Explainable plant disease classification and severity assessment using MobileNetV3-large for crop health monitoring
Year: 2026
Type: article
Abstract: Plant pests and diseases are estimated to cause losses of up to 40% of global crop production, while plant diseases alone are associated with economic losses exceeding US$220 billion annually 1 . Although deep learning-based plant disease classification methods have achieved high accuracy on benchmark datasets, many existing systems focus primarily on disease identification and provide limited information regarding visual explanation and disease burden. In this study, an integrated framework for plant disease classification, visual explanation, and severity estimation is developed using a MobileNetV3-Large backbone on a 38-class PlantVillage-derived benchmark. A three-phase progressive transfer-learning strategy is employed, combining Focal Loss with label smoothing, automated class-balanced weighting, AdamW optimization, and cosine-annealing learning-rate scheduling with linear warm-up. Grad-CAM++ is incorporated to provide spatially localized visual explanations of disease predictions, particularly for images containing multiple symptomatic regions. In addition, an EPPO-informed three-tier operational severity scheme is implemented using HSV-derived pseudo-labels to estimate visible disease burden. On the 38-class PlantVillage-derived benchmark, the proposed framework achieved 98.50% test accuracy, with a macro-F1 score of 0.984 and macro-precision of 0.985, demonstrating competitive classification performance rather than state-of-the-art accuracy. Across five independent random seeds, where applicable, classification performance was 98.49 ± 0.12% accuracy and 98.43 ± 0.12% macro-F1. The severity module achieved 89.1% agreement with the HSV-derived pseudo-labels in the original experiment. These severity results should be interpreted as agreement with an automated reference rather than independent validation against expert agronomic annotations. The proposed framework integrates classification, confidence estimation, visual explanation, and operational severity estimation into a unified inference pipeline, while recognizing the limitations of controlled PlantVillage-derived imagery and the need for future validation on independently collected field data.
Source: Scientific Reports
Institution: Manipal University Jaipur
Language: English
Topic: Smart Agriculture and AI
Subfield: Plant Science
Domain: Life Sciences
SDG: Zero hunger
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7221029517