APPLICATION OF TRANSFER LEARNING IN RAINFALL-INDUCED LANDSLIDE DISPLACEMENT PREDICTION

Landslides are a common geological disaster in mountainous regions, with over 130,000 potential hazards identified in China, more than 70% of which are rainfall-induced. These landslides pose significant threats to lives and property. Since 2019, China has developed an automated geological disaster monitoring network, installing instruments such as rain gauges, soil moisture sensors, and GNSS on landslides to collect, process, analyze, and use real-time data for early warning and prediction. Historical data and models predict the displacement of rainfall-induced landslides over the next three days, playing a crucial role in disaster preparedness.

With the advancement of machine learning and the accumulation of landslide monitoring data, data-driven displacement prediction methods are gaining attention. However, landslides with less than a year of monitoring or without significant early deformation lack sufficient data, leading to poor prediction performance. Common modeling methods based on long-term data of a single landslide are often ineffective.

To address these issues, this study proposes a rainfall-induced landslide displacement prediction method based on transfer learning. This method leverages historical data from landslides with similar geological conditions, terrain features, and disaster types to train models. These models are then applied to predict and provide early warnings for newly monitored landslides. The model supports fine-tuning with new monitoring data, continuously improving prediction performance.

To verify this method, the study used monitoring data from 189 landslides for modeling experiments and performance evaluations. A case study shows that the TCN-Transformer model, trained on a multi-slope integrated dataset, can serve as an effective pretrained model for new monitoring slopes. The three-day average RMSE is reduced by 34.6% compared to models trained on individual slope data, successfully predicting most deformation peaks. Fine-tuning the model with accumulated data from newly monitored slopes further reduced the three-day RMSE by 37.2%, demonstrating significant predictive accuracy. In conclusion, leveraging transfer learning, the proposed method effectively uses available data for rapid deployment and continuous refinement of displacement predictions on newly monitored slopes.