Deep Learning for Mango Disease: Climate-Aware Detection, Fruit Screening, and Few-Shot Recognition from the Field

Abstract

Mango is India's most valuable horticultural crop, yet annual losses from fungal and bacterial diseases exceed INR 15,000 crore. Existing machine learning tools diagnose visible symptoms from photographs but cannot warn farmers before symptoms appear. This paper presents three deep learning models that together form a complete mango disease management system. The first model, ClimaMango, combines a mango leaf photograph with 14 days of real weather data from six UP and Bihar districts to predict disease seven to fifteen days before symptoms appear. On the MangoLeafBD benchmark of 4,000 leaf images across 8 disease classes, it achieves 99.67% test accuracy, which is 75% fewer errors than using images alone. The second model screens mango fruit photographs taken by a standard smartphone camera, achieving 72.44% accuracy across five disease classes on the MangoDHDS dataset with no specialist equipment. The third model, MangoShot, uses few-shot meta-learning to recognise a completely new disease from just five field photographs, achieving 99.89% accuracy. All three models run on a consumer-grade GPU costing approximately INR 30,000 and are designed for smartphone deployment at zero cost to the farmer.