• English 
  • Spanish 

Testing the Suitability of Tree-based Models for Climate Statistical Downscaling

Conference: 10th International Conference on Climate Informatics
Year: 2020
Contribution type: Poster
Poster: Poster.pdf

Statistical downscaling (SD) methods are extensively used to provide high-resolution climate information based on the coarse outputs from Global Climate Models (GCM). In the context of climate change, these methods are designed to learn the relationships that link key large-scale predictor variables (e.g. humidity or geopotential) with the local variables of interest (e.g. precipitation or temperature) over a reference historical period. These relationships are subsequently used to downscale the future simulations provided by the GCMs.

Many methods have been proposed for this task, ranging from relatively simple techniques like linear or generalized linear models to the more complex deep neural networks. In the framework of the European initiative VALUE (http://www.value-cost.eu), Gutiérrez et al. 2018* presented the most extensive intercomparison study to-date, with over 50 standard (i.e. well-established within the climate community) SD methods. Beyond these methods, a number of more sophisticated machine learning techniques have been also applied to SD. However, a fair comparison of these techniques with the more classical ones under a common experimental framework is still missing. This work contributes to fill this gap by testing the suitability of tree-based methods, in particular random forests and gradient boosting.

Our results indicate that random forests and gradient boosting provide a competitive alternative for SD. Moreover, one of their key advantages is the ability to automatically extract the relevant information out of the predictors. This avoids the complex task of manually selecting the most adequate large-scale variables, something which, at present, relies on human expertise and constitutes a substantial source of uncertainty for climate change projections. Our preliminary results highlight the capacity of tree-based methods to handle this issue, leading to robust results independently of the predictor variables considered.