Few-Shot Prediction of Landslide Susceptibility Based on Meta-Learning Paradigm
Objectives:The landslides disaster is one of the most important factors for the construction of major infrastructure in the western China.How to effectively and reliably carry out wide-area landslide sus-ceptibility prediction has always been a frontier difficulty in domestic and foreign studies.However,the present data-driven methods for prediction of landslide susceptibility in large-scale and complex scenarios still face two major issues:(1)The difference between various landslide inducing environments in a wide range of scenarios,would cause the difficulty to applying a single model to account for multiple landslide phenomenon;(2)small sample problem:Complex environmental tasks require models with large capacity and strong representative power,but there is a lack of sufficient landslide samples in practice.Methods:In response to the above problems,this paper takes Qijiang and Fuling District of Chongqing City,China as an example,proposes a local prediction strategy,and introduces the idea of meta-training an intermediate representation suited to be generalized,that can be adapted to the landslide sensitivity prediction task cor-responding to the current local area,with only a very small number of samples in the local area,and num-ber of iterations.Thus,the two issues mentioned can be well settled.Results:The proposed method is dif-ferent from traditional methods such as support vector machines,multilayer perceptrons,and random forests,which require a large number of samples and gradient iterations to train the supervised model.Instead,only a small sample is required to fine-tune the intermediate model,which still improves the global accuracy by 1％-5％,the precision by 1％-3％,the F1-score by 0.5％-6％,and the recall rate is close to the highest level of other methods.Conclusions:The meta-learning paradigm enables few-shot adaptation of landslide susceptibility prediction model with superior statistical performance.
万方数据
维普
