Effect of different mapping units, spatial resolutions, and machine learning algorithms on landslide susceptibility mapping at the township scale

Abstract

With the 1:10,000 geohazard hazard evaluation in Northwest China, the delineation of geohazard hazard zoning has begun to shift to the township scale. This study aims to reveal the impacts of three important uncertainty issues in landslide susceptibility prediction (LSP) at the township scale: mapping units, spatial resolutions, and selection of machine learning algorithms. Taking Chengguan town of Guanghe County, China as an example, the landslide inventory and 9 important conditioning factors were acquired. The normalized frequency ratios of each conditioning factor were calculated under the raster unit at seven resolutions (1, 5, 10, 15, 20, 25, and 30 m) and the slope unit. Four machine learning models [random forest (RF), multilayer perceptron (MLP), support vector machines (SVM), and naive Bayes (NB)] were applied for LSP modeling. The results indicate that slope unit-based models effectively differentiate high- and low-susceptibility zones. In contrast, for raster units, the AUC values of all models significantly improve as the cell size increases from 1 to 30 m, and the mean and standard deviation of the landslide susceptibility index accordingly decrease and increase, respectively. The LSP performance of the four machine learning models in the study region from high to low is RF, MLP, SVM, and NB. In addition, the overlay analysis of landslide susceptibility maps and historical landslides shows that the RF model based on 15 m resolution raster units can obtain the best landslide susceptibility map for the study area. These findings provide critical insights for optimizing landslide susceptibility assessments in township-scale applications, particularly within loess hill regions of Northwest China.