Enhancing landslide risk assessment through the optimization of anomaly detection techniques

The application of machine learning has been instrumental in the quantitative prediction of landslide risk, yet the accuracy and stability of the evaluation results have been undermined by anomalous samples. In this study, Xuanwei City is selected as the research area, where anomaly detection is conducted on the data samples through box plots, LOF, and K-means clustering algorithms, resulting in the identification of 11, 35, and 35 anomalous samples, respectively. After the anomalous data samples are excluded, the prediction results are analyzed through various evaluation metrics, including ROC curves, accuracy, and precision. It is demonstrated that the prediction accuracy of K-means clustering, box plots, and LOF, when combined with the LightGBM algorithm, reaches 87.04%, 87.11%, and 89.81%, respectively, with AUC values of 0.9330, 0.9420, and 0.9458. The use of LOF for anomaly detection is shown to yield the most notable improvements in both prediction accuracy and AUC values.