High resolution 3D mapping of soil organic carbon in a heterogeneous agricultural landscape

Soil organic carbon (SOC) is a key element of agroecosystems functioning and has a crucial impact on global carbon storage. At the landscape scale, SOC spatial variability is strongly affected by natural and anthropogenic processes and linear anthropogenic elements (such hedges or ditches). This study aims at mapping SOC stocks distribution in the A-horizons for a depth up to 105 cm, at a high spatial resolution, for an area of 10 km(2) in a heterogeneous agricultural landscape (North-Western France). We used a data mining tool, Cubist, to build rule-based predictive models and predict SOC content and soil bulk density (BD) from a calibration dataset at 8 standard layers (0 to 7.5 cm, 7.5 to 15 cm, 15 to 30 cm, 30 to 45 cm, 45 to 60 cm, 60 to 75 cm, 75 to 90 cm and 90 to 105 cm). For the models calibration, 70 sampling locations were selected within the whole study area using the conditioned Latin hypercube sampling method. Two independent validation datasets were used to assess the performance of the predictive models: (i) at landscape scale, 49 sampling locations were selected using stratified random sampling based on a 300-m square grid; (ii) at hedge vicinity, 112 sampling locations were selected along transects perpendicular to 14 purposively chosen hedges. Undisturbed samples were collected at fixed depths and analysed for BD and SOC content at each sampling location and continuous soil profiles were reconstructed using equal-area splines. Predictive environmental data consisted in attributes derived from a light detection and ranging digital elevation model (LiDAR DEM), geological variables, land use data and a predictive map of A-horizon thickness. Considering the two validation datasets (at landscape scale and hedge vicinity), root mean square errors (RMSE) of the predictions, computed for all the standard soil layers (up to a depth of 105 cm), were respectively 7.74 and 5.02 g kg(-1) for SOC content, and 0.15 and 021 g cm(-3) for BD. Best predictions were obtained for layers between 15 and 60 cm of depth. The SOC stocks were calculated over a depth of 105 cm by combining the prediction of SOC content and BD. The final maps show that the carbon stocks in the soil below 30 cm accounted for 33% of the total SOC stocks on average. The whole method produced consistent results between the two predicted soil properties. The final SOC stocks maps provide continuous data along soil profile up to 105 cm, which may be critical information for supporting carbon policy and management decisions. (C) 2013 Elsevier B.V. All rights reserved.