Enhancing debris flow susceptibility models using negative samples from hazard-bearing body data

Building a reliable negative sample dataset is crucial for machine learning-based debris flow susceptibility mapping (DFSM). This task is particularly challenging in regions where surveyed and unsurveyed areas are not clearly defined. To partially address this issue, considering that watersheds containing hazard-bearing bodies are more likely to be surveyed, two new negative sample selection methods were proposed that independently utilized road data and building roof data. Using Ya’an City in southern China as the study area, the two proposed methods were compared with a random selection method and a pre-assessment-based method using four machine learning algorithms. The analysis revealed that the pre-assessment-based method yielded the highest area under the receiver operating characteristic curve (AUC) (> 0.92). However, the newly proposed methods produced higher relative susceptibility contrasts among three groups of debris flow watersheds categorized according to the availability of source materials. These methods also demonstrated more coherent DFSM results in two specific regions characterized by distinct geological and topographic conditions. Given the limitations of AUC in evaluating model performance across different sample datasets, the two new methods were considered superior. The findings are expected to provide a valuable reference for enhancing DFSM reliability in regions where unsurveyed areas are not clearly defined.