Cross-Modal Federated Learning for Robust Plant Disease Classification

Abstract

The accuracy of automated plant disease diagnosis is frequently limited by the use of visual symptoms alone, especially when it comes to differentiating between conditions that have a lot of visual similarities. To address this, we propose a new privacy-preserving framework that combines the strengths of multi-modal federated learning (FL) with environmental context. Our system integrates leaf images with synthetic sensor data—such as temperature, humidity, and leaf wetness duration—capturing critical cues that influence disease progression. Actually, this system core is a dualbranch convolutional neural network designed to process both image and environmental features in a way that reflects the biological characteristics of different diseases. Results demonstrates that the multi-modal approach consistently outperforms conventional image-only models across multiple disease categories, and especially true for diseases where environmental factors are very important in how they develop. We further extend the system into a federated learning setting, allowing models to benefit from distributed training while keeping sensitive agricultural data local and private. This makes the framework not only more accurate but also practical for real-world use, where data privacy is essential.