Transferability and Scaling of VNIR Prediction Models for Soil Total Carbon in Florida

Abstract

The assessment of soil total carbon (TC) across large land areas is critical to derive global and regional soil carbon budgets and better understand the interactions between carbon and other biogeochemical cycles. But the cost and time involved in measurements of TC with standard laboratory methods are impractical. Research has suggested that visible/near-infrared (VNIR) diffuse reflectance spectroscopy can provide robust and accurate estimations for TC. The applicability, transfer, and scalability of VNIR-derived soil models are still poorly understood. The objectives of this study in Florida, USA, were to (i) compare two methods to predict soil TC using five fields (local scale) and a pooled (regional scale) VNIR spectral dataset, (ii) assess the model’s transferability among fields, and (iii) evaluate the up- and downscaling behavior of TC prediction models. A total of 560 TC-spectral sets were modeled by partial least squares regression (PLSR) and support vector machine (SVM). The transferability and up- and downscaling of 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 both methods (PLSR and SVM) with R 2 > 0.86, bias < 0.01 %, root-mean-square prediction error (RMSE) = 0.09 %, residual predication deviation (RPD) > 2.70 %, and ratio of prediction error to inter-quartile range (RPIQ) > 4.54. PLSR performed substantially better than SVM to scale and transfer models. Upscaled soil TC models performed somewhat better in terms of model fit (R 2), RPD, and RPIQ, whereas downscaled models showed less bias and smaller RMSE based on PLSR. But no universal trend was found indicating which of the four investigated factors (i–iv) had the most impact that constraints transferability and scalability. The findings from this study have implications for the development of ‘universal’ spectral-based soil models aiming to predict soil properties for a diverse set of different soils formed in different environmental conditions covering a wide range of geographic settings, at its extreme the whole globe. Those ‘universal’ spectral libraries are based on the premise that soil predictions (e.g., soil TC) can be made anyplace because they are built using soil spectral datasets that characterize exhaustively the attribute feature space. This assertion is limited by the fact that a large number of interacting factors of soils, spectra, and environmental properties are needed to represent the exhaustive sample population which has not materialized yet. Given the many factors that can impinge on empirically derived soil spectral prediction models, as demonstrated by this study, more focus on the applicability and scaling of them is needed.