By Aïshael Donata Laury PICARD
Executive summary
Faced with the ravages caused by diseases in Haitian orchards, digital agriculture today offers promising solutions. As part of the Deep Farm project, an intelligent system for automatic detection of mango diseases was developed, relying on networks of convolutional neurons for computer vision. This device quickly identifies seven leaf diseases and four fruit diseases, with precision rates of 97 % on leaves and 75 % on fruits. Integrated into a multi-agent system accessible via smartphone, this decision support tool aims to reduce crop losses by allowing farmers to intervene early. Beyond mango, this modular approach provides a solid basis for the phytosanitary monitoring of other tropical crops, paving the way for a more resilient and productive Haitian agriculture in the face of climate and economic challenges.
Introduction
Global agriculture faces unprecedented challenges: feeding a population of 9.7 billion by 2050, while reducing the environmental footprint of a responsible sector by about one third of global greenhouse gas emissions [1]. In Haiti, these global challenges are coupled with acute local constraints: increased climate vulnerability, recurrent exposure to hurricanes and droughts, digital divide limiting access to modern technologies, and considerable post-harvest losses related to the lack of early detection of plant diseases.
Mango, a strategic culture for food security and the Haitian economy, illustrates this problem perfectly. Anthracnose, a dreadful fungal disease, can destroy up to 60% of crops in the absence of rapid intervention. However, small producers rarely have immediate access to the phytosanitary expertise needed to identify and treat these diseases in time. How, in this context, can artificial intelligence be used by farmers to transform the detection of mango diseases into a fast, reliable and accessible process?
This is precisely the objective of the work carried out within the framework of the Deep Farm project, an international consortium of twelve institutions including the École Supérieure des Technologies Industrielles Avancées (ESTIA) and the École Supérieure d'Informatronics d'Haïti (ESIH). The four-month course was designed to develop computer vision models capable of automatically distinguishing healthy cases from pathological cases on images of leaves and mango fruit, and then integrating these models into an intelligent decision-making system for farmers, students and agricultural technicians.
Methodology: a modular and robust approach
The developed system is based on a modular architecture composed of three specialized convolutive neuron network (CNN) models, each trained for a specific task:
1. Smart filtering model makes it possible to distinguish the relevant images (leaves and mango fruit) from images outside context (other plants, animals, objects). Two approaches were explored: a first version using a single CNN with confidence threshold, and a second, more robust version based on ResNet-50 pre-entered on ImageNet, explicitly incorporating a third class « Non-mango » to improve the ability to reject inappropriate images [2].
2. A leaf model identifies eight classes of pathologies or states: seven diseases (anthracnosis, bacterial canker, weaning wean, blight of the branches, cecidomyia, odium, fumigine) and one class « healthy leaf ». After unsuccessful attempts with EfficientNetB0, a custom designed CNN architecture was developed, gradually stacking convolution blocks (16, 32, 64, then 100 filters) with batch normalization and dropout regularization to ensure robustness and generalization.
3. A Fruity Model treats five classes: four diseases (anthracnose, alternariasis, black rot, peduncle rot) and one class « healthy fruit ». This architecture follows the logic of the leaf model while integrating an additional convolutional layer to better capture the visual variability of fruits (texture, shine, discoloration).
The data sets were compiled from public sources (Kaggle), in the absence of open source Haitian corpus. In order to overcome the lack of diversity and imbalances between classes, several techniques have been implemented, including increasing data (rotations, symmetries, zooms) [3], as well as the use of a loss function of the Focal Loss type [4] and class weight during training. Experimental monitoring was carried out via MLflow, allowing the traceability of experiments, hyperparameters and metrics, and ensuring the reproducibility of the results.
The entire system was exposed via a REST API developed with FastAPI. A containerization with Docker has been put in place to facilitate deployment and integration into the Deep Farm project's multi-agent RAG (Retrival- Augmented Generation). A complete suite of unit and integration tests was developed with Pytest, complemented by functional validations of model predictions, to ensure the robustness of the device.
