Research on the Influence Mechanism of Multi-scale Terrain Automatic River Network Extraction Based on Change Point Detection Technology

Abstract

DEM-based river network extraction is sensitive to drainage thresholds, terrain and data resolution. This study applies four change-point detection (CPD) algorithms to three DEM resolutions (12.5, 30 and 90 m) across mountainous, hilly and plain landscapes (36 experiments). Kernel-based CPD performed best in most combinations, attaining the highest F1-scores in mountains at 90 m (0.89), hills at 30 m (0.88) and plains at 12.5 m (0.75). PELT and Binary Segmentation produced similar thresholds, whereas the Mann–Whitney U test yielded higher values; thresholds varied approximately linearly with DEM resolution. The results show that DEM resolution, terrain and algorithm jointly affect extraction accuracy. Coarser data can suppress false channels in mountains, whereas finer data preserve smaller channels in plains. These findings support terrain-specific selection of DEM resolution and CPD algorithm for automatic river network extraction.