The LRB, located between latitudes 22°S-26°S and longitudes 26°E-35°E, is the fourth-largest transboundary river basin in Southern Africa, draining 416,300 km² (Fig. 1). It is shared by Botswana (19%), Zimbabwe (15%), South Africa (45%), and Mozambique (21%). The LRB hosts 18 million people, predominantly in rural areas and dependent on rainfed agriculture, which is projected to grow to 20 million by 2040 (Dickens et al. 2020). In 2007, South Africa held the largest share of the Limpopo River Basin’s population with about 10.7 million people, while Botswana, though smaller in absolute terms with about 1.2 million people, had a striking concentration, with nearly 69% of its national population residing within the basin (Limpopo Basin Permanent Technical Committee 2010). Economic activities span mining, irrigated agriculture, tourism, energy, forestry, and manufacturing, with irrigated farming the second-largest employer after mining (Sitoe and Qwist-Hoffman 2013). With an average elevation of ~ 840 m, the LRB spans diverse climatic, topographic, and ecological zones. The LRB exhibits strong spatial variability, transitioning from arid conditions in the west with annual rainfall around 200 mm to semi-arid and wetter temperate climates towards the central and eastern regions receiving up to 1,500 mm annually (Limpopo Basin Permanent Technical Committee 2010; Sitoe and Qwist-Hoffman 2013). The mean annual precipitation is approximately 530 mm, with isolated sub-humid pockets in the central basin. Rainfall follows a distinct east–west and north–south gradient, increasing towards the Mozambique coastal plains, and inner parts of South Africa. The wet season extends from October to April, peaking in December through February (> 100 mm/month), while the dry season spans May to September, with minimal rainfall (< 20 mm/month) (Fig. 2). Temperature extremes are notable, with summer daytime temperatures frequently exceeding 40°C, while winter nights can fall to 0°C with profound implications for agricultural activities and associated livelihoods.

Shallow sandy soils dominate the northern parts of LRB, moderately deep sandy to sandy clay loams characterise the south, and deeper sandy soils are prevalent in western and eastern zones. Floodplains in Mozambique have loamy and clay soils, though overall soil water retention remains low. The LRB’s geology comprises significant formations such as the Limpopo Mobile Belt, Kalahari Craton, Karoo System, and Bushveld Igneous Complex (Limpopo Basin Permanent Technical Committee 2010).

This study presents a fully automated, semi-supervised approach for mapping irrigated areas in the LRB. The workflow consists of five main steps:

The methodology was applied for each dry-season month from 2019 to 2024. All data processing and analysis were performed using the Google Earth Engine (GEE) cloud computing platform. An interactive web application that implements the methodology is available at: https://huggingface.co/spaces/IWMIHQ/irrigated-agriculture-extractor. This application allows for greater customisation of the methodology and facilitates its replication in other basins with similar environmental conditions.

A field validation campaign was conducted from 30 September to 5 October 2024 in Limpopo Province, South Africa, to evaluate the accuracy of irrigated area maps. Due to logistical constraints, field validation focused primarily on South Africa, which may limit the generalizability of findings to other LRB countries, including Botswana, Mozambique, and Zimbabwe. Data were collected from 120 agricultural plots using the KoboToolbox mobile data collection platform. The dataset included details on irrigation presence and type (e.g., center-pivot, drip, canal, sprinkler), crop type (e.g., maize, potatoes, citrus, avocado, cabbage), and water source (e.g., groundwater, rivers, canals).

Additionally, ten semi-structured interviews with local farmers provided qualitative insights into irrigation practices. The results revealed that commercial farms predominantly used center-pivot irrigation systems, while smallholder farmers employed a variety of irrigation methods. At the time of the survey, most crops were in the growing or maturity stages, with groundwater identified as the primary irrigation water source.

Figure 5 presents representative photographs from field observations of agricultural land. Images (a) and (c) depict non-irrigated crops, with (a) showing sparse, dry trees on reddish soil and (c) displaying a fenced area with minimal vegetation. In contrast, images (b) and (d) illustrate irrigated crops, with (b) showcasing a lush green field and (d) highlighting well-developed cabbage rows with visible irrigation tubing.

