Construction of small sample seismic landslide susceptibility evaluation model based on Transfer Learning: a case study of Jiuzhaigou earthquake

• Published in: Bulletin of engineering geology and the environment 2022-03, Vol.81 (3), Article 116
• Main Authors: Ai, Xiao, Sun, Baitao, Chen, Xiangzhao
• Format: Article
• Language: English
• Subjects: Earth and Environmental Science; Earth Sciences; Foundations; Geoecology/Natural Processes; Geoengineering; Geotechnical Engineering & Applied Earth Sciences; Hydraulics; Nature Conservation; Original Paper
• ISSN: 1435-9529; 1435-9537
• DOI: 10.1007/s10064-022-02601-6

Traditional machine learning requires big data as the basis, and it is usually difficult to obtain the desired results in areas where there is a lack of data. This study proposes an innovative transfer learning method to establish a seismic landslide susceptibility evaluation model. The process is as follows: (1) a total of 13 influence factors were selected and combined with landslide points triggered by the Wenchuan and Jiuzhaigou earthquakes to form a big dataset and a small sample dataset, respectively; (2) Artificial Neural Network (ANN) was used to train the big dataset and prepare a pre-training model; (3) the Jiuzhaigou seismic landslide susceptibility evaluation model based on transfer learning was established by using the pre-training model. To reflect the advantages of the transfer learning method more intuitively, this study not only tested the accuracy of the evaluation model but also used ANN to train another evaluation model based on the small sample datasets. And the model accuracy was compared with that of the previous model. The results showed that the frequency ratio (FR) accuracy of the model obtained by transfer learning was higher than that of the model directly trained on a small sample dataset. Additionally, the area under curve (AUC) of the model directly trained on a small sample dataset was only 0.84, whereas the AUC of the model obtained by transfer learning was close to 0.90. The study shows that this method can solve the problems associated with traditional machine learning methods when establishing a seismic landslide susceptibility evaluation model.