Efficient and robust deep learning for field-level wheat disease classification on resource-constrained devices

• Deployment-first lightweight CNN for wheat disease detection • Adaptive Channel Reduction improves budget-matched accuracy • 95.67 • 90 • Robust to blur, noise, occlusion with uncertainty-aware inference Wheat is a staple crop in Ethiopia, yet its production is severely constrained by fungal diseases such as stem rust, yellow rust, Septoria, and fusarium head blight, which can cause yield losses and threaten food security. While deep learning has achieved strong performance in plant disease recognition, many models remain impractical for smallholder agricultural settings due to high computational requirements, reliance on cloud connectivity, and large memory footprints. This study presents a deployment-oriented lightweight convolutional neural network for wheat disease classification, designed for resource-constrained mobile environments. A field dataset was collected with the involvement of domain experts affiliated with the Ethiopian Institute of Agricultural Research (EIAR), capturing variability in illumination, occlusion, background clutter, and plant growth stages. The proposed architecture introduces Adaptive Channel Reduction (ACR), a stage-wise channel allocation strategy that preserves mid-level texture representations critical for distinguishing visually similar diseases, and integrates structured pruning with INT8 quantization for efficient on-device inference. Experimental results show that the compressed model achieves 95.67% test accuracy with only 1.16M parameters and a 3.12 MB model size, representing approximately 90% parameter reduction relative to ResNet-18. Compared with standard CNN architectures, the model delivers competitive predictive performance while substantially reducing computational cost, resulting in a favorable accuracy-efficiency trade-off. The INT8 model maintains fast CPU and mobile inference and remains robust to field perturbations, while calibration and selective prediction support reliable human-in-the-loop deployment. These results demonstrate that domain-aware lightweight architectures combined with structured compression enable accurate and practical AI-based wheat disease detection in resource-constrained agricultural environments.