Context aware multimodal plant disease diagnosis using agro ecological metadata and explainable inference

Plant disease detection is a major aspect of precision agriculture, directly impacting crop productivity and food security. Recent deep learning approaches have achieved high accuracy in disease classification from images, but most rely heavily on visual features and remain ill-posed to general changes in illumination, background complexity, image quality, and field conditions. They also have limited agro-ecological information and insufficient interpretability, limiting their practical reliability and users’ trust. This study proposes to overcome these drawbacks by leveraging a multimodal deep learning framework, ContextLeafXAI, inspired by LeafContextReasonNet. The framework considers visual phenotypic aspects alongside agro-ecological information (weather, soil, and geographical information). The proposed reliability-aware gated fusion mechanism adaptively combines information from various sources, and the dual-level explainability strategy, using Grad-CAM and contextual feature attribution, guides lesion localisation and understanding of the role of the environment. Experimental results demonstrate that ContextLeafXAI attains 96.2% accuracy, 95.2% Macro-F1 score, and 0.98 ROC-AUC, which is superior to the unimodal and multimodal baselines evaluated. The framework also shows good cross-dataset generalisation, explanation consistency and moderate computational load. The results suggest that ContextLeafXAI offers a powerful and understandable decision-support solution for plant disease diagnostics across various agricultural scenarios. Code is available at https://github.com/zakeeya12/ContextLeafXAI_Full_Implementation . Not applicable.