Using statistical learning algorithms in regional landslide susceptibility zonation with limited landslide field data

Abstract

Regional Landslide Susceptibility Zonation (LSZ) is always challenged by the available amount of field data, especially in southwestern China where large mountainous areas and limited field information coincide. Statistical learning algorithms are believed to be superior to traditional statistical algorithms for their data adaptability. The aim of the paper is to evaluate how statistical learning algorithms perform on regional LSZ with limited field data. The focus is on three statistical learning algorithms, Logistic Regression (LR), Artificial Neural Networks (ANN) and Support Vector Machine (SVM). Hanzhong city, a landslide prone area in southwestern China is taken as a study case. Nine environmental factors are selected as inputs. The accuracies of the resulting LSZ maps are evaluated through landslide density analysis (LDA), receiver operating characteristic (ROC) curves and Kappa index statistics. The dependence of the algorithm on the size of field samples is examined by varying the sizes of the training set. The SVM has proven to be the most accurate and the most stable algorithm at small training set sizes and on all known landslide sizes. The accuracy of SVM shows a steadily increasing trend and reaches a high level at a small size of the training set, while accuracies of LR and ANN algorithms show distinct fluctuations. The geomorphological interpretations confirm the strength of SVM on all landslide sizes. Our results show that the strengths of SVM in generalization capability and model robustness make it an appropriate and efficient tool for regional LSZ with limited landslide field samples.