Connection methods in landslide susceptibility assessment: Suitability evaluation based on environmental factor type

Connection methods are essential for integrating environmental factors with machine learning models for landslide susceptibility assessments. However, current research does not consider the different characteristics of continuity and discreteness within environmental factors and therefore does not analyze the suitability of various connection methods for different factor types. Moreover, the applicability of connection methods remains unclear when slope units are used as the basic assessment units. This study employed slope units as mapping units. The original data of 15 environmental factors, including 12 continuous and three discrete factors, and two connection methods, i.e., frequency ratio (FR) and modified FR (MFR), were separately used to construct the input datasets for landslide susceptibility modeling. The performance of four widely used machine learning models, random forest (RF), support vector machine (SVM), logistic regression (LR), and multilayer perceptron (MLP), was analyzed to evaluate the suitability of the connection methods for landslide susceptibility mapping. The results show that, in contrast to the decision tree-based RF model, the use of different connection methods for different factor types significantly influences the results of nontree models, including SVM, MLP, and LR. SVM model is the most sensitive to factor types and connection methods. When the MFR is used as the connection method, it improves the mapping results, especially for the SVM model. This shows that it is essential to consider the different characteristics of the data and select an appropriate environmental factor connection strategy to increase the effectiveness of landslide susceptibility evaluation. Furthermore, this study explored the role of connective methods from a sample distribution perspective, providing a theoretical foundation for the more rational and effective integration of environmental factors.