To supplement the in-situ validation data, an additional set of 70 reference samples was derived through visual interpretation of a high-resolution PlanetScope composite (4.77 m spatial resolution, four spectral bands) accessed via Google Earth Engine (GEE). These samples were in areas overlapping with known irrigation schemes, including both smallholder and commercial systems, thereby ensuring that the training and validation datasets reflected the full diversity of irrigation practices in the Limpopo River Basin. Of the PlanetScope-derived samples, 38 were classified as non-irrigated and 32 as irrigated, which, when combined with field observations, yielded a balanced dataset of 98 non-irrigated and 92 irrigated reference points. This integration of field-based data, scheme-level knowledge, and high-resolution imagery strengthened the representativeness of the reference dataset and reduced the likelihood of sampling bias in model training and accuracy assessment.

The locations of the surveyed irrigated fields within Limpopo Province are also shown in Fig. 6.

Water use by irrigated agriculture was quantified using the FAO Water Productivity Open-access Portal, version 3 (WaPOR v3) dataset (FAO 2019). We employed the Level-2 monthly Actual Evapotranspiration and Interception (AETI) product, available at 100 m spatial resolution for the African continent. To integrate with the 10 m irrigated area map, the AETI data were resampled using the nearest-neighbour method, which preserved the original pixel values and thus the integrity of the WaPOR measurements (Lillesand et al. 2015).

The classified irrigated area raster was then converted into vector polygons, and mean AETI values (mm month⁻¹) were extracted for each polygon using a zonal statistics procedure. WaPOR Level-2 AETI values are provided as digital numbers (DN) with a scale factor of 0.1, corresponding to millimetres per month. Conversion to water volume was carried out in two steps:

where is the polygon area in square metres. This ensured that evapotranspiration depth (mm) was consistently converted to volumetric water use (m³). Finally, total irrigation water use for the basin was obtained by aggregating volumes across all polygons.

The composite cropland mask for the LRB covered 4.37 million ha (9.5% of the basin), with South Africa accounting for 3.02 million ha (~ 70%). The remainder was in Botswana (0.59 million ha), Zimbabwe (0.53 million ha), and Mozambique (0.24 million ha) (Fig. 7). The mask, encompassing both rainfed and irrigated areas, formed the baseline for near real-time monthly irrigated area mapping.

Cropland distribution varied substantially among sub-basins (Table 1). The Middle Olifants (659,301.6 ha) and Upper Olifants (582,542.6 ha) together contained over 28% of total cropland, followed by Crocodile (443,986.9 ha), Sand (269,569.4 ha), Shashe (247,169.2 ha), and Notwane (207,940.1 ha). In contrast, the Lower Middle Limpopo (12,529.1 ha), Matlabas (20,954.3 ha), and Shingwedzi (21,533 ha) had minimal coverage.

Validation of the September 2024 irrigated area map using 198 field reference points yielded an overall accuracy of 80% and a Cohen’s Kappa of 0.60 (Fig. 8; Table 2). For the non-irrigated class, producer accuracy was 87.5% and user accuracy 68.5%, with an F-score of 82.4%. The irrigated class achieved a producer accuracy of 75.4% and user accuracy of 90.8% (F-score: 76.8%).

The confusion matrix showed 89/98 non-irrigated and 63/92 irrigated samples correctly classified. Misclassification was more common for irrigated plots (29 labelled non-irrigated) than for non-irrigated areas (9 cases). Many errors occurred in fields where crops were at early phenological stages with high bare-soil fractions, producing spectral confusion with fallow/dryland areas.

McNemar’s test confirmed a significant asymmetry in errors (χ² = 10.89, p < 0.001), indicating a tendency to under-predict irrigated areas.

Table 2 User, producer and overall accuracy for irrigated agriculture for the algorithm models

Figure 9 shows model predictions of irrigated agriculture (highlighted in green) overlaid on true-color composites from PlanetScope imagery. Each pair of panels compares the original satellite view (left) with the corresponding classification output (right). The predictions capture a range of irrigation patterns, including irregularly shaped plots, circular center-pivot systems, and rectangular fields, demonstrating the model’s ability to identify irrigated areas across diverse field geometries and scales.

The monthly distribution of irrigated areas during dry season (May–September) showed clear interannual variability (Fig. 10). The year 2019 recorded the largest areal extent (~ 221,000 ha), significantly higher than subsequent years (ANOVA, F = 8.72, p < 0.001). The lowest seasonal mean occurred in 2020 (~ 171,112 ha), with July and August minima of 116,000 ha and 120,000 ha, respectively. Tukey’s HSD identified 2020 as significantly different (p < 0.05) from 2019, 2022, and 2023. From 2021 to 2023, irrigated area extents partially recovered (from 179,000 to 191,000 ha). In 2024, early-season values exceeded 210,000 ha but declined later in the season. A linear regression of the irrigated areas from 2019 to 2024 showed a downward trend with an average annual decline of ~ 10,200 ha. The month of May showed significantly higher values than those for August or September (paired t-test, p < 0.01), confirming continuation of crops from the previous rainfed season.

