2020/08/11 by Brian Groenke, Luke Madaus, Groenke, Brian +3 · 1 citation
Earth and Planetary Sciences · #Arctic and Antarctic ice dynamics #Climate change and permafrost #Computer Vision and Pattern Recognition (cs.CV) #Cryospheric studies and observations #FOS: Computer and information sciences #I.5.4 #Machine Learning (cs.LG) #Machine Learning (stat.ML)
paper · pdf · doi:10.48550/arxiv.2008.04679
openalex publication_date 2020/08/11 · openalex created_date 2022/07/26 · openalex updated_date 2026/07/28
Downscaling is a landmark task in climate science and meteorology in which\nthe goal is to use coarse scale, spatio-temporal data to infer values at finer\nscales. Statistical downscaling aims to approximate this task using statistical\npatterns gleaned from an existing dataset of downscaled values, often obtained\nfrom observations or physical models. In this work, we investigate the\napplication of deep latent variable learning to the task of statistical\ndownscaling. We present ClimAlign, a novel method for unsupervised, generative\ndownscaling using adaptations of recent work in normalizing flows for\nvariational inference. We evaluate the viability of our method using several\ndifferent metrics on two datasets consisting of daily temperature and\nprecipitation values gridded at low (1 degree latitude/longitude) and high (1/4\nand 1/8 degree) resolutions. We show that our method achieves comparable\npredictive performance to existing supervised statistical downscaling methods\nwhile simultaneously allowing for both conditional and unconditional sampling\nfrom the joint distribution over high and low resolution spatial fields. We\nprovide publicly accessible implementations of our method, as well as the\nbaselines used for comparison, on GitHub.\n