Spatial Heterogeneity Analysis and Machine Learning-Based Forecasting of Land Subsidence in Ho Chi Minh City: A GWR and ConvLSTM Integrated Approach

Abstract

This study develops a framework for simulating and forecasting land subsidence in Ho Chi
Minh City, specifically focusing on District 12. By utilizing InSAR time-series subsidence data
from 2015-2020 alongside influencing factors such as building density, distance to water bodies,
and land use types, the research employs Geographically Weighted Regression (GWR) to analyze
the underlying subsidence mechanisms. Experimental results demonstrate significant spatial
heterogeneity in land deformation, where Land Use and Distance to Water emerge as the most
dominant factors, with average regression coefficients of -0.390 and -0.344, respectively.
Furthermore, the study proposes an integrated forecasting system architecture leveraging
advanced Machine Learning models, including Random Forest, XGBoost, and ConvLSTM deep
learning architectures to predict future surface deformation. Risk zonation results derived from
K-means clustering provide effective visual tools for urban planning and early warning systems
for geological hazards.