Concepts for improving machine learning based landslide assessment

Abstract

The main idea of this chapter is to address some of the key issues that were recognized in Machine Learning (ML) based Landslide Assessment Modeling (LAM). Through the experience of the authors, elaborated in several case studies, including the City of Belgrade in Serbia, the City of Tuzla in Bosnia and Herzegovina, Ljubovija Municipality in Serbia, and Halenkovice area in Czech Republic, eight key issues were identified, and appropriate options, solutions, and some new concepts for overcoming them were introduced. The following issues were addressed: Landslide inventory enhancements (overcoming small number of landslide instances), Choice of attributes (which attributes are appropriate and pros and cons on attribute selection/extraction), Classification versus regression (which type of task is more appropriate in particular cases), Choice of ML technique (discussion of most popular ML techniques), Sampling strategy (overcoming the overfit by choosing training instances wisely), Cross-scaling (a new concept for improving the algorithm’s learning capacity), Quasi-hazard concept (introducing artificial temporal base for upgrading from susceptibility to hazard assessment), and Objective model evaluation (the best practice for validating resulting models against the existing inventory). All of them are followed by appropriate practical examples from one of abovementioned case studies. The ultimate objective is to provide guidance and inspire LAM community for a more innovative approach in modeling.