Transferability and Scalability of Soil Total Carbon Prediction Models in Florida, USA

The applicability, transferability, and scalability of visible/near-infrared (VNIR)-derived soil total carbon (TC) models are still poorly understood. The objectives of this study were to: i) compare models of three multivariate statistical methods, partial least squares regression (PLSR), support vector machine (SVM), and random forest methods, to predict soil logarithm-transformed TC (logTC) using five fields (local scale) and a pooled (regional-scale) VNIR spectral dataset (a total of 560 TC spectral datasets), ii) assess the model transferability among fields, and iii) evaluate their up- and downscaling behaviors in Florida, USA. The transferability and up- and downscaling of the models were limited by the following factors: i) the spectral data domain, ii) soil attribute domain, iii) methods that describe the internal model structure of VNIR-TC relationships, and iv) environmental domain space of attributes that control soil carbon dynamics. All soil logTC models showed excellent performance based on all three methods with R-2 > 0.86, bias < 0.01%, root mean squared error (RMSE) = 0.09%, residual predication deviation (RPD) > 2.70%, and ratio of prediction error to interquartile range (RPIQ) > 4.54. The PLSR method performed substantially better than the SVM method to scale and transfer the TC models. This could be attributed to the tendency of SVM to overfit models, while the asset of the PLSR method was its robustness when the models were validated with independent datasets, transferred, and/or scaled. The upscaled soil TC models performed somewhat better in terms of model fit (R-2), RPD, and RPIQ, whereas the downscaled models showed less bias and smaller RMSE based on PLSR. We found no universal trend indicating which of the four limiting factors mentioned above had the most impact that constrained the transferability and scalability of the models. Given that several factors can impinge on the empirically derived soil spectral prediction models, as demonstrated by this study, more focus on their applicability and scalability is needed.