Exploring the effects of the design and quantity of absence data on the performance of random forest-based landslide susceptibility mapping

This study aims to explore the effects of the design and quantity of absence data on the performance of random forest-based landslide susceptibility mapping. First, the Youfang catchment was selected as the study area, since it has undergone landslides in recent years. Second, the Random, Bioclim, Improved Target Space Exteriorization Sampling (ITSES), one class SVM and domain methods were chosen to generate the absence data. Then, the size of absence data was set to 79, 158, 237, 316, 395, 790, 1580, 3160, 4740, 6320, and 7900, and we repeated the experiment 20 times to get reliable results. The random forest method was selected to construct the landslide susceptibility model. In addition, the model's accuracy and the continuous Boyce index were selected to evaluate the performance of the landslide susceptibility mapping under the various designs and different size of absence data. The results showed that the sampling range and size of the absence data demonstrated significant influences on the accuracy of the landslide susceptibility mapping. When the sampling area was reduced, the size of the absence data needed to be increased (1:5) to balance the presence and absence data in the environmental characteristic space. When the sample area was large (in the random method, it is 99%), the best ratio for the presence to absence data was 1:1. When the sample area was small, the best ratio for the presence to absence data was 1:100. Finally, the generated landslide susceptibility map may be useful for the local government and land use planners in the study area.