Answers are grounded in this paper and cite their sources.
Demonstrates improved landslide prediction in complex geophysical systems, indicating robustness and generalizability.
Accurate real-time landslide prediction is critical for natural hazard mitigation and protecting human lives. Real-time data from different environmental and geotechnical sensors used for landslide monitoring face extreme class imbalance, large sensor-to-sensor variations, and uncertain event timing. Hence, it becomes difficult for standard machine learning algorithms to identify critical slope failures. To solve this problem, the study proposes an Adaptive Attention-Based Stacked Ensemble Learning (ADASEL) framework for decision-level fusion of complex, imbalanced multi-sensor time-series systems. The proposed approach leverages diverse types of sensor data, both environmental and geotechnical sensors. The data collected are combined using a stacked ensemble learning paradigm, which uses base learners and meta-models. The base learners include logistic regression, random forest, lightGBM, extraTrees, XGBoost, CatBoost, and tabNet, while their outputs are optimized by an attention-based multilayer perceptron meta-model. Testing of the approach was conducted based on real-time landslide monitoring systems deployed at different locations within the Indian Himalayan Region. The experiments covered within site, cross-site bidirectional, and leave-one-site-out (LOSO) evaluations to estimate the performance of predictions and the generalizability of the approach. In all experimental setups, ADASEL outperforms the selected baselines on macro-F1, recall, and PR-AUC metrics, with low false alarm ratios and comparable accuracies. The results show that attention-based adaptive stacked fusion approach is effective in terms of its robustness and generalization performance at different monitoring sites. Additional research confirms the validity and interpretability of the approach. Overall, the framework can be used for analyzing complex systems featuring infrequent, yet significant events under conditions of uncertainty and heterogeneity.
No takes yet. Share an insight, caveat, or question.
Chand et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: