Mangoes are the fifth most produced fruit in the world (by tonnage) behind bananas/plantains, citrus fruits, grapes and apples, with a production of 51 million tonnes in 2017 ¹. Burkina Faso contributes between 11% and 18% of mango production in West Africa ². Mangoes are a very important socio-economic and climate issue in Burkina Faso. Mango production and export contribute significantly to the country's income.

Indeed, the mango sector generates more than 15 billion in turnover per year¹. In addition, it contributes to food security in producer areas and vegetation cover, limiting the effects of climate change ². The mango value chain in Burkina Faso is a source of employment. Its activities employ 28,000 people in 21,000 jobs as entrepreneurs or similar (farmers, trackers, retailers, etc.) ², around 350 permanent jobs in processing and packaging companies for fresh exports, and around 7,000 seasonal jobs throughout the chain ³.

The mango sector faces enormous phytosanitary constraints, particularly insect pests ⁴ and also major diseases such as anthracnose, bacterial disease and dieback⁵⁶⁷. These diseases cause significant damage in orchards, such as deterioration in fruit quality, tree death and reduced yields⁸⁹. This damage causes enormous losses, such as loss of income for producers, management costs, reduced productivity, etc.¹⁰. Unfortunately, producers have little understanding of these diseases and their symptoms, epidemiological aspects and diagnostic difficulties. Added to this is the limited dissemination of available control technologies, resulting in low adoption by producers. Nowadays, artificial intelligence is a powerful decision-making tool in agriculture, particularly in the automatic diagnosis of crop diseases ¹¹. In India, automatic diagnosis and classification models for crop pests have been implemented ¹²¹³.

This aspect of AI for agricultural development is still largely unexplored in Burkina Faso, specifically for the diagnosis of mango tree diseases. The aim of this study is to design an effective AI-based automatic diagnosis model for the main mango tree diseases in Burkina Faso.

The images were pre-processed to improve their quality and usability. This mainly involved resizing them to dimensions compatible with YOLOv11, normalising colours and correcting brightness, and increasing the data (data augmentation) through rotation, inversion, cropping and noise to improve the robustness of the model. After pre-processing, batches of 3,000 images for each disease were selected, for a total of 9,000 images for all three diseases (anthracnose, bacterial disease and dieback). These selected images were imported into the Roboflow platform for annotation.

Image annotation consisted of drawing bounding boxes around the symptoms and associating them with a class corresponding to each disease (Fig. 2). This methodical work made it possible to build a systematically annotated dataset that will be used for training.

Supervised transfer learning with the pre-trained YOLOv11 algorithm was used. YOLOv11 (You Only Look Once, version 11) is a convolutional neural network (CNN) designed for real-time object detection and simultaneously enables :

The entire supervised training process was conducted via the Roboflow platform, which provides an integrated environment for managing datasets and training computer vision models. The annotated data was divided into three sets : 75% for training (training set), 20% for validation (validation set), and 5% for testing (test set). Regularisation techniques such as early stopping were applied to avoid overfitting. All data was made available on a platform for easy access.

Model performance was evaluated using standard object detection metrics such as:

Precision : the proportion of correct predictions among all predictions made. It indicates the model's ability to limit false detections.

Recall : this is the proportion of cases correctly detected in relation to all existing positive cases. It reflects the model's ability to not omit any symptoms that are present.

mAP@50 (mean Average Precision at 50% IoU): this is an overall indicator of object detection performance, calculated as the average precision across all classes, with an intersection over union (IoU) threshold set at 50%.

These metrics were used to evaluate the models' ability to correctly identify and locate symptoms of mango tree diseases, while ensuring a balance between accuracy and completeness.

Accurate identification of phytopathological symptoms is a fundamental prerequisite for the development of digital diagnostic tools, particularly those using AI. The symptoms documented in orchards largely corroborate the classic descriptions established by researchers. Indeed, for anthracnose, observations of irregular brown spots on leaves, sunken black lesions on fruit and necrosis of young twigs are consistent with the work of ¹⁷ and ¹⁸, which highlight similar symptoms of anthracnose in humid mango-producing areas, where relative humidity and poor aeration promote infection. The work of ¹⁹has also demonstrated the early appearance of symptoms as early as the flowering stage, directly influencing yield.

