Prediction of Soil Organic Carbon over Landscape

An easily applicable 3D soil mapping approach was applied to predict SOC over landscape.
Upscaled predicted and measured soil organic carbon (SOC) densities of three different datasets compared (Figure 9)

Rong Lang and Xueqing Yang from World Agroforestry Centre co-authored a paper on the prediction of soil organic carbon across continuous soil profile depth in the mountainous subtropic region in China. The study, led by the Institute of Agricultural Sciences in the Tropics in the University of Hohenheim, used a mixed 3D modeling approach to model changes of soil properties with depth and predict the spatial distribution of soil properties at the landscape level. Main determinants of soil organic carbon were land use type, soil depth, and elevation. Subsoil carbon between a depth of 40 and 100 cm comprised >40% of SOC stocks.

Due to the spatial variability of soil resources in rapidly changing landscapes, such as rubber expansion areas in mountainous South East Asia, landscape-based soil organic carbon (SOC) stock assessments need new approaches to obtain cost-effective high-resolution soil maps. 3D modeling presents the opportunity to model changes of soil properties with soil depth and in space in one single model. While most 3D models make use of spatial autocorrelation to create soil maps, it might be feasible for upscaling to neglect the spatial autocorrelation and only model autocorrelation within the soil profiles. We propose a “mixed model over continuous depth” (MMCD), which uses a linear and quadratic term to model changes of soil properties with depth and predicts the spatial distribution of soil properties at the landscape level. As the study area of 43 km2 in South West China was subject to multiple constraints such as sparse road networks, steep terrain, and poor infrastructure, we applied the cost-constrained conditioned Latin hypercube sampling (CCLHS) scheme for soil sampling at 120 locations to a depth of 1 m. The MMCD provides information on the most important drivers of selected soil properties and their relative importance. In this study, SOC was strongly linked to an interaction of elevation with mean horizon depth (p < 0.001) and to the land use type (p < 0.001). An iterative leave-one-third-out evaluation was performed to compare the MMCD to several established 2D and 3D mapping approaches. The MMCD proofed to be as powerful as these established techniques, with an overall modeling efficiency (EF) of 0.72. All tested models had a strong decrease in accuracy with depth, from an EF of about 0.8 in the topsoil to 0.2 at 0.8 to 1m subsoil depth. The MMCD was further used to model highly unbalanced SOC density data with 120 independent topsoil observations and only 11 locations with subsoil observations (EF of 0.75), where the computed prediction intervals (95%) accurately covered the range of legacy measurements. Our approach allowed upscaling of SOC density predictions to the surrounding larger nature reserve of 270 km². The resulting MMCD and 3D maps revealed that on average, 15 and 10% of SOC stocks are expected in the 0.6 to 0.8m and 0.8 to 1m soil depth intervals, respectively. The combination of CCLHS and MMCD is particularly suitable for mountainous subtropical areas with poor road networks. However, this approach requires a strong relationship with the soil property of interest with explanatory environmental covariates, as it does not consider spatial autocorrelation for soil mapping. The advantage of this restriction is that it is easy to apply to highly unbalanced datasets and easy to upscale, given that the environmental covariates in the surrounding area are similar to the calibration area.

Laub, M., Blagodatsky, S., Lang, R., Yang, X., & Cadisch, G. (2018). A mixed model for landscape soil organic carbon prediction across continuous profile depth in the mountainous subtropics. Geoderma330, 177-192.