A Multimodal, Multi-Task, and Uncertainty-Aware Deep Learning Framework for Rice Leaf Disease Detection and Severity Estimation (RICE-MuSTA)

Abstract

Background: Rice is a staple food crop for more than half of the world’s population and plays a central role in food security, particularly in Asian and African countries. However, rice productivity is severely affected by foliar diseases such as rice blast, bacterial blight, sheath blight, and brown spot. Delayed or inaccurate diagnosis of these diseases can lead to yield losses ranging from 10–40%, directly impacting farmers’ livelihoods. While recent advances in deep learning have enabled automated leaf disease detection, most existing approaches remain limited to image-only classification, lack severity assessment, ignore environmental uncertainty, and are unsuitable for real-time field deployment 
Methods: In this work, we introduce RICE-MuSTA, a framework designed to jointly address multimodality, severity estimation, and uncertainty in rice leaf disease monitoring. The framework integrates leaf image features extracted using convolutional neural networks (CNNs), global contextual representations learned via Vision Transformers (ViTs), and microclimate information encoded through a multilayer perceptron (MLP). 
Findings: The proposed model simultaneously performs disease classification and severity estimation while quantifying predictive uncertainty and providing lesion-level explainability using CAM-based heatmaps. To enable practical adoption, a teacher–student distillation and quantization strategy is employed, compressing the model into a lightweight architecture suitable for mobile and edge deployment.
Novelty and applications: Across multiple datasets and evaluation settings, we observed consistent improvements in robustness, interpretability, and deployment efficiency.