Theory and Method of Geographic Information System
LIU Yang, CHU Yunzhi, ZHU Lei, HAO Xiangxia, FEI Taizheng, ZHAO Weidong, YUAN Xiaoyu
The gully edge line is a critical geomorphic feature that demarcates positive and negative terrains on the Loess Plateau.Its accurate and efficient extraction is fundamental for studying gully erosion and landscape evolution,and holds significant importance for guiding ecological restoration in small loess watersheds.To address the existing challenges of insufficient extraction accuracy,limited feature diversity,and the lack of a systematic evaluation framework,which hinder comprehensive and objective model assessment,this study employs machine learning models for positive/negative terrain segmentation and automatic gully edge line extraction in a small watershed of the loess tableland area.The key findings are as follows.(1)In the loess tableland area,the optimal feature subset selected by the random forest(RF)model comprises six features:elevation variation coefficient, B1,B11,B11_cont, B11_mean,and B9_mean.(2)A comparative analysis integrating classification accuracy for positive/negative terrains and gully edge line displacement showed that RF and BP neural network models achieved accuracies of 92.71% and 90.01% ,respectively.These represent improvements of 8.0% and 5.3% over the slope distortion neighborhood judgment method,with a corresponding gully edge offset of 32.67 m (within three raster pixels).This study demonstrates that the optimal feature subset selected by the RF model is well–suited for loess tableland topography,resulting in fewer misclassifications and omissions,thereby providing a scientific basis for the accurate extraction of gully edge lines.