Predicting the unknown: using transfer learning techniques for landslide susceptibility modelling

Landslide susceptibility modelling is an effective way to assist decision-makers in minimizing landslide risk. Developing landslide inventories for susceptibility model training and testing can be of high cost and effort. Previous studies have pointed out that landslide inventories from different areas are able to provide informative knowledge of landslides. However, training models for target areas using data from different regions at different times is challenging due to differences in feature space and/or the data distribution. Traditional machine learning techniques assume that target and source areas have the same data distribution. Therefore, when the data distribution is different, their performance can be degraded. Transfer learning can solve new problems using knowledge extracted from previous experiences. Case-based reasoning (CBR) is a transfer learning method that determines the similarity between source and target areas by considering various attributes (e.g., topography, geology, data structure). The resulting most similar or a weighted combination of models from similar source areas are applied to model susceptibility in the target area. Instead of applying the similarity obtained by CBR, domain adaptation (DA) transfers the knowledge by considering the data distribution between source and target areas. These two techniques are rarely used in landslide assessment studies, yet they have excellent potential to enhance spatial model transfers.

We evaluated the performance of transfer learning using CBR, DA and a CBR-DA combination for landslide susceptibility modelling. Our assessment is based on ten study areas with various spatial resolutions (1 m, 10 m and 25 m) located in Austria, Ecuador and Italy. We explored two modelling scenarios: only one source area available (single-source learning) and multiple source areas (multi-source learning) and compared them to benchmark situations, that is, landslide susceptibility models that are applied to the target area without using transfer learning techniques. Our results clearly showed CBR strategies using single-source learning and multi-source learning were robust and effective in developing highly transferable landslide susceptibility models without any a prior knowledge of landslides in target area. Since-source CBR was the most effective method for model transferring. The proposed method can alleviate the burden of collecting and labelling data, resulting in a more expedited preparation of landslide susceptibility maps for large and data-scarce regions.