Dec 17, 2020
Published in: Research in Research in Mathematics and Public Policy. Association for Women in Mathematics Series (Volume 23, 2020), Chapter 1, pages 1–17 . doi: 10.1007/978-3-030-58748-2_1
Posted on RAND.org on December 22, 2020
This article was published outside of RAND. The full text of the article can be found at the link above.
Policy makers need information about future climate change on spatial scales much finer than is available from typical climate model grids. New and creative methods are being advanced to downscale climate change projections with statistical methods. Important requirements are to reliably downscale the climate parameter means, variability, extremes and trends, while preserving spatial and temporal correlations and permitting uncertainty quantification. In this proof-of-concept study, datasets derived from both observations and climate models were used together to train and test statistical methods. Two machine learning techniques-artificial neural networks and random forests-were tested on the problem of using coarse-scale climate projections (here represented by ERA-Interim reanalyses) to create temperature predictions at specific locations in areas of complex terrain. The methods are trained on and validated by temperature readings from mesonet weather stations in Colorado. This work has implications for fire prevention and water resources management, among other applications.