Development of a Land Subsidence Forecasting Model Using Small Baseline Subset—Differential Synthetic Aperture Radar Interferometry and Particle Swarm Optimization—Random Forest (Case Study: Tehran-Karaj-Shahriyar Aquifer, Iran)

Land subsidence, as a dangerous environmental issue, causes serious damages to farms and urban infrastructure. In this regards, this research was conducted with aimed to assess the efficiency of hybrid algorithm Particle Swarm Optimization–Random forest (PSO-RF) for developing land subsidence prediction model. PSO algorithm was used to select the factors affecting land subsidence, and RF algorithm was used as a classifier. Initially, the subsidence map of the region was obtained using the SBAS-DInSAR method throughout, for 2004 to 2009. The subsidence pattern was V-shaped, with an average of 13.8 cm per year. Then 11 factors dependent to the land subsidence event were prepared as PSO-RF inputs in GIS environment. Then, the weight of each of these factors was calculated using frequency ratio. Finally, 8888 points were randomly extracted from the subsidence map that had effective factors in land subsidence, as well as class 0 (no subsidence) or 1 (subsidence). About 6255 samples were selected for training and 2633 samples for validation of the model. The accuracy of the generated maps was then evaluated using the area under the receiver operating characteristic curve (AUC), RMSE and the accuracy (AC). The PSO-RF approach had a strong predictive accuracy with the smallest prediction error to map the LS hazard subsidence (i.e., AUC training = 93.2%, AUC validation = 89.8%, AC training = 0.86, AC validation = 0.81, RMSE training = 0.43, RMSE validation = 0.55). It was found that the media aquifer was the furthermost effective factor in the land subsidence development and followed by groundwater drawdown and transmissivity and storage coefficient.