A critical view on the suitability of machine learning techniques to downscale climate change projections: Illustration for temperature with a toy experiment

Abstract

Machine learning is a growing field of research with many applications. It provides a series of techniques able to solve complex nonlinear problems, and that has promoted their application for statistical downscaling. Intercomparison exercises with other classical methods have so far shown promising results. Nevertheless, many evaluation studies of statistical downscaling methods neglect the analysis of their extrapolation capability. In this study, we aim to make a wakeup call to the community about the potential risks of using machine learning for statistical downscaling of climate change projections. We

Using Machine Learning to Cut the Cost of Dynamical Downscaling

“…By selecting training years from a century‐scale future projections as well as historical periods, the training set can be made to be representative of all GCM simulated climates, and plausible future climate change would fall within the calibration range of ML, thereby addressing the well‐known issue of poor performance when machine learning downscaling is applied to future projections (Hernanz et al., 2022; Lanzante et al., 2018). A smaller number of years could be selected to further reduce the cost of DD, and a different sampling method could be used, such as selecting the two wettest years, the two driest years and two normal years.…”

“…Compared to DD, the implementation of SD is fast and far less computationally intensive (Wang, Liu, et al., 2021). However, SD models can perform poorly under extrapolation to future climates as few methods account for nonstationary relationships between predictors and predictands under climate change (Hernanz et al., 2022; Hewitson et al., 2014; Lanzante et al., 2018; Salvi et al., 2016; Schoof, 2013), although there are some exceptions (Baño‐Medina et al., 2022; Pichuka & Maity, 2018). Despite the fairly low cost of SD, it is only possible to implement it in regions where fine‐scale observational data are available for training, and typically little is known about its ability to perform downscaling in regions outside of the training domain (Wang, Tian, et al., 2021).…”

Machine Learning‐Based Detection of Weather Fronts and Associated Extreme Precipitation in Historical and Future Climates

“…With regard to changes in specific front types with climate change, we note that Biard and Kunkel (2019) showed DL-FRONT had some difficulty distinguishing between cold and stationary fronts, adding some uncertainty in the projections of each front type. We also acknowledge concerns with extrapolating machine learning algorithms trained on historical data to a climate change context (Hernanz et al, 2022). Using a similar deep learning algorithm trained to detect atmospheric rivers and tropical cyclones in historical climate simulations, Prabhat et al (2021) successfully applied the algorithm to a climate change scenario without tuning, similar to the approach we have taken here.…”

High-resolution grids of daily air temperature for Peru - the new PISCOt v1.2 dataset

“…estimate the value outside the range of the training data limits its use in large regions with low station density 26,27 . Besides, since the relationship between air temperature and auxiliary spatial predictors varies on spatiotemporal scales, recent research has also highlighted the importance of non-stationarity in the spatiotemporal domain by building local models in contrast to global estimation models 13,[28][29][30][31][32] .…”