Identified irrigated agriculture in the Middle Olifants, Crocodile, and Letaba sub-basins of LRB together accounted for the largest share of the basin’s total irrigated area (Fig. 11). An exception occurred in September 2023, when the Crocodile surpassed the Middle Olifants.

Sub-basins such as Lotsane, Changane, Lower Middle Limpopo, Bubi, Bonwapitse, and Mahalapswe consistently exhibited lower irrigated area extents from 2019. Global Moran’s I indicated significant positive spatial autocorrelation each year (I = 0.41–0.55, p < 0.01), with high-value clusters in the central/northeastern basin (e.g., Middle Olifants, Crocodile, Letaba) and low-value clusters in the west and south. These findings demonstrate that irrigated agriculture within the LRB is not randomly distributed but is spatially concentrated in specific sub-basins mainly in South Africa, likely reflecting the combined influence of hydrological availability, irrigation infrastructure, institutions and socio-economic drivers.

Dry-season irrigation water consumption peaked in May and September, with a pronounced minimum in June (Fig. 12). The highest single-month value was recorded in May 2019 (189.8 10⁶ m³), while the lowest occurred in June 2020 indicating transition from rainfed to irrigated season (46.2 10⁶ m³). August consistently exhibited high and stable consumption across all years. Mean dry-season use increased from approximately 103 10⁶ m³ in 2020 to 134 10⁶ m³ in 2024. Across months, May had the highest mean consumption (147.4 10⁶ m³), whereas June recorded both the lowest mean (83.7 10⁶ m³) and the highest interannual variability (CV = 33.6%). A one-way ANOVA indicated significant differences among months (F(4,25) = 8.54, p < 0.001). Post-hoc Tukey’s HSD revealed that August consumption was significantly higher than June (mean difference = 43.3 10⁶ m³, p = 0.026). These results confirm a strong seasonal pattern, with peaks at the start and end of the dry season, stable high usage in August, and a June minimum.

The results reveal a significant decline in irrigated areas, from a mean of 211,281 ha in 2019 to 184,771 ha in 2024. This trend aligns with findings from (Cai et al. 2017) and (Van Niekerk et al. 2018), who reported irrigated areas of 262,000 ha and 218,302 ha in the Limpopo Province, respectively. These estimates were based on coarser 30 m Landsat imagery which probably resulted in an overestimation due to mixed-pixel issues in heterogeneous landscapes (He et al. 2020; Gxokwe et al. 2022). The finer 10 m resolution of Sentinel-2 used in this study better captures smallholder plots, which are critical in the LRB’s fragmented agricultural systems. The observed decline may reflect multiple drivers, including increasing water stress, as reported by (Kapangaziwiri et al. 2021), who noted a doubling of water consumption from 2000 to 2012. The sharp drop in 2020 (171,112 ha) likely resulted from COVID-19 lockdowns across the LRB’s four countries, which disrupted agricultural activities and led to farm abandonment (Paganini et al. 2020). Persistent drought and water scarcity have likely contributed to the limited recovery of irrigated areas post-2020, with water restrictions and reduced reservoir levels disproportionately affecting smallholder farmers (Makhanya 2021; Ferreira et al. 2022; Mukwevho 2023; Mathivha et al. 2024).

Spatially, irrigation is concentrated in the Middle Olifants, Crocodile, Letaba, and Luvuvhu sub-basins, which consistently accounted for the largest irrigated extents, with significant positive spatial autocorrelation (Moran’s I = 0.41–0.55, p < 0.01). This clustering reflects hydrological availability, irrigation infrastructure, and socio-economic factors, as noted by (Sitoe and Qwist-Hoffman 2013). Conversely, sub-basins like Lotsane, Changane, and Lower Middle Limpopo show minimal irrigation, suggesting untapped potential for agricultural expansion.

Intra-annual trends show higher irrigation in May and June compared to August and September, likely due to declining water quotas as the dry season progresses (Mazibuko et al. 2021). Smallholder farmers, reliant on costly irrigation systems, often shift to rainfed agriculture, increasing their vulnerability to drought (Olabanji et al. 2021). The relatively small proportion of irrigated land (0.4% of the basin) indicates room for sustainable expansion, provided water resources are managed effectively.

