Evaluating the use of explicitly spatial models and non-spatial models to estimate soil particle-size fractions and soil texture using Sentinel 1 and covariates

Geostatistical and non-geostatistical methods have been applied in soil science to different types of data, from field collections to proximal- and remotely sensed. However, studies addressing the use of microwave sensors for soil texture estimates are typically constrained to bare soil conditions. Here, we test radar remote sensing data and techniques (backscattering coefficient, radar vegetation index, and pol-decomposition) using four models as non-spatial (linear regression and tree-based regression) and spatial approaches (Empirical Bayesian Kriging Regression-EBKR and Cokriging-CK) to estimate soil particle size fractions and soil texture classes on topsoil layer (10 cm depth). The study area is Ireland spanning from western-eastern coast. The models’ performance was assessed with the leave-one-out cross-validation method. Results showed that CK yielded good predictions for sand, silt, and clay with Nash Sutcliffe Efficiency values of 0.9, 0.6, and 0.5, respectively. However, EBKR was a better approach to capture trends and spatial patterns in the interpolated surface. Importantly, this work provides a methodological framework to inform the design of intensive in situ soil surveys (areas of large uncertainty estimates in the models are those areas that may require detailed soil surveys). It also provides a means to estimate soil properties over large spatial areas and generate new data products for use in model-based approaches that currently employ coarse global scale soil texture products (soil security, climate, hydrological, land surface). The methods outlined here are also relevant to locations where optical remotely sensed data is limited and/or where vegetation cover is present throughout the year.