Advancing ASCAT soil moisture retrievals: benchmarking neural-network models and exploring intraday estimation potential

Soil moisture (SM) is a key variable in land-atmosphere interactions, and numerous efforts aim to produce consistent large-scale SM datasets. Satellite-based retrievals provide valuable complements to physically based approaches, particularly for achieving global coverage. Neural networks (NNs) have demonstrated strong potential for improving SM retrieval accuracy in recent years. This study benchmarks daily SM retrievals from Advanced SCATterometer (ASCAT) observations using multiple NN-based architectures, with varying degrees of localization, a strategy designed to help the models adapt to local conditions. Two model families are evaluated: multilayer perceptions (MLPs) and convolutional neural networks (CNNs). We examine configurations that incorporate physical variable augmentation, geographic coordinate inputs, and explicitly localized designs (pixel-scale MLP and locally-connected CNN) to assess the sensitivity of the model accuracy to input nature and localization strength. In non-localized settings, CNNs consistently yield higher spatial and temporal correlations, reflecting their ability to learn spatial hierarchies and local patterns. In strongly localized designs, the pixel-scale MLP and locally-connected CNN achieve very high overall correlations with substantially reduced local bias, highlighting the value of localized learning for capturing fine‑scale SM variability. In addition to providing improved daily SM estimates, our CNN-based retrieval can also capture intraday variability. This capability is particularly evident during intense precipitation events, offering new perspectives into short-term hydrological dynamics. Looking ahead, future efforts should focus on integrating complementary satellite measurements from other sensors (SMOS, SMAP, AMSR, CIMR) to further improve retrieval accuracy, robustness, and temporal resolution.