Abstract
In arid-zone agriculture, early detection of foliar diseases is critical to preventing catastrophic crop failure under severe environmental constraints. However, existing deep learning models often struggle to distinguish early-stage lesions due to extreme lighting, dust accumulation, and low contrast between damaged tissues and background foliage. To address these challenges, this study proposes a novel Dual-Stream Convolutional Neural Network with Coordinate Attention (DS-CA-Net) specifically optimized for arid-zone crops. The architecture comprises a dual-stream backbone: one stream captures global contextual representations and color variations, while the second stream extracts high-frequency structural details of micro-lesions using dilated convolutions. A coordinate attention mechanism is integrated into both streams to capture long-range spatial dependencies and preserve precise positional information. Evaluated on a newly curated dataset of arid-zone crop leaves, including date palm, pomegranate, and alfalfa, the proposed DS-CA-Net achieved an overall classification accuracy of 98.42% and an F1-score of 98.15%, significantly outperforming standard single-stream architectures such as ResNet-50 and MobileNetV3. Ablation studies confirm that the fusion of multi-scale spatial features and coordinate attention is highly effective in filtering out environmental noise like dust and glare. These findings demonstrate the potential of DS-CA-Net as a robust, edge-deployable diagnostic tool for precision agriculture in desert-fringe farming environments.
