Uncertainties of landslide susceptibility prediction due to different spatial resolutions and different proportions of training and testing datasets

It is of great significance to explore the influences of the spatial resolution and the proportion of model training and testing dataset on the uncertainties of landslide susceptibility prediction(LSP). Taking the landslides in Ningdu County of Jiangxi Province as examples, the frequency ratios of various environmental factors under different spatial resolutions(15, 30, 60, 90 and 120 m) are firstly calculated. Then, the landslide and non-landslide samples are divided into different model training and testing datasets with the proportions of 9/1, 8/2, 7/3, 6/4 and 5/5, and the model input and output variables under 25 combined conditions are obtained. Furthermore, these input and output variables are imported into the Support Vector Machine(SVM) and Random Forest(RF) models to carry out LSP. Finally, the uncertainties of LSP modeling under the 25 combined conditions are discussed using the accuracy assessment as well as the distribution characteristics of landslide susceptibility indexes. The results show that the landslide susceptibility accuracy predicted by the RF model under the spatial resolution of 15 m and training and testing dataset proportion of 9: 1 is the highest, and that the more important environmental factors under each condition are elevation, slope and topographic relief, etc. With decreasing the spatial resolution and/or the proportion of training and testing dataset, the LSP accuracies of both SVM and RF models decrease gradually, and the mean values of landslide susceptibility indexes increase with a decrease of the corresponding standard deviation value. For all combined conditions, as the spatial resolutions and the proportions of training and testing dataset decrease, the LSP accuracies decrease gradually while the corresponding uncertainties increase. It is also indicated that the LSP accuracy of the RF model is better than that of the SVM model under various combined conditions, and that the influence of the spatial resolution on the RF model is significantly greater than that of the proportion of training and testing dataset while there is little difference between the effects of the two factors on the SVM model. © 2021, Science Press. All right reserved.