Multi-source deep data fusion and super-resolution for downscaling sea surface temperature guided by Generative Adversarial Network-based spatiotemporal dependency learning

In this study, we propose a deep learning framework for multi-source deep data fusion and super-resolution for generative adversarial network-based spatiotemporal dependency learning to produce accurately downscaled sea surface temperature (SST) through simultaneously achieving error correction and improvement of spatial resolution. The proposed method is applied to the global ocean and the Korean waters, which is a regional sea, and experiments are conducted to downscale the SST by 2.5 and 5 times, respectively. The multi-source SST data used are numerical reanalysis, multiple satellite composites, and in-situ measurements, and two loss functions of super-resolution and mean square error are applied for adversarial learning. For more reliable performance evaluation, spatially, the global ocean and Korean waters are divided into a number of regional seas classified by characteristics of ocean physics, and temporally, the overall test period is divided into seasons and when extreme events occur. The overall results showed good performance for most experiments when both error correction through data fusion and spatiotemporal dependency learning from consecutive multiple input sequences using low-resolution reanalysis data, high-resolution satellite composite, and in-situ measurements were performed. However, for summer, winter, or extreme event periods, high performance was shown when using low-resolution satellite composite data with the same modality as the target data was used as an input. Furthermore, as a result of a blind test on the trained model with high-resolution target data used as target for the test period as input, the model that learned spatiotemporal dependency learning with error correction through data fusion showed the best and most consistent generalized downscaling performance compared to the test performance.