Landslide areas were predicted using shift-invariant neural network and maximum likelihood method. To predict the areas, elevation, inclination, and convexity, which are closely related to the landslide, were used as explanatory variables. The predictions were calculated using one variable, two variables, and all three variables. The results were then compared between the two methods. To compare the results, two indices were calculated, one is the landslide correct ratio, which is predicted landslide areas per actual landslide-occurred areas, the other is non-landslide correct ratio, which is predicted non-landslide areas per actual non-landslide areas. If the former ratio is much higher than the latter, it is understood that the result may predict areas of excessive landslide. In this study, it is thought that the higher both ratios are at the same time, the more valid the prediction is.
As a result, both ratios were higher than 70% in some cases using neural network. But no case using maximum likelihood method produced cases where were higher than 70%, and the former ratio (predicted landslide areas) was extremely higher than the latter (non-landslide areas). This result suggests that the neural network method is better than the maximum likelihood method in predicting landslide areas.
View full abstract