Physics-informed optimization for a data-driven approach in landslide susceptibility evaluation

Abstract

Landslide susceptibility mapping is an integral part of geological hazard analysis. Recently, the emphasis of many studies has been on data-driven models, notably those derived from machine learning, owing to their aptitude for tackling complex non-linear problems. However, the prevailing models often disregard qualitative research, leading to limited interpretability and mistakes in extracting negative samples, i.e. inaccurate non-landslide samples. In this study, Scoops 3D (a three-dimensional slope stability analysis tool) was utilized to conduct a qualitative assessment of slope stability in the Yunyang section of the Three Gorges Reservoir area. The depth of the bedrock was predicted utilizing a Convolutional Neural Network (CNN), incorporating local boreholes and building on the insights from prior research. The Random Forest (RF) algorithm was subsequently used to execute a data-driven landslide susceptibility analysis. The proposed methodology demonstrated a notable increase of 29.25% in the evaluation metric, the area under the receiver operating characteristic curve (ROC-AUC), outperforming the prevailing benchmark model. Furthermore, the landslide susceptibility map generated by the proposed model demonstrated superior interpretability. This result not only validates the effectiveness of amalgamating mathematical and mechanistic insights for such analyses, but it also carries substantial academic and practical implications.

Keywords

• Physics-informed
• Machine learning
• Bedrock depth
• Scoops 3D
• Landslide susceptibility