GDCM: Generalized data completion model for satellite observations

Abstract

Ocean remote sensing data is crucial in understanding the global climate system. Due to satellite orbital coverage gaps and cloud cover, satellite ocean remote sensing products have significant data gaps. This paper introduces a Generalized Data Completion Model (GDCM) based on deep learning to reconstruct gap-free and cloud-free key oceanic variables such as sea surface temperature (SST), wind speed, water vapor, cloud liquid water, and precipitation rate derived from polar-orbiting satellite sensors including Advanced Microwave Scanning Radiometer 2 (AMSR2), the Special Sensor Microwave Imager (SSMI), and the Moderate Resolution Imaging Spectroradiometer (MODIS). Utilizing Convolutional Neural Networks (CNNs) and attention mechanisms, the GDCM model effectively leverages spatio-temporal information within remote sensing data to fill in missing regions accurately. We use reanalysis data to simulate various data missing scenarios during model training for model development. We tested the model with the US East Coast region's global-coverage AMSR2/SSMI and local-coverage MODIS datasets. The experiments demonstrate that the GDCM model successfully and precisely completes the data for different satellites and types of missing data. To enable the model to capture enough data for the dynamical change patterns, we used seven consecutive days of observation data as inputs to improve the model's data-completion ability, significantly enhancing the handling of MODIS SST missing data due to cloud cover. When the input data's duration increased from one day to seven days, the model's R² value improved from 0.062 to 0.93, and the Root Mean Square Difference (RMSD) decreased from 6.58 to 0.92. Besides the model framework design, we implemented the incremental learning training strategy to enhance the model's data completion capability for different types of missing data, especially for SST data from AMSR2 satellites. The model's completed SST data R² value improved from 0.93 to 0.99, and the RMSD decreased from 2.64 °C to 0.50 °C. The Mean Absolute Difference (MAD) of water vapor data decreased from 0.88 kg/m² to 0.76 kg/m², and the RMSD decreased from 1.39 kg/m² to 1.27 kg/m². This study provides a generalized new solution to the problem of missing ocean data at different resolutions, contributing to a more comprehensive and supporting ocean science research and related applications.