Accurate reconstruction of satellite-derived SST under cloud and cloud-free areas using a physically-informed machine learning approach

Abstract

Sea surface temperature (SST) is an important parameter affecting global climate, weather disasters, and marine resources. Acquiring SST data that covers large areas and spans over long periods is one of the most essential tasks for various scientific research. During the past decades, meteorological satellites (e.g., the Himawari 8) have been able to provide large-scale, high-resolution continuous observations (via a number of visible, nearinfrared, and infrared bands), but have always been affected by active atmospheric activities (i.e., clouds). A detailed literature review on SST analysis or estimation shows that limitations or challenges associated with the existing tools and the state-of-the-art approaches have not been fully resolved yet. Through integrating the knowledge from interdisciplinary domains, hence, we proposed a physically-informed machine learning approach (i.e., a physically-consistent, virtual-gauge approach in the machine learning framework) to elegantly reconstruct daily SSTs under both cloud and cloud-free areas. By this central idea, we developed the TS-RBFNN (i.e., Temporal-Spatial Radial Basis Function Neural Network) and suggested an adequate procedure (with artificial clouds) for model assessment since the data in the cloudy region was unavailable. A systematic study in terms of model implication (i.e., the meaning of network architecture), model validation (i.e., the performance of learning and generalization), and model applications (i.e., in open ocean and coastal seas with different cloud coverage over the four seasons) was conducted. In particular, a pattern similarity analysis (examining SST distributions for several selected sections) and a daily-based error analysis (presenting the variations and distributions of RMSEs for each season) were carried out to clarify the relationship between varying cloud conditions and model performances (inferenced by sunny areas). Overall, the TS-RBFNN would better perform full SST reconstruction with significant improvement up to 60%, compared to the DINEOF (i.e., Data Interpolation Empirical Orthogonal Function). Currently, the TS-RBFNN model is being implemented into the operational system of Taiwan's Central Weather Administration to provide all-weather SST products. In the near future, a long-term societal impact would be expected as the reconstructed SST data could be broadly employed in various scientific applications.