Promising results despite constraints
The overall performance of the test data is satisfactory. The foliar model achieves an impressive overall accuracy of 97%, with a perfect classification (100%) for certain pathologies such as cutting weaving and burning of the branches. The few confusions observed concern mainly cecidomyia, sometimes confused with bacterial canker, as well as a slight confusion between the — pathologies whose visual symptoms may be similar even for an expert human eye.
The fruit model has an overall accuracy of 75%. The filter model, on the other hand, has balanced F1 scores: 0.92 for fruit, 0.98 for leaves, and 0.90 for off-domain images, demonstrating effective discrimination capacity and a robust rejection mechanism for irrelevant content.
These performances were confirmed during experimental tests on new images, especially on local Haitian varieties such as mango « Mrs Francisque », absent from training games. The system correctly identified the health status of the fruit, demonstrating an encouraging capacity for generalization despite the lack of local data in the initial corpus.
The response time observed under local conditions remains below 10 seconds per image, an essential criterion to ensure the practical usability of the system in the field [5]. The modular architecture developed, representing more than 1000 lines of Python code divided into a dozen modules, offers flexibility and maintenance, allowing to adjust or replace each component independently.
Conclusion and prospects
This work demonstrates the technical feasibility of an intelligent system for automatic detection of mango diseases adapted to the constraints of a developing country. By combining computer vision, deep learning and modular architecture, the developed device is a tool to help concrete decision-making to improve agricultural productivity and reduce crop losses in Haiti.
Several areas for improvement are emerging for the future.
Enrichment of the system by adding local data collected directly in Haitian orchards would enable it to refine its ability to recognize the visual specificities of diseases in the local context, particularly in terms of lighting conditions, cultivated varieties and phenological stages.
Furthermore, optimising existing models, with a view to even more efficient on-board deployment on limited-resource devices, is an important lever for improving the usability of the system in the field. Other avenues, such as the exploration of Vision Transformers (ViT), could also be considered as complementary [6].
Finally, the extension of this modular approach to other tropical crops, such as bananas, citrus fruits, rice farming or cocoa, would be fully consistent with the Deep Farm project.
Beyond technical aspects, this project illustrates the need to democratize access to plant protection expertise in rural areas, where farmers often lack the means and time to consult a specialist. By transforming a smartphone into an instant diagnostic tool, artificial intelligence can help bridge the digital divide and support the transition to a more resilient, productive and sustainable Haitian agriculture [7].
References
[1] World Bank Group. (2024). Climate-Smart Agriculture: Building Resilience to Climate Change. Washington, DC: World Bank Publications.
[2] Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). ImageNet Classification with Deep Convolutional Neural Networks. Advances in Neural Information Processing Systems (NeurIPS), 25, 1097-1105.
[3] Zhao, X., Wang, L., Zhang, Y., Han, X., Deveci, M., & Parmar, M. (2024). A review of convolutional neural networks in computer vision. Artificial Intelligence Review, 57. Article 99.
[4] Lin, T.-Y., Goyal, P., Girshick, R., He, K., & Dollár, P. (2017). Focal Loss for Dense Object Detection. Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2980-2988.
[5] Sanga, S.A., Machuve, D., & Jomanga, K. (2020). Mobile-Based Deep Learning Models for Banana Diseases Detection. Engineering, Technology & Applied Science Research, 10(3), 5674-5677.
[6] Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., ... & Houlsby, N. (2021). An Image is Worth 16×16 Words: Transformers for Image Recognition at Scale. International Conference on Learning Representations (ICLR).
[7] Huyen, C. (2022). Designing Machine Learning Systems. Sebastopol, CA: O-Reilly Media.
Aïshael Donata Laury PICARD
Higher School of Computer Science in Haiti
ESIH
aisha.picard@esih.edu
Picard (2025). When artificial intelligence comes to the rescue of Haitian mangoes
