Received: 15 Aug 2020 – Discussion started: 05 Oct 2020
Abstract. Landslides are one of the major geohazards threatening human society. This study was aimed at conducting such a hazard risk prediction and zoning based on an efficient machine learning approach, Random Forest (RF), for Ruijin, Jiangxi, China. Multiple geospatial and geo-environmental data such as land cover, NDVI, landform, rainfall, stratigraphic lithology, proximity to faults, to roads and to rivers, depth of the weathered crust, etc., were utilized in this research. After pre-processing, including digitization, linear feature buffering and value assignment, 19 hazard-causative factors were eventually produced and converted into raster to constitute a 19-band geo-environmental dataset. 155 observed landslides that had truly taken places in the past 10 years were utilized to establish a vector layer. 70 % of the disaster sites (points) were randomly selected to compose a training set (TS) and the remained 30 % to form a validation set (VS). A number of non-risk samples were identified in low slope (< 1–3°) areas and also added to the TS and VS in the similar percentage. Then, RF-based classification algorithm was employed to model the probability of landslide occurrence using the above 19-band dataset as predictive variables and TS for training. After performance evaluation, the RF-based model was applied back to the integrated dataset to calculate the probability of the hazard occurrence in the whole study area. The predicted map was evaluated versus both TS and VS and found of high reliability in which the Overall Accuracy (OA) and Kappa Coefficient (KC) are 91.49 % and 0.8299 respectively. In terms of the risk probability, the predicted map was further zoned into different risk grades to constitute landslide risk map. Modeling results also revealed the order of importance of the 19 causative factors, and the most important ones are the proximity to roads, slope, May–July rainfall, NDVI and elevation. We hence conclude that the RF algorithm is able to achieve the risk prediction with high accuracy and reliability, and this study may provide an operational methodology for geohazard risk mapping and assessment. The results of this study can serve as reference for the local authorities in prevention and early warning of landslide hazard.
The paper is focused on landslides risk mapping and assessment using machine learning, notably, Random Forests algorithm, taking 19
geo-environmental factors into account. A number of innovative procedures were introduced in the research and reliable results of high accuracy obtained. We believe that the risk mapping approach developed in this paper is relevant and extendable to elsewhere, and the results can serve as reference for disaster prevention and early warning for the local government.
The paper is focused on landslides risk mapping and assessment using machine learning, notably,...