Assessing the Suitability of A Posteriori Random Forests for Downscaling Climate Change Projections

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 and under the perfect prognosis approach, these methods learn the relationships that link several large-scale predictor variables coming from a reanalysis (e.g. humidity) with the local variables of interest (e.g. precipitation) over a reference historical period. Subsequently, the so-learnt relationships are applied to GCM predictors to obtain downscaled projections for a future period.

In a recent paper, Legasa et al. (2021) introduced a posteriori random forests (AP-RFs), a modification of classical random forests which make use of all the data in the leaves to estimate any probability distribution. Following the experimental framework proposed in Experiment 1 of VALUE (http://www.value-cost.eu, Gutiérrez et al. 2018), the study showed that AP-RFs obtained reliable stochastic time-series over several locations in Europe using reanalysis predictors. As compared to more classical techniques like generalized linear models (GLMs), this study concluded that AP-RFs are a competitive SD method in terms of different forecast aspects, with one of their key advantages being the ability to automatically perform predictor/feature selection. This avoids the task of manually selecting the most adequate large-scale variables and geographical domain of interest, something which, at present, relies on human expertise and constitutes a substantial source of uncertainty for downscaling climate change projections.

Nevertheless, an assessment of the suitability of AP-RFs for producing local climate change projections from GCM predictors is still lacking. This work aims to fill this gap by providing a fair comparison of AP-RFs with GLMs and state-of-the-art convolutional neural networks (CNNs), which were recently shown to provide satisfactory results for this task (Baño-Medina et al. 2021). We build on VALUE’s Experiment 2a and train the different methods considered using ERA-Interim “perfect” predictors. Afterwards, the EC-Earth model is used to generate downscaled projections for 86 locations distributed across Europe under a strong emission scenario, the RCP8.5. 

Our preliminary results suggest that AP-RFs generate plausible downscaled future projections of precipitation. In particular, differently to traditional GLMs, which are very sensitive to the predictor set considered and may produce implausible climate change projections (Manzanas et al. 2020), this technique yields delta changes consistent with those obtained from both the raw EC-EARTH outputs and the CNNs.

References
Baño-Medina, J., Manzanas, R. & Gutiérrez, J.M. On the suitability of deep convolutional neural networks for continental-wide downscaling of climate change projections. Clim Dyn 57, 2941–2951 (2021). doi: https://doi.org/10.1007/s00382-021-05847-0

Gutiérrez, J.M., Maraun, D., Widmann, M. et al. An intercomparison of a large ensemble of statistical downscaling methods over Europe: Results from the VALUE perfect predictor cross-validation experiment. Int. J. Climatol. 2019; 39: 3750– 3785. doi:  https://doi.org/10.1002/joc.5462

Legasa M.N., Manzanas R., Calviño, A. et al. A Posteriori Random Forests for Stochastic Downscaling of Precipitation by Reliably Predicting Probability Distributions. Submitted to Water Resources Research.

Manzanas, R., Fiwa, L., Vanya, C. et al. Statistical downscaling or bias adjustment? A case study involving implausible climate change projections of precipitation in Malawi. Climatic Change 162, 1437-1453 (2020). doi:  https://doi.org/10.1007/s10584-020-02867-3