Refinement of NOAA AMSR-2 Soil Moisture Data Product: 1. Intercomparisons of the Commonly Used Machine-Learning Models

Soil moisture (SM) is an important variable for hydrological, meteorological, and agricultural studies and applications. The SM Operational Products System (SMOPS) was developed by the National Oceanic and Atmospheric Administration (NOAA)-National Environmental Satellite, Data, and Information Service (NESDIS) to operationally provide an integrated satellite SM data product. The Advanced Microwave Scanning Radiometer 2 (AMSR2) SM retrieval is an important component of the currently operational SMOPS. This study is proposed to refine the AMSR2 data product using an optimal machine-learning model, and this first paper of the two-part series is to intercompare the six commonly used machine-learning models including multiple linear regression (MLR), regression tree (RRT), random forest (RFT), gradient boosting (GBR), extreme gradient boosting (XGB), and artificial neural network (ANN). Results indicate that all of the six approaches can preserve the reference data information beyond the training time period, which ensures they predict past and future satellite retrievals without a new training procedure. Relative to other models, the XGB method is more successful to respect to the reference data SM Active Passive (SMAP) and the in situ observations from the U.S. Department of Agriculture Soil Climate Analysis Network (SCAN). It has a good implication for the implementation of the XGB model to refine the AMSR2 SM retrievals in the second paper.

Automated summary

This study proposes an optimal machine-learning model to refine the AMSR2 data product, and compares six commonly used machine-learning models. The XGB method is found to be more successful in preserving the reference data information compared to other models.