Water consumption peaked in May (147.4 10⁶ m³) and September, with a minimum in June (83.7 10⁶ m³, CV = 33.6%), reflecting the LRB’s strong seasonality (Fig. 2). These patterns correspond to crop phenological stages, with May aligning with the maturity of summer crops and planting of winter crops, and September marking the growing stage of new crops, both requiring high water inputs (Al-Kaisi and Broner 2009; Chadalavada et al. 2021). The low water use in June, despite relatively high irrigated area, likely coincides with the planting phase of winter crops, which have lower ET demands (Sihlobo 2022). The increase in mean dry-season water use from 103 10⁶ m³ in 2020 to 134 10⁶ m³ in 2024, despite declining irrigated area, suggests intensifying demand, possibly due to drought-driven reliance on irrigation, as noted by (Botai et al. 2020), or improvements in irrigation efficiency, such as revitalized schemes and treated effluent use (Limpopo Basin Permanent Technical Committee 2010). The 2020 minimum likely reflects reduced agricultural activity during COVID-19 restrictions (Paganini et al. 2020).

The framework demonstrates strong operational potential by combining automation, scalability, and high spatial resolution. Leveraging Google Earth Engine enables basin-wide monthly mapping without the need for continuous in-season field data, while post-classification refinements reduce misclassification from perennial crops and steep terrain. The integration of WaPOR evapotranspiration adds an essential link between mapped extents and actual water use, enhancing its value for allocation planning.

Validation against independent field observations yielded an overall accuracy of ~ 80% (κ ≈ 0.60), a moderate level of agreement suitable for basin-scale operational monitoring. The model exhibits a conservative bias: irrigated fields are more likely to be undercounted, especially at early phenological stages with high bare-soil fractions, while non-irrigated areas are classified more reliably. This reduces the risk of overestimating irrigation but may lead to omission of marginal or early-season activity.

Despite these strengths, some limitations remain. Field validation was concentrated in South Africa’s Limpopo Province, which may restrict generalizability across the basin’s diverse agro-ecological settings. The reliance on resampled WaPOR data (100 m to 10 m) could introduce minor spatial inaccuracies, even though nearest-neighbor methods were applied to mitigate this. Furthermore, while the semi-supervised approach reduces dependency on costly ground surveys, some level of field data will always be required to benchmark results and calibrate models in new contexts.

Future improvements could involve integrating radar or thermal datasets to better capture irrigation signals under cloudy conditions or soil-dominated canopies and applying temporal deep learning approaches that exploit seasonal dynamics. Broader validation across Botswana, Mozambique, and Zimbabwe, supplemented by high-resolution commercial imagery and irrigation scheme inventories, would further strengthen accuracy and applicability. Overall, the high-resolution, semi-supervised framework marks a clear advancement over coarse-scale mapping, balancing automation, interpretability, and operational feasibility, while acknowledging the need for continued refinement and context-specific validation.

The high-resolution, monthly irrigated area maps and associated water use estimates produced by this framework provide an evidence base that was previously unavailable for the Limpopo River Basin. By linking spatially explicit irrigation extents with water consumption, the system enables managers to monitor irrigation dynamics throughout the dry season, identify peaks in demand, and detect shifts in cultivation patterns over time. This near real-time information supports more equitable allocation of scarce water resources, as authorities can see not only how much area is irrigated but also how much water is consumed in different sub-basins.

The capacity to capture smallholder fields at 10 m resolution is particularly valuable. These plots are often underrepresented in coarser-scale assessments, yet they play a central role in household food security and rural livelihoods. By including both small-holder and commercial irrigation in the same monitoring system, the methodology provides a more complete picture of agricultural water use and helps planners design interventions that are socially inclusive.

At the same time, the spatial distribution maps highlight irrigation hotspots, such as the Middle Olifants and Crocodile sub-basins, where water demand is concentrated and resource competition is most acute. This allows basin managers to target monitoring, efficiency improvements, or enforcement efforts where they are most needed. Conversely, the detection of underutilized areas offers opportunities for sustainable expansion, provided hydrological assessments confirm that water resources are available (Perrin et al. 2012).

In the longer term, embedding this monitoring system within the Limpopo Digital Twin initiative will ensure that outputs are directly accessible to decision-makers, farmers, and researchers. This integration provides a platform for evidence-based agricultural and water planning, enhancing resilience in a climatically vulnerable basin.