[M-GI31] Earth and planetary informatics and data utilization

[MGI31-P05] Evaluation of Automatic Slope Failure Identification Method Using CVA and Random Forest

*Mitsunori UEDA¹, Tatsuya NEMOTO¹, Venkatesh RAGHAVAN¹ (1.Osaka Metropolitan University)

Keywords:Slope Failure, Change Vector Analysis, Random Forest, Automatic Identification, Feature Selection

Slope failure is a phenomenon in which soil and rocks slide down a slope and distribution maps of slope failure areas are created for the purpose of solving geological problems such as disaster prevention, urban planning, and land development. However, visual interpretation requires a lot of time and effort and there are problems with objectivity and reproducibility due to differences in the experience and interpretation of scientists and engineers. Therefore, there is a need to develop an automatic slope failure area identification method using machine learning. Conventionally, the machine learning models for extracting slope failure areas using topographic information have focused on the mechanical characteristics of slope failures and used the corresponding topographical information as input data for learning. However, the appropriate learning features for identifying slope failure areas are not known. In addition, to understand the characteristics of slope failures, it is necessary to investigate not only the location and scale but also morphological features such as scarp and main body at the slope failure area. In this study, we used Change Vector Analysis (CVA) and Random Forest Classifier (RFC) to construct an automatic slope failure area identification model, extract scarp and main body in slope failure areas, verify the accuracy of the model and evaluate learning features. CVA is a method to express the temporal changes in paired images of two periods as vectors. RFC is machine learning method that performs classification by arranging multiple decision trees in parallel and taking a majority vote. We calculated topographical information using Digital Elevation Model (DEM) before and after the slope failure and analyzed the temporal changes using CVA. The strength and angle of the change vector of the topographical information were used as learning features for RFC to classify the area into scarp, main body, and non-slope failure area. The target slope failure areas were caused by heavy rain that occurred in Hyogo Prefecture in 2014. 1 m DEM was generated from the aerial laser survey data taken before and after the slope failure. The calculated topographic data are Slope angle, Slope aspect, Laplacian, Normal vector of topographic surface, Variance of normal vector, Topographic classification, Topographic roughness and Convergence index. In the results of CVA, non-slope failure areas were concentrated in areas with low change vector strength values. In addition, the scarp was found to have angle of change vector concentrated at specific values. This suggests that the strength of the change vector is useful for identifying the slope failure area or non-slope failure and the angle of the change vector is useful for identifying morphological features. The accuracy of the RFC model was evaluated using the kappa coefficient, which showed a substantial agreement of 0.75. In order to confirm the important features for identifying slope failure areas, we calculated the contribution rate of each feature in the RFC model. As the results, the vector angle in the paired image of the normal vector of the topographic surface and its variance showed the highest value, followed by the vector strength in the elevation and Laplacian. This shows that geometric characteristics are important for identification. Regarding the normal vector and its variance (angle), angles less than 100° contributed to learning, and angles greater than 100° hardly contributed to learning. For elevation and Laplacian (strength), a constant contribution rate was shown for all values. For this reason, in order to simplify the machine learning model and improve its robustness, it is necessary to select features that focus on the feature threshold. In the future, it will be necessary to investigate the best combination of learning features and to consider improving the extraction accuracy of slope failure areas by using other machine learning methods.