Soil Texture and Organic Carbon Fractions Predicted from Near‐Infrared Spectroscopy and Geostatistics

Abstract

Core Ideas Near‐infrared spectroscopy (NIRS) was appropriate for predicting soil texture and C. Both linear and nonlinear multivariate models could be used for NIRS calibration. Soil texture was predicted with greater precision than organic C fractions. Near‐infrared spectroscopy and kriging were a useful combination for assessing spatial variation. Near‐infrared spectroscopy (NIRS) and geostatistics are relatively unexplored tools that could reduce the time, labor, and costs of soil analysis. Our objective was to efficiently determine lateral and vertical distributions of soil texture and soil organic C (SOC) fractions in an agroforestry system (a 7‐ha field) on a Coastal Plain site in North Carolina. To predict selected properties from a large number of soil samples collected from this field, NIRS was calibrated against laboratory‐determined properties. Support vector machines was a multivariate model that performed better than partial least squares to obtain greater precision with NIRS for all soil properties. To predict soil properties with precision across the field, geostatistical modeling with maximum likelihood and ordinary kriging was used. When we combined the two modeling processes, the root mean square error (RMSE) and the RMSE relative to the dataset mean (%RMSE) were 67 g kg‐1 for sand (9.3% RMSE), 34 g kg‐1 for clay (22.7% RMSE), 1.63 g kg‐1 for total organic C (26.7% RMSE), 0.67 g kg‐1 for particulate organic C (36.1% RMSE), and 24 mg CO2–C kg‐1 3 d‐1 for the flush of CO2 (29% RMSE). We conclude that the combination of NIRS and kriging produced acceptable errors and therefore could be used to predict the spatial distribution of soil texture and SOC fractions in this agroforestry system to allow efficient assessment of management changes with time and better predict small‐scale input requirements.