Neural network approaches for enhanced landslide prediction: a comparative study for Mawiongrim, Meghalaya, India

The present study is an in-depth examination of machine learning strategies for landslide prediction that use three primary methods: Recurrent Neural Network (RNN), Gated recurrent unit (GRU), and long short-term memory (LSTM). The research investigates their varied benefits and limits to comprehending landslide prediction at a location situated in Mawiongrim, Meghalaya, India. RNN, LSTM, and GRU are compared and evaluated using large datasets from data collected using the real-time system set up in the area. The conclusion suggests combining RNN, LSTM, and GRU provides comprehensive accuracy and precision for landslide prediction. After thoroughly examining these variables, it was determined that LSTM outperforms RNN and GRU in terms of result prediction for slope displacement. Additionally, the study provides prediction models for matric suction, a critical factor in forecasting landslides caused by rainfall. In comparison to LSTM and GRU, it was discovered that the RNN model could predict the matric suction parameter more accurately. By examining the error metrics RMSE, MAPE, and MAE measures, it can be determined that the prediction models outperform the currently used ones. Of the three models selected, the LSTM model gives the best prediction for slope inclination with 82% accuracy. For matric suction prediction RNN model gives the best prediction with 67% accuracy. When used in conjunction with the sensor network system and prediction models, the landslide prediction system can continuously and in real-time monitor several landslide parameters, including soil mass displacement and matric suction in soil strata. It is possible to set up alerts and short messages to notify administrative staff members in various roles of an approaching landslide and advise them to leave the area to prevent any catastrophic destruction.