Reduction of Persistent Stress-Equivalent Wind Biases With Machine Learning and Scatterometer Data

The numerical weather prediction (NWP) stress-equivalent 10-m wind (U10S) forecasts are used as a common forcing for ocean models; however, these forecasts present local and systematic biases when compared to the observational data. The scatterometer wind observations are being assimilated by European Centre for Medium-Range Weather Forecasts (ECMWF) Integrated Forecasting System (IFS), but even after the assimilation, the sea-surface wind biases are still present. A previous approach to reduce such biases was based on correcting the forecasts with the mean differences between scatterometer observations and the NWP output accumulated over a certain period of time. However, this approach shows performance degradation for the periods when fewer scatterometers are available and in the operational framework. To overcome these limitations, we propose the use of machine learning (ML) to predict such biases using other atmospheric and oceanic NWP variables as inputs, so that the observational data are only required during the training. In this work, we show the results for the preliminary ML models trained on a small subset of data that use U10S scatterometer–NWP differences as the target. The predicted corrections applied to the ECMWF fifth reanalysis dataset ERA5 show error variance reduction up to 9.86% on a test subset globally when compared to Advanced Scatterometer (ASCAT-A) and up to 6.25% against independent scatterometer HSCAT-B, thereby reducing the local biases. The best performance is seen in the extra tropics with error variance reduction up to 10.6%.