A 1.5 m resolution land cover map of Southern China created with Superpixel U-Net

Abstract

Accurate land cover mapping is essential for supporting ecological protection and resource management. However, in regions with complex and fragmented landscapes, such as Southern China (SC), existing medium- and high-resolution land cover products exhibit limited accuracy. To address this challenge, we developed Superpixel_U-Net, a multi-head segmentation algorithm that combines semantic segmentation and superpixel segmentation, to generate a 1.5 m land cover map for SC (the SCLC map). Towards this purpose, we acquired 6.1 TB very-high-resolution (VHR) ESRI World imagery and manually annotated 43,000 VHR images into eight land cover classes. Vegetation and water indices derived from Sentinel-2 imagery were used as auxiliary inputs to enrich spectral features. The resulting SCLC map achieved Pixel Accuracy (PA) values of 88.83% and 90.66% in the pixel-level and patch-level validations respectively, and showed a strong agreement with official survey reports (R² = 0.88). This dataset provides a solid foundation for further analyses that require accurate delineation of land cover types in SC.

Affiliated Institutions

• Qinghai University CN
• Sun Yat-sen University CN
• Peking University CN

Related Publications