A novel super-resolution urban green space segmentation network generating 0.2m resolution urban green space results using low-resolution imagery

Accurate mapping of urban green space (UGS) is crucial for urban planning and ecology. High-resolution drone images are often used for UGS identification, but their high cost limits large-scale mapping. To address this, we propose SR-UGSnet, a super-resolution segmentation framework that reconstructs low-resolution remote sensing images using spatial redundancy from high-resolution data. The framework combines CNN and Transformer-based semantic segmentation to refine these images, improving segmentation from 0.8 m to 0.2 m resolution. SR-UGSnet shows robustness against mislabeled training data and achieves an overall accuracy (OA) of 83.21%-83.65% and IOU scores of 62.38%-64.23%. Our approach outperforms other greenfield mapping products, demonstrating the potential of using low-resolution imagery for high-resolution UGS segmentation.