Improvement of MiRS Sea Surface Temperature Retrievals Using a Machine Learning Approach

Abstract

We report on the development of a machine learning approach to improving sea surface temperature (SST) retrievals based on satellite-based microwave channel measurements at frequencies higher than 23 GHz. The approach uses a deep neural network (DNN) trained using Microwave Integrated Retrieval System physical retrievals as inputs and collocated European Centre for Medium-Range Weather Forecasts analyses for training and validation. The DNN was designed to characterize SST retrieval residual and then used to correct the original retrieval. Evaluation based on one year of independent data showed reduction in retrieval residual standard deviation from 3.22 to 1.80 K in January and 3.02 to 1.92 K in July and reduction in mean residual from 0.30 to 0.08 K in January and 0.61 to 0.22 K in July. Comparisons with multilinear regression and machine learning approaches that used measured brightness temperatures as inputs were significantly less effective in retrieving SST directly, although the DNN used brightness temperature also showed improvements. This indicates that physical retrieval provides valuable information useful in characterizing retrieval residual beyond that of the measured radiances. The DNN approach also effectively removed scan angle dependence of retrieval residuals—an important consideration with cross-track instruments. Sensitivity tests indicated that skill declines with time as time increases from training month, but that skill in the same month, one year later is nearly the same as that of the original training month. This suggests that it may be sufficient to pretrain a stratified model with monthly or seasonal dependence using one full annual cycle, which could then be used in subsequent years with continued good performance.