Determination of high-resolution soil texture profile at the hillslope scale

Abstract

Large-scale quantification of soil physical properties is challenging due to their inherent spatial and temporal variability. This variability determines hydrologic and biogeochemical behavior of soils and influences their ecosystem responses. This study investigated electrical resistivity tomography (ERT) in an Artificial Neural Network (ANN) framework to test spatially variable linear and non-linear models to predict in-situ soil textures up to a depth of 6 m for the Piedmont Physiographic Region in Georgia. Soil resistivity was measured for several months across a hillslope and used as input for the ANN along with relative depth of investigation and lake levels, since our study sites were located on a shoreline. The models were cross validated and tested at independent sites. Overall, soil texture was strongly correlated to depth (-0.7 to 0.85) while resistivity had a significant yet weaker correlation (-0.27 to 0.33). The correlation between depth and texture changed markedly in the saprolite layers such that ANN-based models performed better than multivariate regression, capturing the linear relationship between depth and texture in the shallow layers and the correlation between resistivity and texture for deeper depths. Within the ANN framework, depth alone generated an R2 = 0.55 and RMSE = 9.12 % for an independent site validation and resistivity improved the metrics marginally to 0.6 and 8.47 %, respectively. The R2 and RMSE for the deeper saprolite depths, however, improved from 0.01 and 6.36 % to 0.31 and 5.94 %, respectively. The predictions were also robust when based on a single ERT survey and time of data collection (i.e., dry vs wet season). The ANN model predictions were sensitive to the structure of model inputs with the best model importing resistivity and lake levels first, and the second layer processing the obtained information together with depth. These ERT based ANN models provided finer information than what is currently available from SSURGO while covering deeper portions of the subsurface and maintaining high vertical resolution.