An Explainable AI Framework (XAI-CoffeeNet) with Deep Learning and Vision Transformers for Saudi Coffee Disease Recognition and Severity Assessment

This paper presents XAI-CoffeeNet, an explainable artificial intelligence framework with three core innovations for automated coffee disease management in Saudi Arabia's Jazan high-altitude plantations (1300–1400 m): (i) a hybrid CNN–Vision Transformer architecture with adaptive confidence-based feature fusion that dynamically balances local lesion features and global disease pattern recognition, addressing the limitation of single-architecture approaches that capture only local or global features independently; (ii) a dual LIME–SHAP explainability framework providing complementary local instance-level and global feature-level interpretability, validated against expert-annotated disease regions (localization IoU = 0.912), addressing the gap of single-method XAI systems that provide incomplete model transparency; and (iii) an integrated quantitative severity assessment module producing continuous disease progression scores correlated with trained agronomy annotations (r = 0.924), enabling precision treatment decisions beyond binary disease classification. Coffee farming in the Jazan region faces critical challenges from plant diseases that devastate crop yield in these isolated high-altitude plantations, where the scarcity of trained pathologists and inaccessible mountainous terrain make conventional expert monitoring impractical and economically unsustainable. The framework incorporates dual LIME and SHAP explainability mechanisms and is validated across 18 farms in the Jazan region under varying altitude and environmental conditions. XAI-CoffeeNet achieves 96.7% accuracy, 95.1% F1-score, and 0.989 AUC in multi-class detection of four disease categories (Healthy, Coffee Leaf Rust, Coffee Berry Disease, and Bacterial Blight). The severity assessment module yields a correlation of 0.924 with trained agronomy annotations. The dataset was split at the image level rather than farm level; reported metrics should therefore be interpreted as upper-bound estimates. In observational before–after field trials, system adoption was associated with a 37% reduction in pesticide usage and a 19% improvement in crop yield from 48.3 to 30.4 L/ha/season and 612 to 728 kg/ha respectively, across 18 farms over an 8-month observational period.