SM-RDC: A Three-Step Downscaling Framework for Daily 1-km Seamless SMAP Product Generation Over the Tibetan Plateau

Soil moisture (SM) is a crucial variable for regulating global climate change. However, SM products derived from microwave remote sensing often suffer from low resolution and incomplete coverage, hindering their use in regional hydrology and precision agriculture. To address these challenges, this study developed a three-step downscaling framework, SM-RDC, which consists of processes of reconstruction, downscaling and calibration. In the reconstruction step, a temporal-spatial 3D convolutional (TS3D) network was proposed to fill gaps in the 9 km SMAP time series by leveraging high-resolution auxiliary data, producing continuous, high-quality SM labels. For downscaling step, an Attention-CNN (ACNN) model incorporating the convolutional block attention module was designed to robustly extract spatial and channel features from eleven auxiliary variables, covering meteorology, climate, vegetation, energy, land cover, topography, and soil properties, to predict 1-km SMs. In the calibration step, a residual correction method was applied to refine the downscaled results. The framework was tested by downscaling the SMAP product from 9km to 1km resolution over the Tibetan Plateau. Results illustrated that the proposed SM-RDC framework is capable of producing daily 1-km SMs with high accuracy and seamless spatial coverage, with an average ubRMSE of 0.056 compared to in-situ observation data, and an average ubRMSE of 0.036 against the original SMAP data, demonstrating its potential for enhancing climate monitoring and hydrological applications.