A data-driven approach to precipitation parameterizations using\n convolutional encoder-decoder neural networks

Abstract

Numerical Weather Prediction (NWP) models represent sub-grid processes using\nparameterizations, which are often complex and a major source of uncertainty in\nweather forecasting. In this work, we devise a simple machine learning (ML)\nmethodology to learn parameterizations from basic NWP fields. Specifically, we\ndemonstrate how encoder-decoder Convolutional Neural Networks (CNN) can be used\nto derive total precipitation using geopotential height as the only input.\nSeveral popular neural network architectures, from the field of image\nprocessing, are considered and a comparison with baseline ML methodologies is\nprovided. We use NWP reanalysis data to train different ML models showing how\nencoder-decoder CNNs are able to interpret the spatial information contained in\nthe geopotential field to infer total precipitation with a high degree of\naccuracy. We also provide a method to identify the levels of the geopotential\nheight that have a higher influence on precipitation through a variable\nselection process. As far as we know, this paper covers the first attempt to\nmodel NWP parameterizations using CNN methodologies.\n