Land Subsidence Susceptibility Mapping in Kakinada, India using Machine Learning Techniques

Abstract

Land subsidence poses significant risks to the infrastructure and ecosystems in coastal urban areas, making it a critical concern for regions like Kakinada, India, a rapidly developing coastal city. To address this problem in Kakinada, this research integrated Interferometry-based land subsidence inventory data with seven key conditioning factors: elevation, lithology, geomorphology, land use, Normalised Difference Vegetation Index (NDVI), drainage density, and distance to roads. Five machine learning models were employed, namely Extreme Gradient Boosting (XG Boost), Random Forest, Gradient Boosting Machine (GBM), Logistic Regression, and Multilayer Perceptron, for susceptibility mapping. Sixty per cent of the data was used for model training, 20% for hyperparameter tuning, and 20% for testing model accuracy. The models’ ability to make accurate predictions was assessed through the area under the receiver operating characteristic curve (AUC-ROC) analysis. Results demonstrated that the GBM model achieved superior performance, with 82% accuracy and an AUC score of 0.88, while Random Forest showed comparable effectiveness. The analysis identified 50 square kilometres at high risk for future subsidence. The findings of this study will help urban planners identify the most vulnerable regions and implement strategies for sustainable development.