Sub-Pixel Semantic Segmentation for Precision Agriculture: Identifying Micro-Defects in Crop Foliage

Abstract: Early detection of micro-defects in crop foliage—including sub-millimetre chlorotic lesions, fungal hyphae boundaries, and early-stage necrotic patches—remains one of the most challenging problems in computational precision agriculture. Conventional semantic segmentation models operate at the native pixel resolution of remotely-sensed or macro-lens imagery and consistently fail to recover boundary-level structural detail at scales below one pixel. This paper presents Sub-Pixel Segmentation Network (SPSNet), a novel deep-learning architecture that incorporates learned sub- pixel convolution layers, spectral-channel attention, and a multi-scale hierarchical decoder to segment foliage defects at resolutions exceeding the sensor's native sampling grid. Our model is trained and evaluated on FoliageDefect-22K, a purpose-built annotated dataset comprising 22,400 high-resolution images across six crop species and nine defect categories collected under controlled and field conditions. SPSNet achieves a mean Intersection-over-Union (mIoU) of 91.7%, surpassing the nearest competing method by 8.5 percentage points, while maintaining near-real-time inference at 18.3 frames per second on a single NVIDIA A100 GPU. Ablation studies confirm the individual contribution of each architectural component. These results establish sub-pixel segmentation as a viable and practical tool for deployment in precision agricultural monitoring systems.

Keywords: precision agriculture, semantic segmentation, sub-pixel convolution, foliage micro-defects, deep learning, crop health monitoring, spectral attention