Applying Deep Learning to Sea Surface Temperature Retrieval

Sea surface temperature (SST) is one of key parameters used for weather forecast and climate research, such as hurricane and cyclone forecast, Ocean climate, global warming, El Niño/La Niña and so on. The Global SSTs are generated using infrared or microwave remote sensing from polar orbiting or geostationary meteorological satellites. The satellite-derived SST algorithm is based on radiative transfer theory and was simplified to multi-channel regression approach, in which the regression coefficients are trained by prior estimated SST, in-situ SST or radiative transfer model simulated SST. This empirical regression SST algorithm has been comprehensively employed for national and international SST retrieval since it was firstly proposed by McClain et al., [1985], and until now it is still main SST algorithm in the world, although there may be physical SST retrieval explored in the SST community.

With the development of the state-of-art artificial intelligence, the deep learning (DL) method gradually become popular algorithm and is applied to most of science and technical fields, including atmosphere and ocean remote sensing. The advantage of the DL method is that we don’t need to understand and use complicated radiative transfer equation or tons of empirical regression coefficients to retrieve SST. It thus make the SST retrieval higher efficiency. In this study, we designed and developed a DL algorithm applied to SST retrieval using Python/Tensorflow tools. The input data are Sensor Data Record (SDR) data of the Visible Infrared Imaging Radiometer Suite (VIIRS) onboard Suomi National Polar-orbiting Partnership (S-NPP) - the one of Joint Polar Satellite System (JPSS) satellites. The well-validated NOAA Advance Clear-sky Processer over Ocean (ACSPO) SST was used as label data. The ACSPO clear-sky mask was used to identify clear-sky pixels for the DL training. The preliminary result showed that the DL trained SST were comparable with ACSPO SST for both mean and standard deviation. The statistics were getting worse when the test data set are far from the training data, which indicates that more training data is needed to prevent the DL-SST accuracy decay when the retrieve SST are far from test data.