An End-to-End Multi-Task ResNet-FPN Network for Simultaneous Plant Type, Disease Type, and Severity Classification

Abstract

Accurate and rapid diagnosis of plant diseases is critical for food security and sustainable agriculture. This paper proposes a multi-task deep learning framework that simultaneously classifies plant species, disease type, and three severity levels (mild, medium, severe) from leaf images. The model employs a ResNet101 backbone enhanced with a Feature Pyramid Network (FPN) to extract multi-scale features, followed by three task-specific classification heads. Severity labels are automatically generated using a semantic segmentation approach based on adaptive thresholding in LAB and HSV color spaces, eliminating the need for manual annotation. To address class imbalance, the dataset is balanced via oversampling with on-the-fly data augmentation. Extensive experiments on the PlantVillage dataset demonstrate that the multi-task model achieves near‑perfect plant classification (macro F1 0.9995) and disease classification (macro F1 0.9978), and 99.1% macro F1 for severity estimation, outperforming single-task baselines while using 66% fewer parameters. Comprehensive evaluation including confusion matrices, per-class precision/recall/F1, ROC curves, and sample predictions are provided. The proposed framework offers an integrated solution for plant health monitoring with high accuracy and efficiency.