A Self-Supervised Framework for Refined Reconstruction of Geophysical Fields via Domain Adaptation

Abstract

Reconstructing fine-grained, detailed spatial structures from time-evolving coarse-scale geophysical fields has been a long-standing challenge. Current deep learning approaches addressing this issue generally require massive fine-scale fields as supervision, which is often unavailable due to limitations in existing observational systems and the scarcity of widespread high-precision sensors. Here, we present AdaptDeep, a self-supervised framework for refined reconstruction of geophysical fields via domain adaptation from the coarse-scale source domain to the fine-scale target domain. This method incorporates two pretext tasks, cropped field reconstruction and temporal augmentation-assisted contrastive learning, to leverage spatial and temporal correlations in the target domain. A global propagation structure is proposed in the feature extraction network to leverage bidirectional information for enhanced long-range dependencies and robustness against estimation errors. In experiments, AdaptDeep correctly identifies local, fine structures and significantly recovers 81.2% detailed information in sea surface temperature fields.