A Meta-Learning Approach of Optimisation for Spatial Prediction of Landslides

Abstract

Optimisation plays a key role in the application of machine learning in the spatial prediction of landslides. The common practice in optimising landslide prediction models is to search for optimal/suboptimal hyperparameter values in a number of predetermined hyperparameter configurations based on an objective function, i.e., k-fold cross-validation accuracy. However, the overhead of hyperparameter optimisation can be prohibitive, especially for computationally expensive algorithms. This paper introduces an optimisation approach based on meta-learning for the spatial prediction of landslides. The proposed approach is tested in a dense tropical forested area of Cameron Highlands, Malaysia. Instead of optimising prediction models with a large number of hyperparameter configurations, the proposed approach begins with promising configurations based on several basic and statistical meta-features. The proposed meta-learning approach was tested based on Bayesian optimisation as a hyperparameter tuning algorithm and random forest (RF) as a prediction model. The spatial database was established with a total of 63 historical landslides and 15 conditioning factors. Three RF models were constructed based on (1) default parameters as suggested by the sklearn library, (2) parameters suggested by the Bayesian optimisation (BO), and (3) parameters suggested by the proposed meta-learning approach (BO-ML). Based on five-fold cross-validation accuracy, the Bayesian method achieved the best performance for both the training (0.810) and test (0.802) datasets. The meta-learning approach achieved slightly lower accuracies than the Bayesian method for the training (0.769) and test (0.800) datasets. Similarly, based on F1-score and area under the receiving operating characteristic curves (AUROC), the models with optimised parameters either by the Bayesian or meta-learning methods produced more accurate landslide susceptibility assessment than the model with the default parameters. In the present approach, instead of learning from scratch, the meta-learning would begin with hyperparameter configurations optimal for the most similar previous datasets, which can be considerably helpful and time-saving for landslide modelings.