Enhancing Global Surface Soil Moisture Estimation From ESA CCI and SMAP Product With a Conditional Variational Autoencoder

Abstract

High-quality soil moisture (SM) estimation is crucial for various applications, including drought monitoring, environmental assessment, and agricultural management. Advances in remote sensing technology have enabled the retrieval of near real-time Earth surface SM using both active and passive sensors. However, the ESA climate change initiative (CCI) SM product, which combines data from multiple sensors, sacrifices spatial-temporal resolution and coverage due to satellite orbit constraints and retrieval algorithms. To address this issue, an SM reconstruction approach based on a conditional variational autoencoder model was developed, leveraging the high spatial resolution of SMAP L4 data and the accuracy of CCI fused products across different land cover types. This method resulted in the creation of a global three-day SM product at 0.0625$^{\circ }$ spanning from 2015 to 2021. The reconstructed SM product underwent rigorous validation against global core SM sites and sparse observation networks. The evaluation employed multiple metrics, including the global unbiased root mean square error (ubRMSE) and correlation coefficient (CC). The validation yielded results, with ubRMSE values of approximately 0.029 and 0.071 m$^{3}$/m$^{3}$, and CC values of around 0.863 and 0.743 for core SM sites and sparse observation networks. This reconstructed product offers global coverage and enhanced accuracy compared to existing benchmarks.

Keywords

• Conditional variational autoencoder (CVAE)
• ESA climate change initiative (CCI)
• product reconstruction
• SMAP L4
• surface soil moisture (SM)