A comparison of geographic datasets and field measurements to model soil carbon using random forests and stepwise regressions (British Columbia, Canada)

We used geographic datasets and field measurements to examine the mechanisms that affect soil carbon (SC) storage for 65 grazed and non-grazed pastures in southern interior grasslands of British Columbia, Canada. Stepwise linear regression (SR) modeling was compared with random forest (RF) modeling. Models produced with SR performed better than those produced using RF models (r2 = 0.56–0.77 AIC = 0.16–0.30 for SR models; r2 = 0.38–0.53 and AIC = 0.18–0.30 for RF models). The factors most significant when predicting SC were elevation, precipitation, and the normalized difference vegetation index (NDVI). NDVI was evaluated at two scales using: (1) the MOD 13Q1 (250 m/16-day resolution) NDVI data product from the moderate resolution imaging spectro-radiometer (MODIS) (NDVIMODIS), and (2) a handheld multispectral radiometer (MSR, 1 m resolution) (NDVIMSR) in order to understand the potential for increasing model accuracy by increasing the spatial resolution of the gridded geographic datasets. When NDVIMSR data were used to predict SC, the percentage of the variance explained by the model was greater than for models that relied on NDVIMODIS data (r2 = 0.68 for SC for non-grazed systems, modeled with SR based on NDVIMODIS data; r2 = 0.77 for SC for non-grazed systems, modeled with SR based on NDVIMSR data). The outcomes of this study provide the groundwork for effective monitoring of SC using geographic datasets to enable a carbon offset program for the ranching industry.