DBPFNet: a double branch parallel fusion neural network method for land subsidence susceptibility mapping with InSAR observation data

Current machine learning methods for land subsidence susceptibility mapping (LSSM) predominantly focus on the spatial features of land subsidence conditioning factors (LSCFs), overlooking the sequence relationships that merger after the superposition of these factors. This often leads to unreliable LSSM results. To address this limitation, this paper proposes a novel double-branch parallel fusion neural network, termed DBPFNet, which integrates multi-factor sequence and spatial features to improve LSSM accuracy. The Beijing Plain is selected as the study area. InSAR-derived land subsidence data are used as positive samples, and 12 LSCFs are chosen for analysis. Convolutional neural network (CNN) is employed to learn multi-factor spatial features, while long short-term memory (LSTM) is used to learn multi-factor sequence features. The spatial and sequence features are fuzed by two full connections to generate the LSSM. Experimental results demonstrate that the proposed DBPFNet model significantly outperforms CNN, LSTM and transformer models in terms of performance, yielding highly accurate LSSM result. The high susceptibility areas are predominantly located in the central region of Beijing Plain. Key factors influencing land subsidence in the study area include groundwater, altitude, build density, precipitation and river density.