Abstract

A soil moisture estimation method was developed for Sentinel-1 synthetic aperture radar (SAR) ground range detected high resolution (GRDH) data to analyze moisture conditions in a gently undulating and heterogeneous subarctic area containing forests, wetlands, and open orographic tundra. In order to preserve the original 10-m pixel spacing, PIMSAR (pixel-based multitemporal nonlocal averaging) nonlocal mean filtering was applied. It was guided by multitemporal statistics of SAR images in the area. The gradient boosted trees (GBT) machine learning method was used for the soil moisture algorithm development. Discrete and continuous in situ soil moisture values were used for training and validation of the algorithm. For surface soil moisture, the root mean square error (RMSE) of the method was 6.5% and 8.8% for morning and evening images, respectively. The corresponding maximum errors were 34.1% and 33.8%. The pixelwise sensitivity to the training set and method choice was estimated as the variance of the soil moisture values derived using the algorithms for the three best methods with respect to the criteria: the smallest maximum error, the smallest RMSE value, and the highest coefficient of determination ( R2 ) value. It was, on average, 6.3% with a standard deviation of 5.7%. Our approach successfully produced instantaneous high-resolution soil moisture estimates on daily basis for the subarctic landscape and can further be applied to various hydrological, biogeochemical, and management purposes.