A mango-disease study in Burkina Faso combines a substantial local image collection with an external test on new trees. Published in Scientific Reports on 4 November, it reports 50.3% mean average precision for the leading detection model and 104 successful symptom detections among 115 external images at a 70% confidence threshold, about 90.4%. Both results contribute evidence, but they use different datasets and scoring rules. Understanding each helps explain what the study has established and how its proposed mobile tool could develop.[1]
The work begins from a substantial regional agricultural problem. Mango orchards face bacterial disease, anthracnose and branch dieback, and identifying problems early enough for appropriate action can be difficult without specialist support. The researchers visited 40 orchards and collected 11,001 images showing characteristic symptoms: 4,527 of anthracnose, 3,038 of bacterial disease and 3,436 of dieback. These are images of selected symptomatic material, not a random survey of every healthy and unhealthy tree in the region.[1]
How the model was trained
The team selected 3,000 images from each disease class, giving 9,000 balanced originals. It split those into 5,400 training, 2,700 validation and 900 test images before augmentation. Transformations were then applied only to the training set, expanding it to 10,800 images; validation and test images remained unaugmented. This produced 14,400 images in the final workflow. The ordering matters: keeping transformed versions of a training photograph out of the held-out sets avoids that direct route to an inflated test result. The models used pretrained YOLOv11 components.[1]
The principal reported performance results were V2-3M with mean average precision at intersection-over-union threshold 0.50, or mAP@50, of 50.3%; precision of 55.1%; recall of 49.1%; and F1 score of 51.9%. The Results section reports the other configurations at mAP@50 of 23.5% and 46.8%; the abstract gives the first of those as 23%. V2-3M leads on this reported metric, although the study's statistical comparisons did not establish significant differences among configurations on precision, recall or F1. Its figures still leave room for missed targets and false alarms.[1]
In object detection, mAP@50 is not simply the percentage of diseased trees diagnosed correctly. It evaluates model predictions across confidence settings and measures whether predicted boxes overlap with annotated symptom regions sufficiently. Precision asks what share of the model's positive detections are correct; recall asks what share of the annotated targets are found. Neither by itself tells a grower how many trees across a farm have the right diagnosis. An image can contain several lesions, confusing background objects and more than one potential target.[1]
Testing beyond the original orchards
The separate test used 115 images from additional trees in Bobo-Dioulasso that were outside the original 40 orchards. The authors report 104 correct symptom detections at a 70% confidence threshold, or approximately 90.4%. This is a useful external check on the selected model. It is also a smaller assessment with a different scoring procedure from the main mAP comparison, so it cannot replace that result or establish a 90.4% diagnostic success rate across all farms. The value lies in having tested the model beyond the original collection as well as on the held-out images.[1]
A further extension would carry the same locked model across more regions, seasons, smartphone cameras, light conditions and cultivars. Healthy leaves and visually similar non-disease damage would help establish how often an alert is unwarranted; ambiguous or mixed disease cases would test how the model behaves when there is no simple answer. The authors call for wider datasets and cross-regional validation. Those steps would build on the external orchard check already performed, rather than start external evaluation from scratch.[1]
The economic consequences depend on more than numerical classification. False reassurance could delay a treatment or inspection. False alarms could trigger unnecessary pesticide use or expense. An extension worker would need to know whether the model can abstain when uncertain and whether it recommends the correct management step under local conditions. A classification engine that names a disease cannot independently select the right treatment, pesticide dose or timing for every orchard.[1]
From symptom detection to a mobile tool
The authors propose integrating disease-management recommendations and deploying the model through a mobile application. That is a prospective development path. The publication does not establish that the app is already operating in Burkina Faso, that farmers have adopted it or that yields and incomes improved. Nor should the paper's use of the word 'accuracy' beside one of its model metrics be used to turn precision or mAP into a measured farmer-level diagnostic success rate.[1]
The study leaves a concrete resource and a development path: locally collected symptom images, a documented model comparison and an external check on new trees. A useful extension service could combine that detection capability with appropriate management guidance and referral when an image is ambiguous. The main detection metrics and the smaller external result answer different questions along that path. Preserving both gives researchers and prospective users a more useful account of progress than reducing the work to one accuracy headline.[1]
Sources & further reading
- Evaluation of state-of-the-art models for improving the diagnosis of mango tree diseases through supervised learning of symptom images in Burkina Faso Scientific Reports / Dianda and colleagues · 4 Nov 2026 · Accessed 4 Nov 2026