Lightweight scalable deep learning framework for real time detection of potato leaf diseases

Potato leaf diseases, if left undetected, threaten food security in agricultural economies and cause substantial crop losses. To address this critical challenge, we developed an AI-based system called the Enhanced Single Shot Multibox Detector(EF-SSD), a variant of SSD that integrates multiscale feature fusion and Squeeze-and-Excitation attention to improve fine-grained lesion detection. The enhanced model processes high-resolution images (512$$\times$$512 pixels) and analyzes leaves at ten magnification levels, enabling it to identify even minor signs of infection. The inclusion of Squeeze-and-Excitation filters allows the system to focus more effectively on characteristic disease patterns, increasing detection precision. After scanning the leaves, the system applies advanced image processing techniques to localize disease regions and assess their severity. We evaluated EF-SSD using 2,500 labeled potato leaf images representing healthy plants and cases of early and late blight. The proposed model achieved a mean Average Precision (mAP) of 97% at 0.5 IoU, an F1-score of 95%, and an Intersection over Union (IoU) of 89%, outperforming advanced detectors such as YOLOv5, YOLOv8, RetinaNet, and Faster R-CNN across all metrics. It also delivers real-time inference at 47 FPS, confirming its suitability for on-field deployment. An ablation study further demonstrates the effectiveness of SE blocks and extended feature hierarchies in enhancing detection accuracy. These outcomes highlight EF-SSD’s potential as a reliable, efficient, and scalable tool for smart agriculture and early crop disease management.