With regard to bacterial disease, the photographed symptoms captured, such as translucent oily spots, leaf perforations and exudates on fruit, are typical of infection by Xanthomonas campestris pv. mangiferaeindicae. These symptoms were described by ⁶ in Burkina Faso and West Africa, who highlighted the rapid progression of bacterial lesions. Recognition of these visual characteristics is crucial in order to differentiate this disease from anthracnose, whose lesions can sometimes overlap on the fruit. With regard to decline, symptomatic analysis reveals leaf wilting, progressive browning of twigs, and fruit abortion ²⁰. This description corresponds to the symptoms associated with complex fungal infections, particularly Lasiodiplodia spp, as reported by ²¹.

All of this symptomatic data thus constitutes a consistent database for training AI models, particularly through deep learning, to identify major diseases. This has been demonstrated in the work of ²² on the automatic identification of wheat diseases in Europe using AI algorithms. Also, based on the training of approximately 8,438 images of symptoms acquired on PlantVillage, ²³ have set up a system for the automatic detection of diseases in vineyards and mango trees in India. This will constitute a database of the symptoms of the main diseases affecting mango trees and, above all, a significant step forward in overcoming the lack of phytosanitary expertise and speeding up decision-making in the field by producers.

The performance achieved by the V2-3M, V3-3M and V4-3M models demonstrates differentiated capabilities in detecting the symptoms of major mango tree diseases from annotated images with accuracies greater than 50%. The V2-3M model, with an mAP@50 of 50.3%, appears to be the most effective, confirming the importance of good parameterisation and optimised architecture in automatic detection tasks. Its performance combines relatively balanced precision (55,1%) and recall (49,1%), indicating that it can be used as a starting point for a functional prototype of a mobile plant health diagnosis tool ²⁴. These metrics show a good compromise between the model's ability to correctly locate symptoms and its ability to minimise both false positives and false negatives.

On the other hand, the poor performance of the V3-3M model (mAP@50 = 23,5% ; recall = 26,7%) despite similar accuracy (56,8%) highlights limitations in the generalisation of learning. This model seems to struggle to capture symptomatic diversity, particularly due to reduced sensitivity to morphological and chromatic variations in symptoms.

The metrics of the V4-3M model, which are intermediate (mAP@50 = 46,8%) and reflect a certain robustness, albeit less marked than V2-3M, show that improved performance depends on a compromise between architectural complexity, annotation quality and training image diversity. This finding is consistent with the observations of ²⁵, who showed that training models on highly diversified image databases significantly improves the robustness of artificial intelligence tools applied to agriculture.

Furthermore, the satisfactory performance of the V2-3M model is in line with recent research into leaf disease detection using deep learning. For example, ²⁶ achieved a mAP of 51% on tomato diseases using CNN architectures. ¹² achieved 99% accuracy for a cashew anthracnose diagnosis model using tools such as Mobilenet and Inception. In terms of performance, the V2-3M model is the most effective. Its ability to balance accuracy (55,1%) and recall (49,1%), combined with an mAP@50 of 50,3%, makes it the most suitable for initial operational application. The V4-3M model, although fairly close, remains slightly less robust, while V3-3M should be ruled out due to its insufficient results. Thus, V2-3M provides a solid basis for an automated diagnostic system, with prospects for improvement.

These results confirm the feasibility of using artificial intelligence models for the automatic detection of major mango tree diseases, while highlighting the need to continue optimising algorithms and enriching training databases²⁷. The integration of such a tool into a mobile application could strengthen early detection capabilities and improve the responsiveness of phytosanitary interventions in Burkina Faso.

Anthracnose, bacterial disease and dieback are major phytosanitary constraints affecting mango trees. Unfortunately, producers have little knowledge of the epidemiological aspects of these diseases or the control techniques available. This study was conducted with the aim of designing an accurate automatic diagnostic model for the main diseases affecting mango trees using artificial intelligence. Training the collection of symptomatic images showed that the V2-3M model performed better in terms of metrics for automatic disease detection.

These relevant results open up new prospects for integrating AI tools into agroecological practices, particularly for phytosanitary monitoring and management in orchards. It would be wise to integrate treatment recommendations into the model, deploy it on a mobile application, and promote it for the benefit of producers, all of which will contribute to effective disease management in mango orchards.