Efficient recognition of cone karst landforms through deep learning: insights from multi-source data fusion in southwest China

Abstract

Accurate mapping of cone karst hills, a typical landform in tropical-subtropical karst landscapes, is essential for understanding karst geomorphology and supporting environmental management. However, the complex topography of these areas poses significant challenges for conventional mapping approaches. This study presents an advanced deep learning approach for cone karst hill identification in China's Guizhou Karst Plateau, integrating multi-source data, including high-resolution Digital Elevation Models (DEM), DEM-derived data and spectral information with U-Net and DeepLab V3+ architecture. The results demonstrate that the U-Net model consistently outperformed DeepLab V3+, achieving higher accuracy and adaptability across all data configurations for cone karst hills recognition. The optimal configuration (DEM, slope and local relief) showed notably superior performance compared to other tested configurations. Rigorous validation in the Xinyi and Anlong regions of southwestern Guizhou confirmed the method's reliability and transferability. This research establishes a scalable and transferable processing workflow that enables high-precision, large-scale mapping of individual cone karst hills, with direct applications in geomorphological research, sustainable land management, and conservation planning for fragile karst ecosystems.