ResNet101-FPN Based Deep Learning Framework for Multi-Class Plant Leaf Disease Classification

Abstract

Plant diseases significantly threaten global agricultural productivity and food security, causing annual crop yield losses of up to 30%. Early and accurate detection of plant diseases is essential for implementing timely intervention strategies and ensuring sustainable agricultural practices. Recent advances in deep learning have revolutionized automated plant disease detection using leaf images. This study proposes a robust deep learning framework based on ResNet101 integrated with a Feature Pyramid Network (FPN) for multi-class plant disease classification. The model is trained and evaluated using the comprehensive PlantVillage dataset containing 38 distinct plant disease classes across multiple crop species. The Feature Pyramid Network extracts multi-scale features from different layers of the backbone network, enabling the model to capture both fine-grained texture details and high-level semantic information essential for distinguishing visually similar disease symptoms. The proposed architecture employs transfer learning with ImageNet pre-trained weights and is optimized using the Adam optimizer with sparse categorical crossentropy loss. Experimental results demonstrate exceptional performance with macro-averaged precision of 99.45%, recall of 99.54%, and F1-score of 99.50% across 10,861 test samples. Comprehensive evaluation using confusion matrix analysis, Receiver Operating Characteristic (ROC) curves, and feature map visualization confirms the robustness and discriminative capability of the proposed approach. The results indicate that the ResNet101-FPN model provides a reliable, scalable, and deployable solution for automated plant disease diagnosis in precision agriculture systems.

Code and data availability

The paper uses the public PlantVillage dataset, but that is a third-party benchmark from prior work (Hughes and Salathé), not a paper-specific asset. No author code, trained models, data deposits, or availability statements are provided in the supplied blocks.

No evidence-backed public reproduction asset is currently recorded.