A LIGHTWEIGHT DEEP LEARNING ARCHITECTURE FOR REAL-TIME CROP DISEASE DIAGNOSIS IN RESOURCE-CONSTRAINED RURAL INFRASTRUCTURES

Abstract

Agricultural productivity across sub-Saharan Africa continues to face severe constraints from crop diseases, which account for yield losses estimated at 20–40% annually in smallholder farming systems. Automated, timely disease identification presents a viable pathway to mitigating these losses, yet existing deep learning models demand computational resources that far exceed those available in rural field settings. This paper presents AgriEdge, a novel lightweight deep learning architecture designed specifically for real-time crop disease classification on resource-constrained edge hardware. AgriEdge introduces a modified one-dimensional convolutional attention mechanism (1D-CAM) that replaces the standard multi-head self-attention of Vision Transformers, reducing the parameter count by 72% relative to ViT-Small while retaining a macro F1-score of 98.5% on a benchmark agricultural image dataset comprising five disease classes across 16,873 annotated samples. The model, quantized to INT8 precision, achieves an on-device inference latency of 18.3 ms at a batch size of one on a Raspberry Pi 4B, a commercially available single-board computer affordable to smallholder cooperatives. Comparative evaluations against six established architectures, including ResNet-18, MobileNetV3-Small, EfficientNet-Lite0, and ViT-Small, confirm that AgriEdge occupies a superior position on the F1-score–parameter trade-off curve. The system is validated under realistic deployment conditions reflecting low-bandwidth, low-power rural settings. These findings support the broader agenda of inclusive agricultural digitization and economic revitalization in underserved communities across the African continent.

Keywords: edge computing, crop disease classification, lightweight deep learning, convolutional attention mechanism