Integration of Fuzzy Logic and Neural Networks for Explainable Early Diagnosis of Rice Plant Diseases

Abstract

Early diagnosis of rice leaf diseases remains challenging due to subtle symptom manifestation, uncontrolled illumination, heterogeneous backgrounds, and the limited interpretability of purely data-driven models. This study proposes an explainable hybrid framework integrating a Mamdani Fuzzy Inference System (FIS) with an Artificial Neural Network (ANN) for early rice leaf disease diagnosis under real-field conditions. The framework combines engineered symptom descriptors extracted from segmented leaf regions (GLCM texture and HSV color features), acquisition-time environmental measurements, and a fuzzy-derived disease severity cue to mitigate symptom ambiguity while preserving rule-based interpretability. Experiments were conducted on 8,000 field-acquired rice leaf images collected from multiple locations, covering Healthy, bacterial leaf blight, brown spot, and leaf smut classes. Evaluation followed a leakage-controlled, location-disjoint protocol. Across five independent runs, the proposed FIS–ANN achieved an average accuracy of 91.3 ± 0.6% and a macro-F1 score of 90.8 ± 0.7%, significantly outperforming a feature-based ANN and a fine-tuned ResNet-18 baseline (paired McNemar test, p < 0.05). Per-class analysis shows consistent recall improvements for visually overlapping diseases, and additional evaluation on mild-severity samples confirms maintained sensitivity at early disease stages. Field deployment experiments using smartphone-acquired images from unseen locations further demonstrate robust generalization with low on-device inference latency. These results indicate that integrating fuzzy severity reasoning into a lightweight neural classifier provides a practical balance between performance, interpretability, and computational efficiency, supporting early disease screening and mobile decision-support applications in precision agriculture.