Trevan Flynn;Robert Kostecki;Ansa Rebi;Taqi Raz

soilgrids;carbon stock;discrete;isdasoil;sum;dataset

Soil carbon stock research has gained prominence in environmental studies, especially given its significance as one of the largest terrestrial carbon reservoirs, particularly amidst climate change concerns. Accurate predictions necessitate comprehensive soil profile measurements, which are resource intensive. To address this, depth functions are employed to derive continuous estimates, aligning with GlobalSoilMap project standards. However, global datasets employing depth functions in raster format have not been widely utilized, which could further lower financial costs and improve accuracy in data-scarce regions. Moreover, research into aggregating depth functions for total carbon stock estimation remains limited, offering opportunities to streamline the methodology. The aim of this study was to apply equal-area splines to estimate soil carbon stocks, utilizing SoilGrids and iSDAsoil (Innovative Solutions for Decision Agriculture) datasets in a quaternary catchment (317 km2) in KwaZulu-Natal, South Africa (~30° 48′ E and 29° 18′ S). Both datasets were resampled to a 250 m resolution, and the splines were interpolated to a depth of 50 cm per pixel. Various aggregation methods were employed including, the cumulative sum (definite integral), discrete sum (sum of 1 cm spline predictions) and the mean carbon stock (mean to 50 cm). Quantitative evaluation was performed with 310 external soil samples. SoilGrids showed higher predictions (100 to 546 kg) than iSDAsoil (66.9 to 225 kg m-2) for cumulative sums. Discrete sums also exhibited higher predictions for SoilGrids (293 to 789 kg m-2) compared to iSDAsoil (228 to 557 kg m-2). SoilGrids with discrete sum aggregation closely matched previous studies, estimating total carbon stock for the catchment at 5,789 to 7,126 tons, albeit with spatial variations. However, when evaluating with an external dataset, the results were not satisfactory for any method according to Lin's concordance correlation coefficient (CCC; correlation of a 1:1 line), with all models obtaining a CCC below 0.01. Similarly, the root mean squared error (RMSE), representing the difference between predicted and true values, showed all models had a RMSE greater than 58 kg m-2. It was concluded that SoilGrids and iSDAsoil are spatially inaccurate in the catchment but can still provide information into the total carbon stock content. Improvements of this method include obtaining more soil samples for the datasets, incorporating local data into the spline implementation, making the method more computationally efficient and accounting for discrete horizon boundaries.