A Multi-Task Deep Learning Framework for Simultaneous Detection of Leaf Disease and Nutrient Deficiency in Oryza sativa

A Multi-Task Deep Learning Framework for Simultaneous Detection of Leaf Disease and Nutrient Deficiency in Oryza sativa

Authors

  • M. Kaliappan, S Vimal, Gaurav Dhiman, S. Pandiarajan

Keywords:

rice leaf disease detection, nutrient deficiency classification, multi-task learning, EfficientNet, transfer learning, precision agriculture, deep learning, convolutional neural networks.

Abstract

Rice (Oryza sativa) constitutes the dietary staple for more than 3.5 billion people worldwide, yet annual crop losses attributable to foliar diseases and macro-nutrient deficiencies consistently erode global food security. Conventional diagnostic practice demands trained agronomists performing field visits, a resource-intensive and delay-prone process ill-suited to the temporal dynamics of infection progression and nutritional stress. This paper presents a novel multi-task deep learning framework that simultaneously(AgroMTL-Rice) identifies leaf diseases and nutrient deficiencies from a single RGB leaf image, thereby enabling dual-diagnosis within a single forward pass. The proposed architecture couples a pre-trained EfficientNetB0 feature extractor with a symptom-disentanglement-inspired dual-branch head that separates texture-oriented disease semantics from colour-physiology-oriented nutrient semantics. Two publicly available Kaggle datasets, the Rice Leaf Diseases dataset (bacterial leaf blight, brown spot, leaf smut) and the Nutrient Deficiency Symptoms in Rice dataset includes nitrogen, phosphorus, and potassium deficiency are integrated under a unified multi-task label construction scheme, enabling cross-domain knowledge transfer between the two diagnostic tasks. Experimental evaluation on a stratified hold-out test set yields a weighted-average F1-score of 0.98 for the disease branch and 0.92 for the nutrient branch, with macro-average AUC values reaching 1.00 for all disease classes and 0.98–1.00 for nutrient classes. Fine-tuning of the top 40 EfficientNetB0 layers further sharpens discriminative boundaries. Ablation studies confirm that the shared-projection layer reduces nutrient classification error by 14.3% relative to single-task baselines. AgroMTL-Rice demonstrates competitive or superior performance compared to recent state-of-the-art single-task convolutional and transformer-based methods while requiring a single model deployment, substantially reducing inference overhead in resource-constrained agricultural IoT scenarios.

Downloads

Published

2026-09-24

Issue

Section

Articles

Citation Check

Loading...