5:15 PM - 7:15 PM
[HTT15-P03] Enhancing Earthquake-Induced Landslide Prediction with Transfer Learning: Comparison of Feature and Parameter Representation Transfer Models for Niigata, Japan
Keywords:Earthquake-Induced Landslide Risk Assessment , Transfer Learning, Feature Representation Transfer, Parameter Representation Transfer, GIS
This study presents a comparative analysis of feature representation transfer and parameter representation transfer techniques for predicting earthquake-induced landslides. Leveraging transfer learning, the research adapts a pre-trained Multilayer Perceptron (MLP) model, developed using a robust dataset from the 2015 Nepal earthquake comprising 3,700 landslide points and an equal number of non-landslide points, to the unique geological and seismic conditions of the 2004 Niigata earthquake region. This adaptation employed a balanced dataset of 200 landslide points and 200 non-landslide points as fine-tuning. Experimental results reveal that, under small-sample conditions, a standalone MLP model achieved an accuracy of 87.5% on the Niigata data, incorporating feature representation transfer improved accuracy to 93.8%, and parameter representation transfer further enhanced performance to 95.3%. In addition, a seismic landslide susceptibility map was produced to delineate areas in Niigata that are potentially susceptible to future earthquake-induced landslides.
The validation process demonstrates that transfer learning can significantly streamline model training and improve predictive accuracy in regions with limited training samples. These findings underscore that parameter representation transfer with comprehensive fine-tuning of model parameters provides a more effective adaptation of predictive models across regions with distinct seismic characteristics than feature representation transfer. Overall, this study highlights the potential of advanced transfer learning techniques to bridge data gaps and enhance landslide susceptibility assessments, emphasizing their promising application in global disaster risk management frameworks.
The validation process demonstrates that transfer learning can significantly streamline model training and improve predictive accuracy in regions with limited training samples. These findings underscore that parameter representation transfer with comprehensive fine-tuning of model parameters provides a more effective adaptation of predictive models across regions with distinct seismic characteristics than feature representation transfer. Overall, this study highlights the potential of advanced transfer learning techniques to bridge data gaps and enhance landslide susceptibility assessments, emphasizing their promising application in global disaster risk management frameworks.