<p>Landslide is one of the world’s most hazardous geological disasters caused by geological, hydrological, and meteorological conditions. Conventional early warning systems are mostly based on threshold models or single sensor inputs, which do not clearly represent the nonlinear and dynamic slope failure behavior. To address these issues, this research work introduces a Multi-Sensor Long Short-Term Memory (LSTM) model for accurately predicting landslide occurrences. The system incorporates heterogeneous sensors, such as rain gauges, piezometers, inclinometers, and tensiometers. With the help of such sensors, some of the characteristics such as water level, pore water pressure, slope displacement, and soil moisture can be obtained, which provides an overall description of the slope stability condition. Leveraging the updated features such as the mean of inclinometer readings, derived displacement sign and displacement absolute magnitude, the Multi-Sensor LSTM networks capture long-term dependencies as well as nonlinear relationships among different sensor modalities, thereby improving prediction accuracy. Oversampling is considered before model training to correct class imbalance associated with landslide datasets, guaranteeing balanced representation of low, medium, and high-risk classes. Hyperparameter tuning is incorporated to enhance the model performance by systematic parameter optimization. Adam optimizer is utilized to optimize the model. The model has been tested with a data set from landslide-prone areas of Meghalaya, India, that are extremely vulnerable to rainfall-induced slope failures. The results indicate that the combination of oversampling and hyperparameter tuning improves the identification of minority risks. Thus the model is more accurate, reduces false alarms, and provides more timely predictions than traditional statistical and machine learning methods.</p>