A robust gap-filling approach for European Space Agency Climate Change Initiative (ESA CCI) soil moisture integrating satellite observations, model-driven knowledge, and spatiotemporal machine learning

Abstract

Spatiotemporally continuous soil moisture (SM) data are increasingly in demand for ecological and hydrological research. Satellite remote sensing has potential for mapping SM, but the continuity of satellite-derived SM is hampered by data gaps resulting from inadequate satellite coverage, snow cover, frozen soil, radio-frequency interference, and so on. Therefore, we propose a new gap-filling approach to reconstruct daily SM time series using the European Space Agency Climate Change Initiative (ESA CCI). The developed approach integrates satellite observations, model-driven knowledge, and a machine learning algorithm that leverages both spatial and temporal domains. Taking SM in China as an example, the reconstructed SM showed high accuracy when validated against multiple sets of in situ measurements, with a root mean square error (RMSE) and a mean absolute error (MAE) of 0.09-0.14 and 0.07-0.13ĝ€¯cm3ĝ€¯cm-3, respectively. Further evaluation with a 10-fold cross-validation revealed median values of the coefficient of determination (R2), RMSE, and MAE of 0.56, 0.025, and 0.019ĝ€¯cm3ĝ€¯cm-3, respectively. The reconstructive performance was noticeably reduced both when excluding one explanatory variable and keeping the other variables unchanged and when removing the spatiotemporal domain strategy or the residual calibration procedure. In comparison with gap-filled SM data based on a satellite-derived diurnal temperature range (DTR), the gap-filled SM data from bias-corrected model-derived DTRs exhibited relatively lower accuracy but higher spatial coverage. Application of our gap-filling approach to long-Term SM datasets (2005-2015) produced a promising result (R2Combining double low line0.72). A more accurate trend was achieved relative to that of the original CCI SM when assessed with in situ measurements (i.e., 0.49 versus 0.28, respectively, in terms of R2). Our findings indicate the feasibility of integrating satellite observations, model-driven knowledge, and spatiotemporal machine learning to fill gaps in short-and long-Term SM time series, thereby providing a potential avenue for applications to similar studies.