Enhancing the Spatial Resolution of Monthly Gravity Solutions Derived from SLR and DORIS

Abstract

Monthly gravity-field solutions derived from satellite laser ranging (SLR) and Doppler Orbitography and Radiopositioning Integrated by Satellite (DORIS) provide an observational record that extends well before the GRACE era but resolve only a limited range of spatial wavelengths. In this study, we implement deep learning methods to generate monthly gravity fields up to degree 60 from SLR/DORIS products (up to degree 10) by targeting COST-G GRACE(-FO) solutions. The proposed framework combines coefficient-level tokenization, a Transformer encoder, a degree-wise autoregressive gated recurrent unit (GRU) decoder, temporal conditioning, and a task-specific high-degree low-order loss. The final product is obtained by ensembling 20 independently trained models; its performance is evaluated on the held-out 2022--2023 test period. When assessed against COST-G ground truth in terms of equivalent water height, the ensemble achieves an area-weighted root-mean-square error (RMSE) of 4.06 cm and a spatial grid correlation of 0.986. It reduces grid-wise RMSE by 28.4\% relative to a recent SLR/DORIS reconstruction and achieves the best results for seven of the eight principal spectral, spatial, and basin-scale metrics. Application to the complete 1984--2023 SLR/DORIS record yields a consistent GRACE-like gravity field variation product and offers the potential to extend multiple downstream applications back through time.