Geospatial Machine Learning Models for Integrated Landslide Risk and Asset Exposure Assessment

Abstract. Landslides increasingly threaten communities, infrastructure, and ecosystems, especially in areas with steep gradients, heavy precipitation, and unregulated land utilization. This research employs a hybrid geospatial and machine learning (ML) methodology to evaluate landslide susceptibility and asset exposure in Western Maharashtra, an area noted for its significant vulnerability. Three machine learning models—Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Back Propagation Neural Network (BPNN)—were trained utilizing topographic, hydrological, land use, and soil variables. XGBoost attained the best accuracy at 91.17%, generating a susceptibility map divided into five risk groups. To assess exposure, the susceptibility outputs were combined with building footprint data, indicating significant threats to residential areas, infrastructure systems, and agricultural land. Google Earth Engine (GEE) facilitated satellite-based analysis, whilst Google Colab enabled model training and validation. The results indicate robust model performance; however, limitations include reliance on static input data and the lack of real-time environmental monitoring. Future endeavors intend to integrate dynamic datasets, advanced deep learning architectures, and IoT-based early warning systems. The research highlights the need of combining geospatial analysis with machine learning methods for sustainable disaster risk mitigation and informed spatial planning.