An Explainable and Severity-Aware Deep Learning Framework for Plant Leaf Disease Diagnosis with Organic Management Recommendations

Abstract

Timely plant disease management requires reliable disease recognition, an estimate of disease severity, interpretable visual evidence, and actionable agronomic guidance. This paper presents an explainable and severity-aware decision-support framework that integrates these functions within a unified workflow. A MobileNetV2-based classifier recognizes 15 bell-pepper, potato, and tomato disease/healthy classes from PlantVillage images, while a crop-gated severity branch evaluates potato and tomato using six categorical Leaf Disease Severity (LDS) levels. Grad-CAM provides class-discriminative visual explanations, and the predicted disease serves as a deterministic key for retrieving organic-management recommendations from a curated JSON knowledge base. The outputs are integrated through a Streamlit interface. The disease module records 94.62% accuracy, 94.31% macro-precision, 93.97% macro-recall, and 94.08% macro-F1; the six-level severity module records 88.74% accuracy, 87.96% macro-precision, 87.51% macro-recall, and 87.68% macro-F1. Comparative analysis is presented against ResNet50, EfficientNetB0, and a standalone MobileNetV2 baseline. The framework emphasizes lightweight inference, explicit separation of disease and severity supervision, qualitative explainability, and traceable knowledge retrieval