Construction of Landslide Detection Model Using Optical Satellite Images and DEMs by Machine Learning with Multi-instance Decision Tree Model

Abstract

An accurate detection of landslides was achieved herein with the application of a decision-tree model with explanatory variables calculated from satellite images and digital elevation models (DEMs) learned from two cases of landslides in Japan with different causes. The study areas were (1) Atsuma, Hokkaido, where the 2018 Hokkaido Eastern Iburi earthquake occurred, and (2) Asakura, Fukuoka, where the heavy rainfall in northern Kyushu in July 2017 caused damage.

In the decision-tree analysis, explanatory variables were calculated from optical satellite images and DEMs. Based on these data, three models were generated: the ‘Atsuma model’ using the Atsuma case as supervised data, the ‘Asakura model’ using the Asakura case as supervised data, and the ‘Atsuma+Asakura model’ using pixels extracted from the Atsuma and Asakura cases as supervised data. Each model was applied to test data that were randomly extracted from the target areas of Atsuma and Asakura, and the accuracy of the landslide detection was quantitatively evaluated with Cohen’s Kappa, F-measure, and overall accuracy.

The analysis results demonstrated that the Atsuma+Asakura model had the same level of detection accuracy as the Atsuma model, with Cohen’s Kappa of 0.701 and overall accuracy at 91.7% in the Atsuma case. In the Asakura case, the Atsuma+Asakura model’s detection accuracy had a Cohen’s Kappa of 0.559 and overall accuracy at 90.0%, which were at the same level as those of the Asakura model. These results indicate that the decision-tree model trained on multiple cases has high detection accuracy and generalization performance and is effective for detecting landslides caused by multiple factors. The effectiveness of the parameters calculated from the optical satellite images and DEMs for detecting landslides was also quantitatively confirmed by ROC curves.