Neural Spatiotemporal Interpolation for Gap Filling of GNSS-R Soil Moisture Data

Abstract

Spaceborne Global Navigation Satellite System Reflectometry (GNSS-R) provides L-band observations suitable for soil moisture retrieval and monitoring, but the resulting datasets are inherently sparse in space and time. Many applications, however, benefit from access to seamless, gap-free soil moisture datacubes. This study introduces Neural Spatiotemporal Interpolation (NSTI), a deep learning framework for gap filling of GNSS-R-derived soil moisture spatiotemporal datacubes. NSTI formulates gap filling as a regression problem in which missing values are estimated from known values in their spatiotemporal neighborhoods through a single, spatially transferrable neural network model. The framework is evaluated on CYGNSS-derived surface soil moisture aggregated at 36 km and 9 km resolutions. Across both resolutions, NSTI achieves gap-filling errors comparable to state-of-the-art approaches while requiring substantially fewer trainable parameters. Validation against data from SMAP (ubRMSE 0.049 m³/m³) and ERA5-Land (ubRMSE 0.056 m³/m³) further shows that the gap-filled datacubes preserve agreement with external reference datasets, when compared to the original sparse product. Together, these results demonstrate that spatiotemporal soil moisture structure can be consolidated into a unified and scalable deep learning framework for gap filling of sparse GNSS-R data.