Intercomparison of machine learning methods for statistical downscaling: the case of daily and extreme precipitation

Abstract

Statistical downscaling of global climate models (GCMs) allows researchers to study local climate change effects decades into the future. A wide range of statistical models have been applied to downscaling GCMs but recent advances in machine learning have not been explored. In this paper, we compare four fundamental statistical methods, Bias Correction Spatial Disaggregation (BCSD), Ordinary Least Squares, Elastic-Net, and Support Vector Machine, with three more advanced machine learning methods, Multi-task Sparse Structure Learning (MSSL), BCSD coupled with MSSL, and Convolutional Neural Networks to downscale daily precipitation in the Northeast United States. Metrics to evaluate of each method's ability to capture daily anomalies, large scale climate shifts, and extremes are analyzed. We find that linear methods, led by BCSD, consistently outperform non-linear approaches. The direct application of stateof-the-art machine learning methods to statistical downscaling does not provide improvements over simpler, longstanding approaches. * vandal.t@husky.neu.edu † evan.kodra@risq.io ‡ a.ganguly@neu.edu limiting models to coarse spatial and temporal scale projections. While the coarse scale projections are useful in understanding climate change at a global and continental level, regional and local understanding is limited. Most often, the critical systems society depends on exist at the regional and local scale, where projections are most limited. Downscaling techniques are applied to provide climate projections at finer spatial scales, exploiting GCMs to build higher resolution outputs. Statistical and dynamical are the two classes of techniques used for downscaling. The statistical downscaling (SD) approach aims to learn a statistical relationship between coarse scale climate variables (ie. GCMs) and high resolution observations. The other approach, dynamical downscaling, joins the coarse grid GCM projections with known local and regional processes to build Regional Climate Models (RCMs). RCMs are unable to generalize from one region to another as the parameters and physical processes are tuned to specific regions. Though RCMs are useful for hypothesis testing, their lack of generality across regions and extensive computational resources required are strong disadvantages.

Double‐structured sparse multitask regression with application of statistical downscaling

“…One benchmark approach for statistical downscaling is the bias correction and spatial disaggregation (BCSD) method (Wood, Maurer, Kumar, & Lettenmaier, ; Bürger, Murdock, Werner, Sobie, & Cannon, ), which utilizes quantile mapping to perform bias correction between the coarse‐resolution climate models to the fine‐resolution projection. Vandal, Kodra, and Ganguly () showed that machine learning methods are competitive versus traditional statistical downscaling methods. Here, we compare the statistical downscaling performance of the DSSMR method versus BCSD and three machine learning methods studied in the work of Vandal et al (): task‐wise LASSO, ridge regression, and elastic net.…”

“…Vandal, Kodra, and Ganguly () showed that machine learning methods are competitive versus traditional statistical downscaling methods. Here, we compare the statistical downscaling performance of the DSSMR method versus BCSD and three machine learning methods studied in the work of Vandal et al (): task‐wise LASSO, ridge regression, and elastic net.…”

Quantifying Uncertainty in Discrete-Continuous and Skewed Data with Bayesian Deep Learning

“…Many statistical models have been explored for downscaling, from bias correction spatial disaggregation (BCSD) [6] and automated statistical downscaling (ASD) [15] to neural networks [33] and nearest neighbor models [16]. Multiple studies have compared different sets of statistical downscaling approaches on various climate variables and varying temporal and spatial scales showing that no approach consistently outperforms the others [5,14,35]. Recently, Vandal et al presented improved results with an alternative approach to downscaling by representing the data as "images" and adapting a deep learning based super-resolution model called DeepSD [36].…”

“…Root Mean Square Error (RMSE) and bias are compared to understand the average daily effects of downscaling. To analyze extremes, we select two metrics from Climdex (http:// www.clim-dex.org) which provides a suite of extreme precipitation indices and is often used for evaluating downscaling models [4,35]:…”

Statistical downscaling of global climate models with image super-resolution and uncertainty quantification

“…Though BCSD is a simple approach, it has been shown to perform well compared to more complex methods [12,83]. Furthermore, we have shown that BCSD performs similarly, or better, when compared to off-the-shelf ASD approaches [115].…”