Identifying Mango Leaf Diseases with Advanced Deep Learning Approaches and Convolutional Neural Networks - MangoMed AI

Overview

India produces 50% of the world's mangoes. Yet diseases like Anthracnose and Bacterial Canker routinely destroy 10–39% of yields because early detection requires expert eyes that most farmers don't have access to. MangoMedAI is an IEEE-published deep learning system that detects and classifies 8 mango leaf diseases with 98.97% accuracy and an F1 score of 99.10%, using a fine-tuned EfficientNet-B0 model trained on 12,046 leaf images. Built with FastAI and PyTorch, it outperforms multiple existing approaches in both accuracy and deployment efficiency.

About This Research

The problem isn't just agricultural, it's economic. Mango Bacterial Canker alone can cause 10–100% yield loss. Powdery mildew has damaged 23% of unsprayed mango trees globally. Anthracnose has caused losses of up to 39%. Farmers in the field can't diagnose these diseases with the naked eye, and getting an expert to every plantation isn't realistic at scale.

MangoMedAI addresses this with a computer vision system that can identify whether a mango leaf is healthy or infected with one of 8 diseases, from a single photograph.

The system was built by fine-tuning a pre-trained EfficientNet-B0 model on a dataset of 12,046 mango leaf images spanning 9 classes (1 healthy, 8 disease types), using the FastAI framework on top of PyTorch. The training pipeline used aggressive image augmentation, random flipping, rotation, zoom, and contrast enhancement, to build robustness against real-world variation in leaf photography.

Three architectures were benchmarked before selecting EfficientNet-B0:

— ResNet-18 was rejected early: 84% accuracy was not remotely good enough for a system that could be used in real agricultural decisions.

— GoogLeNet achieved ~99% accuracy but required 463 minutes per epoch, nearly 4x slower than EfficientNet's 128 minutes. For a system that needs to be deployable on mobile or edge devices, this training cost makes it impractical. It was rejected not because it failed, but because it couldn't scale.

— EfficientNet-B0 was selected: 98.97% accuracy, 99.10% F1 score, 128 minutes per epoch. The right tradeoff between performance and deployment viability.

Final metrics across 12,046 test images: — Accuracy: 98.97% — Precision: 99.05% — F1 Score: 99.10% — Error Rate: 0.40%

The confusion matrix showed 10,601 true positives and only 91 false negatives, meaning the system almost never misses a diseased leaf. For an agricultural detection tool, false negatives (missed diseases) are the dangerous failure mode, so this is the metric that matters most.

The model outperformed all comparable existing methods reviewed in literature, including AlexNet-based approaches (91.2% accuracy) and ResNet variants.

Published at the 2025 IEEE International Conference on Next Generation Communication & Information Processing (INCIP-2025). DOI: 10.1109/INCIP64058.2025.11020297