressions (MLR). Soil Tillage Res. 2019;190:202–8. https://doi.org/10.1016/j.still.2019.01.011.

93.

Soltanolkotabi M, Javanmard A, Lee JD. Theoretical Insights Into the Optimization Landscape of Over-Parameterized Shallow Neural Networks. IEEE Trans Inf Theory. 2019;65(2):742–69. 10.1109/TIT.2018.2854560.

94.

Mena F, et al. Adaptive fusion of multi-modal remote sensing data for optimal sub-field crop yield prediction. Remote Sens Environ. 2025;318:114547. https://doi.org/10.1016/j.rse.2024.114547.

95.

Samuel Naik B et al. Sep., Enhanced Tobacco Yield Prediction Using Spatial Information and Exogenous Variable-driven Machine Learning Models, J. Sci. Res. Reports, vol. 30, no. 9, pp. 733–749, 2024, [Online]. Available: https://hal.science/hal-05066559

96.

Johnson MD, Hsieh WW, Cannon AJ, Davidson A, Bédard F. Crop yield forecasting on the Canadian Prairies by remotely sensed vegetation indices and machine learning methods. Agric Meteorol. 2016;218–9. https://doi.org/10.1016/j.agrformet.2015.11.003.

97.

Kavipriya J, Vadivu G. Exploring Crop Yield Prediction with Remote Sensing Imagery and AI, in 2024 International Conference on Advances in Computing, Communication and Applied Informatics (ACCAI), 2024, pp. 1–5. 10.1109/ACCAI61061.2024.10601854

98.

Salami ZA, Babamuratov B, Ayyappan V, Prabakaran N, Khudayberganov K, Singh CV. Modelling Crop Yield Prediction with Random Forest and Remote Sensing Data. Nat Eng Sci. 2025;10(2):67–78. 10.28978/nesciences.1763843.

99.

Li F, Miao Y, Chen X, Sun Z, Stueve K, Yuan F. In-Season Prediction of Corn Grain Yield through PlanetScope and Sentinel-2 Images. Agronomy. 2022;12(12):1–18. 10.3390/agronomy12123176.

100.

Jiang C, Guo X, Li Y, Lai N, Peng L, Geng Q. Multimodal Deep Learning Models in Precision Agriculture: Cotton Yield Prediction Based on Unmanned Aerial Vehicle Imagery and Meteorological Data. Agronomy. 2025;15(5):1–19. 10.3390/agronomy15051217.

101.

Lu J, et al. Deep Learning for Multi-Source Data-Driven Crop Yield Prediction in Northeast China. Agric. 2024;14(6). 10.3390/agriculture14060794.

102.

Bhokarkar Vaidya R, Ambedkar B, Kale Babasaheb Ambedkar K, Vaidya R, Nalavade D, Kale K. Hyperspectral Imagery for Crop yield estimation in Precision Agriculture using Machine Learning Approaches: A review, Researchgate.Net, vol. 9, no. 9, pp. 777–789, 2021, [Online]. Available: https://www.researchgate.net/profile/Renuka-Vaidya/publication/359699400_Hyperspectral_Imagery_for_Crop_yield_estimation_in_Precision_Agriculture_using_Machine_Learning_Approaches_A_review/links/62492b0a21077329f2ed62cb/Hyperspectral-Imagery-for-Crop-yiel

103.

Ashfaq M, Khan I, Shah D, Ali S, Tahir M. Predicting wheat yield using deep learning and multi-source environmental data. Sci Rep. 2025;15(1):1–20. 10.1038/s41598-025-11780-7.

104.

Schwalbert RA, Amado T, Corassa G, Pott LP, Prasad PVV, Ciampitti IA. Satellite-based soybean yield forecast: Integrating machine learning and weather data for improving crop yield prediction in southern Brazil. Agric Meteorol. 2020;284:107886. https://doi.org/10.1016/j.agrformet.2019.107886.

105.

Karongo J, Mwaniki JI, Ndiritu J, Mokaya V. Sorghum yield prediction based on remote sensing and machine learning in conflict affected South Sudan. Sci Rep. 2025;15(1):1–16. 10.1038/s41598-025-89030-z.

106.

Trentin C, Ampatzidis Y, Lacerda C, Shiratsuchi L. Tree crop yield estimation and prediction using remote sensing and machine learning: A systematic review. Smart Agric Technol. 2024;9:100556. https://doi.org/10.1016/j.atech.2024.100556.

107.

Haddaway NR, Page MJ, Pritchard CC, McGuinness LA. PRISMA2020: An R package and Shiny app for producing PRISMA 2020-compliant flow diagrams, with interactivity for optimised digital transparency and Open Synthesis, Campbell Syst. Rev., vol. 18, no. 2, p. e1230, Jun. 2022. https://doi.org/10.1002/cl2.1230

Yes