Lightweight CNN Architectures for Real-Time Crop Disease Detection on Low-End Smartphones for Rural Deployment

Crop disease diagnosis remains a critical bottleneck for smallholder farmers, particularly in rural and semi-urban regions where agricultural extension services are limited and internet connectivity is unreliable.While convolutional neural networks (CNNs) have achieved high accuracy in plant disease classification, most existing research prioritizes accuracy alone, using large architectures such as ResNet and VGG that are impractical for deployment on low-end smartphones.This study addresses that gap by evaluating MobileNetV3-Small, EfficientNet-Lite0, and a custom pruned/quantized CNN against a ResNet50 baseline for multi-class crop disease classification using the PlantVillage dataset.Beyond accuracy and F1-score, models are benchmarked using model size, CPU-only inference latency, and memory footprint under conditions representative of budget Android devices with no GPU and offline operation.The best-performing lightweight model is further optimized using posttraining quantization, and its accuracy-efficiency tradeoff is analyzed before and after compression.The study proposes an accuracy-per-MB and accuracy-permillisecond framework to compare deployment feasibility across architectures.Since verified experimental measurements were not available during document preparation, numerical results remain to be experimentally determined.