A cascaded parallel deep learning framework for precise potato canopy density mapping via edge-aware feature fusion

High-resolution mapping of canopy density for field crops such as potatoes serves as the crucial technical foundation for canopy coverage evaluation. However, the dense overlapping of potato canopies and the severe optical homogenization between senescent vegetation and soil backgrounds in complex farmland environments cause conventional semantic segmentation models to frequently suffer from blurred boundaries and micro-topological fragmentation. To overcome the limitations of mere module stacking in traditional methods, this study proposed a novel edge-aware and multi-scale interactive deep learning framework based on potato canopy images. This approach eliminated background ambiguity by introducing SIOX to guide color priors and deeply integrates a gated convolutional edge extraction mechanism with the bidirectional U-Net++ architecture. This integration enabled the iterative synergy of deep and shallow semantics within a multi-scale feature space to accomplish the detection of potato canopy density under complex backgrounds. Experiments demonstrated that the proposed model exhibits broad applicability and exceptional robustness in highly variable environments. Its mIOU and B-IOU reached 79.64% and 60.55% respectively, significantly outperforming state-of-the-art networks. Furthermore, the model achieved high prediction accuracy across various density gradients based on high-fidelity masks. Compared to baseline models such as DeepLabv3+, HRNet and SegFormer, the proposed model showed notable improvements in both mIOU and segmentation accuracy. Combined with the PyQt5 interactive system constructed in this study, the automated quantification and visualization of potato canopy coverage have been achieved, providing a potential methodological reference for intelligent potato vine killing equipment. The training program can be obtained from https://github.com/HT-JL/Potato-Research.git .