Urban road collapse risk assessment based on the extended xLSTM Network

Abstract

Rapid urbanization has substantially increased the complexity of urban underground spaces. This complexity leads to frequent road collapse incidents that pose significant threats to the safety and property of urban residents. Therefore, accurate methods of performing early road collapse risk assessments are crucial for preventing these incidents and emergency preparedness. In this study, road collapse incident data for 2016–2021 were collected for Foshan, Guangdong Province, a city in southern China. Utilizing InSAR time-series data from Sentinel-1 satellites, ground subsidence maps were generated, and the publicly accessible Ground Subsidence Trend-Based Urban Road Collapse Risk Dataset (GSTURCRD) was constructed. A novel risk assessment method for urban road collapse based on an extended long short-term memory (xLSTM) network was proposed. This method introduces two new LSTM variants, the scalar LSTM (sLSTM) and the matrix LSTM (mLSTM), incorporating exponential gating and an innovative matrix memory structure. These variants are integrated using residual connections to form a comprehensive network architecture that enables effective learning and representation of the temporal features. The experimental results from the dataset demonstrate that the proposed method significantly outperforms the original LSTM network and traditional machine learning methods regarding assessment capability (its accuracy was 0.886, and its recall was 0.857). Furthermore, the method's effectiveness was validated by an analysis of actual incidents that occurred in Foshan; thus, its ability to generate accurate and timely detections and provide early warnings for high-risk road sections in urban areas was confirmed.