An Explainable Software-Based Framework for Pre-Symptomatic Grape Disease Risk Assessment Using MobileNetV2, Grad-CAM, and Weather Data

Viticulture faces severe yield losses from major fungal diseases such as Downy Mildew, Powdery Mildew, and Black Rot. Existing deep learning classification models predominantly rely on visible symptomatic lesions on leaves or fruit clusters. However, waiting for visual symptoms to manifest often delays chemical mitigation beyond the optimal therapeutic window, resulting in irreversible crop degradation. This paper proposes a software-based, explainable framework for pre-symptomatic grape disease risk assessment. The architecture combines finegrained leaf image feature extraction using MobileNetV2 with micro-climate weather telemetry (temperature, relative humidity, leaf wetness duration, and rainfall forecast). To ensure clinical transparency for agronomists and farmers, Explainable AI (XAI) is integrated via Gradientweighted Class Activation Mapping (Grad-CAM), generating visual saliency heatmaps that highlight sub-visual cellular stress indicators and necrotic margins. A dynamic Epidemiological Micro-Climate Risk Index (EMRI) fuses deep learning probabilities with real-time weather analytics. Evaluated on multi-source datasets including PlantVillage and field-captured vineyard imagery, MobileNetV2 achieves a classification accuracy of 97.84%. Based on a comprehensive literature review, a critical research gap is identified: current agricultural systems treat image classification and micro-climate forecasting as disjoint workflows without explainable decision support. The proposed framework fills this gap through an integrated web and mobile platform featuring automated risk alerts and actionable advisory generation