This study aimed to assess the contribution of RADARSAT-2 polarimetric synthetic aperture radar (SAR) data to digital mapping of soil surface texture of Rouville County, near Montreal, Canada. First, compositional data transformation using isometric log ratio (ilr) was applied to soil texture data to transpose the simplex into multidimensional real space, which is better suited to multivariate analysis. Thereafter, two assumptions on the relationships between RADARSAT-2 data and the ilr components were evaluated: (i) as linearly dependent by applying cokriging (CK) and regression kriging (RK); and (ii) as nonlinearly dependent by applying the ε-insensitive and nonlinear support vector regression (SVR). The results were compared with ordinary kriging (OK). The environmental variables used to define the covariates were monopolarization SAR channels, the parameters extracted from entropy/anisotropy/ mean α polarimetric, Freeman and Durden, and Touzi decompositions. Using 283 soil samples for training and 89 for validation, the results showed that the root mean square error (RMSE) of prediction obtained by OK was 8% for silt, 13% for clay, and 13% for sand. The SVR produced the best prediction accuracy compared with the geostatistical interpolation techniques. Compared with OK, the improvement of the digital mapping accuracies (in terms of RMSE reduction) with SVR was significant: RMSE was reduced by 18% for sand, 17% for silt, and 35% for clay. It was followed by RK, with RMSE reduction ranging from 6 to 13%, and then CK (3-5%). Using SAR polarimetric data extracted from a RADARSAT-2 image as covariates was found to be very useful for digital mapping of soil surface texture.
ASJC Scopus subject areas
- Soil Science