Abstract
<jats:p>The purpose of the present study is to describe a new algorithm based on a neural network approach (Convolutional Neural Network with encoder-decoder architecture, CNN) for cloud liquid water path (LWP) estimation over the global ocean from MTVZA-GYa satellite-based microwave radiometer observations. The input data for the network are the antenna temperatures measured in 10 MTVZA-GYa channels. The CNN was trained on a sample of 2 723 000 spatiotemporally collocated pairs of antenna temperatures and reference cloud LWP values from the ERA5 reanalysis. The data were selected for individual days across different seasons of 2024–2025 over the Pacific and Atlantic oceans, excluding polar areas. Verification of the retrieved LWP was performed against the spatiotemporally nearest “reference” LWP values from the ERA5 reanalysis and with LWP estimates from the AMSR2 microwave radiometer onboard the Japanese GCOM-W1 satellite. The root-mean-square deviation values calculated for various areas of the Atlantic and Pacific oceans within the latitude zone of ±60° for June 30, 2025 range from 0.05 to 0.075 kg/m² depending on the region and observation time. A visual comparison of the LWP fields retrieved from the MTVZA-GYa and AMSR2 data showed a good agreement in the patterns of high and low values for both kinds of estimates. The verification confirmed the operational capability of the proposed method for analyzing MTVZA-GYa data and demonstrated a satisfactory quality of the retrieved cloud LWP fields.</jats:p>