Generating Surface Soil Moisture at 30 M Spatial Resolution Using Both Data Fusion and Machine Learning Toward Better Water Resources Management at the Field Scale
Ahmed Samir Abowarda,Liangliang Bai,Caijin Zhang,Di Long,Xueying Li,Qi Huang,Zhangli Sun
DOI: https://doi.org/10.1016/j.rse.2021.112301
IF: 13.5
2021-01-01
Remote Sensing of Environment
Abstract:Soil moisture has a considerable impact on the hydrological cycle, runoff generation, drought development, and water resources management. Soil moisture products provided by passive microwave remote sensing possess coarse spatial resolutions ranging from 25 to 50 km, unable to reflect large spatial heterogeneity in soil moisture caused by complex interactions among meteorological forcing, land cover, and topography. Meanwhile, active microwave remote sensing can provide higher spatial resolution than passive sensors that may reach 1 km but with lower temporal resolution of 6-12 days (e.g., Sentinel-1). Better water resources management, particularly for the agricultural sector, requires spatiotemporally continuous soil moisture estimates at the field scale (e.g., 30 m x 30 m) to reflect its high spatiotemporal variability across heterogeneous land surfaces. In this study, both data fusion and random forest models along with a range of remote sensing, reanalysis, and in situ data were jointly used to generate spatiotemporally continuous surface soil moisture (SSM) at 30 m x 30 m. First, both daily normalized difference vegetation index (NDVI) and surface albedo at 30 m x 30 m were generated by fusing reflectance products of the MODerate resolution Imaging Spectroradiometer (MODIS) and Landsat. Second, China Meteorological Administration Land Data Assimilation System (CLDAS, 0.0625 degrees x 0.0625 degrees) land surface temperature (LST) was fused with both MODIS LST and Landsat LST to generate spatially complete LST maps (at 11 a.m. local solar time for each day) at 30 m x 30 m. Last, random forest models were developed to generate spatiotemporally continuous SSM of 30 m x 30 m using the fused variables at fine spatial resolution (e. g., NDVI, surface albedo, and LST), SSM background fields, and ancillary variables such as precipitation and soil texture as the model inputs. Compared with original SSM of the European Space Agency (ESA) Climate Change Initiative (CCI) Version 4.4 SSM, Soil Moisture Active Passive (SMAP) Level-4 SSM, and CLDAS SSM, the downscaled SSM using these products as background fields was improved significantly in terms of accuracy and spatial distribution. Moreover, the integration of multiple SSM background fields improved the performance of the downscaled SSM significantly in terms of spatiotemporal consistency and accuracy compared with that using a single SSM background field. Overall, the downscaled SMAP_L4 + CLDAS SSM showed the best performance at four sites (i.e., Weishan, Huailai, Hujiatan, and Paihuai) out of seven sites on the North China Plain with R, bias, MAE, RMSE, and ubRMSE ranging from 0.70-0.84, -0.034-0.012 cm (3)/cm(3), 0.025-0.044 cm(3)/cm(3), 0.031-0.050 cm(3)/cm(3), and 0.022-0.042 cm(3)/cm(3), respectively. The proposed framework maximizes the potential of data fusion, random forest models, and in situ data in deriving spatiotemporally continuous SSM estimates at 30 m x 30 m, which should be valuable for water resources management at the field scale.