U-Net MobileNetV2 detects crop diseases with 98.2% accuracy in 14.7ms

Plant diseases cause 20–40% of global agricultural yield loss—over USD 220 billion annually. Kumar and Kar from arXiv have developed a unified deep learning pipeline that integrates semantic segmentation, regression-based severity estimation, and disease classification into a single framework. The system categorizes stress into four levels (Low to Very High) based on infected leaf area. They tested four architectures—U-Net (MobileNetV2), SegFormer, FCN, and PSPNet—on the Apple Tree Leaf Disease Segmentation dataset (1,641 samples, six classes).

U-Net with MobileNetV2 emerged as the clear winner: 98.20% pixel accuracy, 0.70 mean IoU, and 99.41% detection accuracy, all at 14.7ms per image—fast enough for real-time field deployment. SegFormer managed a competitive 0.66 mIoU, while FCN and PSPNet lagged at ~0.49 mIoU. More importantly, the computed stress index correlates strongly with human experts (r=0.968, R²=0.937), proving the system's reliability for automated decision support in precision agriculture. This enables farmers and agronomists to quantify disease severity quickly, objectively, and at scale.