A Posteriori Random Forests for Stochastic Downscaling of Precipitation by Predicting Probability Distributions

Abstract

Despite being the main tool used nowadays to simulate the evolution of the climate system, the spatial resolution of the current global climate models (GCMs)-typically up to around a hundred kilometers-is still insufficient for most practical applications (see e.g., Doblas-Reyes et al., 2013 and references therein). To alleviate this limitation, statistical downscaling (SD) methods aim to build models linking a set of key large-scale predictors (e.g., geopotential or winds) with the target predictand (e.g., precipitation or temperature) over the area of interest (see e.g., von Storch et al., 1993). Under the perfect prognosis (PP) approach (

Assessing Three Perfect Prognosis Methods for Statistical Downscaling of Climate Change Precipitation Scenarios

“…For this purpose, we update the split function used in Legasa et al. (2022), which is tasked with splitting the predictors' space to provide predictive samples of precipitation, to account for the mixed nature of the Bernoulli‐Gamma distribution by considering a mixture of the Gamma deviance and the binary cross‐entropy. Specifically, we define the split function to be, for a set of predictive precipitation observations $$\left\{{y}_{i}\right\}$$ falling on a leaf, $$$\stackrel{\text{Bernoulli}\,\text{Entropy}}{\overbrace{-\overline{p}\mathrm{log}\,\overline{p}-\left(1-\overline{p}\right)\mathrm{log}\left(1-\overline{p}\right)}}+\stackrel{\text{Gamma}\,\text{Deviance}}{\overbrace{2\sum\limits _{{y}_{i}^{+}\in \left\{{y}_{i}\right\}}\left(-\mathrm{log}\left(\frac{{y}_{i}^{+}}{{\overline{y}}^{+}}\right)+\frac{{y}_{i}^{+}-{\overline{y}}^{+}}{{\overline{y}}^{+}}\right)}},$$$ where $$\overline{p}$$ is the proportion of wet days in $$\left\{{y}_{i}\right\}$$, $${y}_{i}^{+}$$ is the intensity/rainfall on wet days, and $${\overline{y}}^{+}$$ the mean precipitation intensity for the wet days.…”

“…(2020) and Legasa et al. (2022), which conducted a comprehensive assessment of the suitability of different settings for CNNs and APRFs for statistical downscaling, respectively. Therefore, the present study provides a representative overview of the merits and demerits of the different techniques considered for our target task.…”

“…Nevertheless, Legasa et al. (2022) tested APRFs using reanalysis predictors, thoroughly assessing the calibration/training stage, and thus a relevant next question is whether this technique is also suitable for SD of climate change scenarios, that is, using GCM predictors. This issue is often referred to as transferability (Dayon et al., 2015; Hernanz et al., 2022).…”

“…On top of their interpretability and their skill capturing non-linear predictor-predictand relationships, RFs automatically perform feature/predictor selection, thus avoiding the complex, time-consuming and often human-guided task of pre-defining an informative set of predictors. Nevertheless, Legasa et al (2022) tested APRFs using reanalysis predictors, thoroughly assessing the calibration/training stage, and thus a relevant next question is whether this technique is also suitable for SD of climate change scenarios, that is, using GCM predictors. This issue is often referred to as transferability (Dayon et al, 2015;Hernanz et al, 2022).…”

Customized Deep Learning for Precipitation Bias Correction and Downscaling

“…A. Panofsky & G. W. Brier, 1968;Thrasher et al, 2012;Wood et al, 2002) and recent machine learning based approaches such as random forests (X. Legasa et al, 2022;Long et al, 2019;Mei et al, 2020;Pour et al, 2016), support vector machines (Tripathi et al, 2006) and artificial neural networks (Schoof & Pryor, 2001;Vandal et al, 2019). Recently, advances in deep learning have made a significant impact on many fields and have been proven superior to traditional machine learning methods because of their powerful abilities in learning spatiotemporal feature representation in an end-to-end manner (Ham et al, 2019;Reichstein et al, 2019;Shen, 2018).…”