Optimized Landslide Prediction Using LSTM and Random Forest Tuned by Black Hole Algorithm

Abstract

Landslides are among the most destructive natural hazards in mountainous regions, particularly in the Eastern Himalayas, where they cause severe damage to infrastructure, disrupt settlements, and result in significant human and economic losses. Accurate landslide susceptibility mapping remains challenging because conventional geospatial and machine learning approaches often fail to capture the complex spatial and temporal factors influencing landslide occurrence. This study proposes a hybrid framework integrating Long Short-Term Memory (LSTM) networks, Random Forest (RF) classifiers, and the Black Hole Algorithm (BHA) to improve landslide susceptibility assessment. The LSTM model analyzes temporal variables such as rainfall and Normalized Difference Vegetation Index (NDVI), while the RF model evaluates spatial factors including slope, elevation, lithology, and soil characteristics. BHA is employed to optimize model hyperparameters and enhance prediction performance. The proposed framework was validated using the Landslide Recent Incidents–India (2016–2020) dataset, comprising 32 geo-environmental attributes. Its predictive performance was evaluated using accuracy, precision, recall, and F1-score, and further verified through comparison with existing state-of-the-art landslide prediction models. The resulting susceptibility maps categorized areas into five risk classes: Very Low, Low, Moderate, High, and Very High. The proposed model achieved 99% accuracy, with 91% precision, 89% recall, and 90% F1-score, effectively identifying high-risk regions such as Assam, Sikkim, and Manipur. These results demonstrate that the proposed LSTM–RF framework optimized by BHA provides an accurate and reliable solution for landslide susceptibility mapping, supporting disaster risk reduction, infrastructure planning, and sustainable land-use management.