Spectral Geometry-Informed Neural Interpolation for GNSS-R Soil Moisture Retrieval

Abstract

GNSS-Reflectometry (GNSS-R) is a promising source for soil moisture (SM) information, but its sparse and irregular sampling makes interpolation difficult, particularly near sharp spatial discontinuities. This paper presents a geometryinformed neural interpolation framework for reconstructing SM fields from sparse and irregular GNSS-R observations. The proposed method combines spectral positional encoding, graphbased geometric regularization, and edge-aware refinement to better preserve sharp spatial transitions during reconstruction. Within this framework, Sobel and Scharr operators are evaluated as inference-time gradient estimators. Experiments on CYGNSS SM data, using SMAP as ground truth, showcase that while both models yield accurate SM field reconstructions, the Scharrbased variant, namely LB-PE Scharr, is better suited for high-quality and sharper edge enhancement, with a slightly lower interpolation error than the Sobel-based counterpart. The results highlight the benefit of Scharr filtering for interpolation of GNSS-R irregular data.