6 - Land-Subsidence Spatial Modeling Using the Random Forest Data-Mining Technique

Abstract

The main objective of the present research is land-subsidence spatial modeling and its assessment using a random forest data-mining algorithm in Kerman Province, Iran. For this purpose, a land-subsidence inventory map was prepared using extensive field surveys in the study area. In the next step, the spatial relationship between land-subsidence locations and 10 effective factors, including slope percent, aspect, elevation, lithology units, distance from river, piezometric wells data, land use, plan curvature, topographic wetness index (TWI), and distance from faults were considered by frequency ratio (FR) theory and the weight of each factor’s classes were determined. Also, the results of the relationship between land subsidence and effective factors using random forest data-mining techniques indicated that three factors, including piezometric wells, elevation, and distance from faults had the most importance on land-subsidence occurrence in the study area, respectively. Finally, the land-subsidence susceptibility map was prepared using an RF model and “randomForest” package in the R statistical software. The results of the RF model evaluation using 30% of unused locations in the modeling process and based on a receiver operating characteristic (ROC) curve showed that the final land-subsidence susceptibility map shows excellent accuracy with an AUC value of 93%. Therefore, the mentioned map can be used for optimizing management of water resources and preventing the reoccurrence of this phenomenon in the study area.