A dust-resistant adaptive CNN for edge-based date palm disease detection in Arid climates: application to Ha’il’s agricultural sector

Abstract

This study introduces a dust-resistant, edge-deployable Convolutional Neural Network (CNN) tailored for the automated diagnosis of date palm leaf diseases in arid agricultural environments, with a specific focus on Ha’il, Saudi Arabia. Addressing the persistent challenge of dust interference that frequently degrades visual quality in field diagnostics, the proposed Dust-Resistant Adaptive Network (DRA-Net) architecture combines a dust-adaptive preprocessing layer, a lightweight disease-specific attention mechanism, and an EfficientNet backbone to tackle the dual challenge of visual degradation due to airborne particulate matter and the subtle morphological cues of plant pathologies. The model is trained on a diverse, field-sourced dataset encompassing physiological deficiencies, fungal infections, pest infestations, and healthy leaf samples, with synthetic dust augmentation applied to simulate real-world interference. Designed for low-latency operation on edge devices, the proposed pipeline integrates dust-aware preprocessing, pathology-focused attention, and an edge-optimized training strategy to ensure robust deployment in resource-limited settings. DRA-Net demonstrates strong generalization across pathology classes and maintains high diagnostic fidelity under dust-prone conditions. Its performance surpasses conventional CNN baselines, achieving an average precision of 98.7%, recall of 98.8%, F1-score of 98.6%, and mAP@0.5 of 98.7%. These contributions collectively underscore DRA-Net’s effectiveness as a practical solution for early, reliable disease detection in environmentally challenging agricultural landscapes.