Answers are grounded in this paper and cite their sources.
Retrieving atmospheric parameters (e.g., XCO₂ ) from satellite hyperspectral data faces two fundamental physical challenges: spectral ambiguity, where sensor noise overwhelms weak absorption features, and covariate shift, where observing geometries nonlinearly distort spectral distributions. To address these, we propose a Physics-Enhanced Transformer framework integrating physical constraints as inductive biases. Specifically, a Gradient-Guided Smoothing (GPS) module leverages spectral gradients to dynamically balance denoising and absorption feature fidelity. Furthermore, a Physics-Aware Spectral Calibration (GSC) module explicitly models observing geometry effects to decouple environmental covariates from the target gas signal. Experiments on a large-scale OCO-2 dataset demonstrate state-of-the-art performance across two operational scenarios. Under strict out-of-time extrapolation (training 2020–2022, testing 2023), the model exhibits strong physical generalization with an RMSE of 1.519 ppm, overcoming natural inter-annual baseline shifts. In a spatiotemporal gap-filling scenario, it achieves an exceptional RMSE of 1.033 ppm. Independent spatial holdout validation at TCCON ground sites reveals high consistency with the Optimal Estimation Method (OEM) supervision product. Crucially, the network acts as a robust nonlinear statistical filter; it successfully isolates aleatoric uncertainty propagated from instrument noise in traditional pixel-wise inversions, serving as a highly effective error-correction and spatiotemporal smoothing surrogate.
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: