Global Monitoring of Sea Surface Temperature Using Data from the MTVZA-GYa Satellite Microwave Radiometer

The target instrumentation of the polar-orbiting spacecraft (SC) of the METEOR-M series (SC No. 1, 2, 2-2, 2-3, 2-4, and subsequent) includes the MTVZA-GYa microwave radiometer with scanning and atmospheric sounding functions. Its measurements are designed for the remote determination of geophysical parameters of the atmosphere and underlying surface, including global monitoring of sea surface temperature (SST). The radiometric channels of the MTVZA-GYa scanner operate in atmospheric transparency windows at frequencies of 10.6, 18.7, 23.8, 31.5, 36.5, 42.0, 48.0, and 91.65 GHz with vertical and horizontal polarization. Starting from the METEOR-M No. 2-3 spacecraft (launched in 2019), the measurement capabilities of MTVZA-GYa have been experimentally expanded to include scanning channels at 6.9 and 7.3 GHz.

This report discusses the development and application of a modified multilayer perceptron (MLP) neural network algorithm, updated compared to [1], for retrieving SST from MTVZA-GYa data acquired by the METEOR-M No. 2-4 spacecraft (launched in 2024). The input data for the MLP are antenna temperatures (Ta) measured in five vertically polarized MTVZA-GYa channels at 10.65, 18.7, 23.8, 31.5, and 36.5 GHz, which, according to theoretical estimates, are most sensitive to SST variations. The decision to avoid using brightness temperatures (Tb) as predictors is due to significant systematic discrepancies between Ta and Tb, as well as the intention to eliminate additional errors introduced by external calibration procedures during the conversion from the Ta to Tb scale. The neural network algorithm employs a four-layer unidirectional architecture with the RELU activation function on three hidden layers and a linear activation function on the output layer. The hidden layers contain 40, 30, and 15 neurons, respectively. The network was trained on a dataset of 6,000 "pairs" comprising spatially and temporally collocated antenna temperatures (Ta) and SST values derived from marine buoy observations in the ICOADS database. The training dataset was limited to the region between 60°S and 60°N latitude and 75°W and 30°E longitude, using data from four days: July 15, August 15, September 30, and October 10, 2024.

Examples of global SST maps derived from MTVZA-GYa data for November 1, 2024, are presented, alongside comparisons with climatological mean SST values from the ERA5 reanalysis. Verification of satellite SST estimates was performed by comparing them with near-coincident marine buoy observations after applying a filtering procedure to exclude outliers deviating from climatological mean SST by more than ±3 standard deviations. The standard deviations were calculated for each grid node of the ERA5 climatological field. This filtering retains over 65% of the satellite SST estimates. The root mean square error (RMSE) of SST retrieval across the globe within the ±60° latitude zone on November 1, 2024, ranges from 1.0 to 2.0°C, depending on the sensing region. Considering the relatively high instrumental noise level of MTVZA-GYa data, these results confirm the operational viability of the proposed approach.

1. Масляшова А.О., Успенский А.Б. Картирование температуры поверхности океана по данным микроволнового радиометра МТВЗА-ГЯ со спутника «Метеор-М» № 2-4 // Материалы 22-й Международной конференции «Современные проблемы дистанционного зондирования Земли из космоса». Москва: ИКИ РАН, 2024. C. 46. DOI 10.21046/22DZZconf-2024a