Hybrid Deep Learning Framework for Landslide Risk Prediction Using Meteorological and Geospatial Parameters

Abstract

Landslide prediction plays an important role in disaster risk reduction and protecting infrastructure and human lives when performed accurately and in a timely manner. Despite good performance, machine learning techniques are hindered by severe class imbalance in landslide datasets and the opacity of complex predictive models. To address these challenges, this study proposes an integrated framework for landslide risk prediction through systematic comparative analysis. The training data were first balanced using the Synthetic Minority Over-sampling Technique with meteorological and geospatial parameters. We then used Logistic Regression as a baseline, Random Forest as a machine learning benchmark, and two deep learning architectures, namely 1D-CNN and LSTM. A performance-weighted hybrid ensemble was also considered to combine the predictive capabilities of the individual models. Under the evaluated model configurations, Random Forest and the optimized 1D-CNN emerged as the strongest models, achieving test-set macro F1-scores of 0.91 and 0.90, respectively, with comparable cross-validation performance. The optimized 1D-CNN was further analyzed using SHapley Additive exPlanations to provide class-specific explanations of model predictions and identify the features driving landslide risk assessment. The outcomes illustrate an interpretable and reproducible workflow that combines class balancing, comparative model assessment, and explainable artificial intelligence for landslide risk prediction. The framework, however, needs to be validated using independent, real-world landslide inventories before it can be considered for operational early-warning applications.