Prediction of Soil Carbon in the Conterminous United States: Visible and Near Infrared Reflectance Spectroscopy Analysis of the Rapid Carbon Assessment Project

Abstract

Visible and near infrared reflectance spectroscopy (VisNIr) has been used across a number of spatial scales to predict soil organic carbon (OC) content. The rapid Carbon Assessment Project (raCA) is a nationwide project that collected 144,000+ soil samples from across the conterminous United States for C stock mapping using VisNIr. The objective of this study was to calibrate and validate the VisNIr soil OC and total C (TC) models with ~20,000 samples from raCA. Models were developed with either partial least squares regression (PlSr) or the Artificial Neural Network (ANN) model. Four auxiliary variables-raCA region, land Use land Cover, Master Horizon, and Textural Class-were tested to stratify the dataset and to develop local models. The results showed that OC and TC models calibrated with ANN (R 2 > 0.94; rPD > 4.0) outperformed those of PlSr (R 2 = 0.83; rPD = 2.4). For PlSr, local models developed with all four auxiliary variables exhibited improvement in prediction accuracy, and the improvement was only marginal for ANN models. Master Horizon and Textural Class appeared to be more effective in stratifying samples into homogeneous groups because they gave an overall lower root mean squared error of prediction (rMSE P ) for the validation samples. For the majority of the local Textural Class models, the rMSE P of OC prediction ranged from 0.5 to 1.5%. To maximize the applicability of the raCA spectral library on external soil samples, PlSr local models developed from Master Horizon or Textural Class appeared to be more favorable.Abbreviations: ANN, Artificial Neural Network; IC, inorganic carbon; LULC, land use land cover classes; OC, organic carbon; PLSR, partial least squares regression; RaCA, Rapid Carbon Assessment Project; RPD, ratio of performance to deviation; RPIQ, ratio of performance to interquartile range; TC, total carbon; VisNIR, visible and near infrared reflectance spectroscopy. S oil organic carbon (OC) is a key soil property that plays many critical roles, from agriculture production to biogeochemical cycling to ecosystems functioning (Adhikari and Hartemink, 2016;Chen et al., 2000;Lal, 2004; van Wesemael et al., 2010). The capability of soils to sequester C and therefore regulate atmospheric CO 2 concentration and mitigate climate change is widely recognized and has been an active area of research (Grunwald et al., 2011;Lal, 2004;West and Post, 2002). Up-to-date, baseline soil C stock maps across different scales are a very useful tool for researchers, stakeholders, and policymakers for a wide variety of applications, ranging from best land management practices to natural resource conservation to C auditing (de Gruijter et al., 2016;Minasny et al., 2011).Producing large-scale (national or continental) soil C maps (e.g., for the purpose of C inventory and understanding of soil C sourcing and sinking) is highly challenging. Many previous studies have used legacy soil data (Aitkenhead and Coull, 2016;Minasny et al., 2013;Mulder et al., 2016 Core Ideas• We modeled ~20,000 samples in raCA from...

Soil Organic Carbon Across Mexico and the Conterminous United States (1991–2010)

“…Our simulated annealing analysis highlights SOC relationships (positive and negative) with climate variables (precipitation and land surface temperature), elevation (and terrain derivatives), and vegetation greenness (productivity and seasonality) that are consistent with previous literature describing SOC drivers across diverse environmental conditions (Evans et al, 2011;Hobley et al, 2015). In addition, when applying the simulated annealing framework to the independent datasets, a MODIS surface reflectance short wave infrared band (M06MOD4; Table S1) was one of the first five important variables predicting SOC, which is consistent with the infrared based methods used by the USDA for developing of the RaCA dataset (Wijewardane et al, 2016). The prediction capability of the infrared spectra (e.g., near infrared and midinfrared) for SOC can be attributed to the strong spectral absorption characteristics of soil organic matter and BD (the main components of SOC) in the infrared spectral bands .…”

“…We calculated model residuals against two fully independent datasets across both countries ( n =9239). Across CONUS we used 6,179 SOC estimates (2010) from the Rapid Carbon Assessment Project (RaCA, Soil Survey Staff and Loecke, ; Wijewardane et al, ) and 3,060 (2009–2011) SOC estimates from top soil samples extracted from the Mexican National Forest and Soils Inventory of the Mexican Forest Service (2009–2011; Figure S2). These independent datasets have been collected using different sampling designs and using different SOC calculation methods from our initial training dataset (INEGI and ISCN).…”

“…For example, SOC predictions and estimates across Mexico and the conterminous United States (CONUS) are available from a variety of methods and in different formats. These include soil type polygon maps, field observations, and reflectance spectroscopy analysis, as well as global SOC variability gridded surfaces based on environmental correlation methods (e.g., Bliss et al, ; Hengl et al, ; Hengl et al, ; Wijewardane et al, ). Further examples include the use of linear geostatistics for the interpolation of SOC across Mexico (Cruz‐Cárdenas et al, ) and SOC modeling efforts across the United States (Padarian et al, ).…”

A genetic algorithm‐based stacking algorithm for predicting soil organic matter from vis–NIR spectral data

“…The present study accomplished a similar decrease without using additional information. A performance of R 2 = 0.50 was reported in the Chinese vis–NIR SSL (Shi et al, ), whereas the results for the global SSL gave an R 2 of 0.57 (Ramirez‐Lopez et al, ); the best results in the literature were reported in the USA with R 2 = 0.94 (Wijewardane et al, ). A detailed review of results reported in the literature is given in Viscarra Rossel et al ().…”

Mid‐infrared spectroscopy for prediction of soil health indicators in the United States

“…While this methodology is an excellent way of increasing sampling density without greatly increasing analytical costs, the models may not be broadly applicable. Several research groups have been working to build large national and even international spectral libraries and associated databases with the intention of turning diffuse reflectance spectroscopy into a routine production tool (Baldock, Hawke, Sanderman, & Macdonald, 2014;Dangal, Sanderman, Wills, & Ramirez-Lopez, 2019;Nocita et al, 2014;Terhoeven-Urselmans, Vagen, Spaargaren, & Shepherd, 2010;Viscarra Rossel et al, 2016;Wijewardane, Ge, Wills, & Loecke, 2016, 2018.…”