Edge AI for Real-Time Crop Disease Detection: A Lightweight Deep Learning Framework with Knowledge Distillation for Resource-Constrained Environments

Plant diseases threaten global food security, causing over 20% annual yield losses and approximately USD 220 billion in economic damage worldwide. While deep convolutional neural networks (CNNs) have shown remarkable efficacy in automated plant disease diagnosis, their deployment on resource-constrained edge devices is hindered by computational demands, memory footprints, and energy consumption. This study introduces AgriNet-Lite, a lightweight deep learning architecture synergistically integrating depthwise separable convolutions, an efficient channel attention mechanism, and knowledge distillation from a high-capacity teacher network. The framework was evaluated on the PlantVillage benchmark dataset (54,305 images, 38 disease classes, 14 crop species). AgriNet-Lite achieved 98.72% classification accuracy with macro-averaged precision of 0.986, recall of 0.984, and F1-score of 0.985, while maintaining a compact model size of 2.1 million parameters (8.1 MB). Edge deployment on Raspberry Pi 4 demonstrated inference latency of 28 ms per image (35.7 FPS). Ablation studies confirmed individual contributions of channel attention (+2.30% accuracy) and knowledge distillation (+1.70% accuracy), with cumulative improvement of 3.92 percentage points. Field validation under natural conditions using a custom-collected tomato leaf dataset (n = 1,247) achieved 94.2% accuracy despite substantial domain shift. Comparative analysis against MobileNetV2, EfficientNet-B0, Tiny-LiteNet, and CNN-SEEIB demonstrated superior accuracy-efficiency trade-offs. AgriNet-Lite establishes a pragmatic, deployable solution for on-device plant disease diagnosis, particularly suited for smallholder farmers in regions with limited computational infrastructure and intermittent internet connectivity.