Series or Parallel? An Exploration in Coupling Physical Model and Machine Learning Method for Disaggregating Satellite Microwave Soil Moisture

Remotely sensed soil moisture (SM) dataset with well accuracy and fine spatiotemporal resolution is very valuable in various fields. Downscaling is a promising way to obtain such an SM dataset. There are currently two basic methodologies in the downscaling with satellite datasets, i.e., the machine learning (ML) methods and the physical or semiphysical (PH) models. This study focuses on exploring feasible ways to integrate them for boosting the performance of downscaling. Here, three parallel modes, i.e., arithmetic average (Ari), geometric average (Geo), and weighted average (Wei), and two series modes, i.e., first PH then ML ( PH-ML) and first ML then PH (ML-PH), were developed and evaluated. A representative of PH, DSCALE_mod16, and three ML algorithms, random forest (RF), extreme gradient boosting (XGBoost), and LightGBM, were used in this study. Microwave SM datasets from SM Active and Passive (SMAP) Mission, satellite datasets from Moderate Resolution Imaging Spectroradiometer (MODIS), and in situ SM observations spanning from April 2015 to December 2018 were employed in the evaluation. Spatial dynamic range, energy conservation, and precision preservation analyses were conducted as evaluation methods. Results demonstrated that the PH-ML series mode outperformed the other coupling modes, which was even better than the better one of the PH and ML. The PH can first generate a valuable initial estimate, and the initial estimate's advantages can be preserved, while its errors can be suppressed by the ML. Resultantly, the ideology of first using the PH to obtain an initial estimate and then putting the initial estimate into the ML is suggested for further microwave SM downscaling.

Automated summary

This study focuses on exploring the integration of machine learning methods and physical or semiphysical models to improve downscaling performance. The results demonstrate that the PH-ML series mode, which involves obtaining an initial estimate from PH and then inputting it into ML, outperforms other coupling modes.