InSAR time-series analysis and susceptibility mapping for land subsidence in Semarang, Indonesia using convolutional neural network and support vector regression

Abstract

Global sea-level rise due to climate change is a critical problem for coastal cities. One of the coastal cities in Indonesia, Semarang, is in danger of being submerged by seawater due to flooding, which is exacerbated by land subsidence. Time-series analysis based on persistent scatterer (PS) interferometry usually suffers from sparse measurement points (MP), and thus does not completely cover the study area with low coherence. In this study, we utilized improved combined scatterers interferometry with optimized point scatterers (ICOPS) to increase the MP density. We also compared the optimization of the ICOPS method based on support vector regression (SVR) with convolutional neural network (CNN) algorithms. The ICOPS-CNN showed better model performance and MP density, and was less time-consuming. Further land subsidence analysis was conducted in Semarang using susceptibility mapping. The land subsidence susceptibility map from the standalone deep learning (DL) algorithm (CNN = 0.833) outperformed the standalone machine learning (ML) algorithm (SVR = 0.820). The hybrid DL algorithm with grey wolf optimizer (GWO) (CNN-GWO = 0.907) exhibited the highest accuracy, followed by the hybrid ML algorithm (SVR-GWO = 0.843). The accuracy was consistent with the comparison between the susceptibility maps and the ICOPS measurements, resulting in a susceptibility map based on CNN-GWO with the highest accuracy (0.889). Land subsidence monitoring using ICOPS based on CNN and susceptibility mapping based on a hybrid DL algorithm (CNN-GWO) can be used by local governments to improve urban development planning in Semarang and its surrounding areas.