4.7 Article

True global error maps for SMAP, SMOS, and ASCAT soil moisture data based on machine learning and triple collocation analysis

Journal

REMOTE SENSING OF ENVIRONMENT
Volume 298, Issue -, Pages -

Publisher

ELSEVIER SCIENCE INC
DOI: 10.1016/j.rse.2023.113776

Keywords

Machine learning; Triple collocation analysis; Remotely sensed soil moisture; Time -varying geophysical data; Quality control

Ask authors/readers for more resources

This article discusses the importance of quantifying the accuracy of satellite-based soil moisture data and the limitations of existing statistical methods. It then fills the spatial gaps in TCA results using machine learning and provides spatially complete error maps for satellite-based soil moisture data products. Additionally, SHAP values are used to examine the impact of various environmental conditions on the quality of satellite-based soil moisture retrievals.
Quantifying the accuracy of the satellite-based soil moisture (SM) data is important for a number of key applications, such as: combining satellite-basedSM products for long-term SM analyses, assimilating SM data into land surface models, and providing quality flags to mask bad quality SM data. A range of statistical methods have been proposed to estimate error statistics for large-scale SM datasets including the: instrumental variable (IV) method, triple collocation analysis (TCA), and quadruple collocation analysis (QCDA). While requiring only two input products, the IV method also imposes an additional assumption that one input product possesses serially uncorrelated errors -thus limiting its scope compared to TC. Likewise, QCDA requires four independent SM data products that are difficult to obtain and may not always be available for analysis. Nonetheless, TCA-based methods still cannot provide truly global error maps for satellite SM products due to the limited number of independent SM products and difficulties with baseline TCA assumptions. Moreover, temporal sampling requirements for TCA are often impractical because of low SM retrieval skill in forested and arid areas - as well as in regions prone to radio frequency interference.Here, we seek to fill significant spatial gaps in TCA results using machine learning (ML) and therefore provide spatially complete error maps for the satellite-based SM data products derived from the Soil Moisture Active Passive (SMAP), Soil Moisture and Ocean Salinity (SMOS), and Advanced Scatterometer (ASCAT) systems. Furthermore, we use SHapley Additive exPlanations (SHAP) values, a model-agnostic technique for interpreting ML models, to examine the impact of various environmental conditions on the quality of satellite-based SM retrievals.Globally, and across all three products, 72.0% of missing error information in a TCA-based analysis, due to either the lack of valid data or the inability of TCA to provide reliable results, can be reconstructed from the ensemble prediction mean of the ML models. Overall, we found that 22.7% (a.m.) and 34.2% (p.m.) of the Earth'sSM dynamics (between 60 degrees S to 60 degrees N) have not been investigated properly across all three satellite missions.

Authors

I am an author on this paper
Click your name to claim this paper and add it to your profile.

Reviews

Primary Rating

4.7
Not enough ratings

Secondary Ratings

Novelty
-
Significance
-
Scientific rigor
-
Rate this paper

Recommended

No Data Available
No Data Available