Semantic Segmentation of Small-Scale Urban Green Spaces Using LiDAR and Deep Learning

抄録

Urban green spaces (UGSs) are increasingly being recognized for their role in addressing biodiversity loss and climate change, as highlighted by the IPBES and IPCC. However, conventional remote sensing methods struggle to accurately detect small-scale UGS such as residential planting. This study aimed to develop a method for high-precision extraction of small-scale UGSs using airborne remote sensing data and deep learning. Focusing on Tama City, Tokyo, we created 8 datasets by combining RGB imagery with light detection and ranging (LiDAR)-derived features—normalized Digital Surface Model (nDSM), Number of Returns (NOR), and Intensity. These were input to semantic segmentation using SegNet to verify the accuracy of green space extraction. The results showed that incorporating nDSM, NOR, and Intensity significantly improved the extraction accuracy compared to the RGB-only input (p < 0.05). The combination of NOR and Intensity was particularly effective in detecting small-scale UGS. Conversely, the extraction accuracy decreased for green spaces with complex morphologies and low circularity. Our findings highlight the value of LiDAR features, especially NOR and Intensity, in the high-precision extraction of small UGSs and demonstrate the potential of open data for supporting future urban planning and environmental policy.