PARAN: A novel physics-assisted reconstruction adversarial network using geostationary satellite data to reconstruct hourly sea surface temperatures

Abstract

Diurnal variability in sea surface temperature (SST) significantly influences ocean-atmosphere thermal interactions. Conventional numerical methods for reconstructing hourly SST are limited by high computational demands and difficulties in accommodating real-time atmospheric and oceanic conditions. Therefore, this study introduces the Physics-Assisted Reconstruction Adversarial Network (PARAN), a novel deep-learning framework that reconstructs high-resolution (2 km), hourly SST using geostationary satellite data. By integrating numerical model knowledge and leveraging generative adversarial network architectures, PARAN achieves enhanced reconstruction performance. We applied the PARAN framework to the Northwest Pacific region, generating a continuous, hourly subskin SST layer from January 2019 to December 2021. The network was validated against high-resolution satellite data and in situ buoy observations, demonstrating substantial accuracy improvements over existing high-resolution numerical models (GLO12v4 and HYCOM). We obtained higher correlation coefficients (0.994 and 0.982), negligible biases (0.007 and -0.165), and lower root mean square errors (0.435 and 0.766) than those of existing models when comparing PARAN with drifting and mooring buoys, respectively. In particular, PARAN effectively captured detailed spatiotemporal SST variations, including diurnal warming, and showed robustness under both clear and cloudy conditions.