Evaluating consistency across multiple NeoSpectra (compact Fourier transform near‐infrared) spectrometers for estimating common soil properties.

Rapid and cost‐effective techniques for soil analysis are essential to guide sustainable land management and production agriculture. This study aimed at evaluating the performance and consistency of portable handheld Fourier‐transform near‐infrared spectrometers and the NeoSpectra scanners in estimating 12 common soil physical and chemical properties including pH; organic carbon (OC); inorganic carbon (IC); total nitrogen (TN); cation exchange capacity (CEC); clay, silt, and sand fractions; and exchangeable potassium (K), phosphorus (P), calcium (Ca), and magnesium (Mg). A diverse set of samples (n = 600) were retrieved from a national‐scale soil archive of the Kellogg Soil Survey Laboratory of USDA‐NRCS and scanned with five NeoSpectra scanners. Predictive models for the soil properties were developed using partial least squares regression (PLSR), Cubist, and memory‐based learning (MBL). Cubist outperformed PLSR and MBL, with the best prediction performance for clay, OC, and CEC (R2 > 0.7), followed by IC, sand, silt, and Mg (R2 > 0.6), and then pH, TN, and Ca (R2 > 0.5). K and P were predicted somewhat poorly with R2 of 0.48 and 0.22. All five NeoSpectra yielded comparable near‐infrared (NIR) spectral data and the PLSR models for the soil properties (in terms of model regression coefficients). However, the consistency assessment showed that the model performance was significantly decreased when the training and testing spectra were from different NeoSpectra scanners. It is concluded that NeoSpectra scanners could be rapid and cost effective for estimating certain soil properties, and calibration transfer should be considered for applications where multiple devices are involved and high estimation accuracy from NIR data is required. Core Ideas: Soil samples (n = 600) were retrieved from Kellogg Soil Survey Laboratory and analyzed with 5 NeoSpectra scanners.Partial least squares regression (PLSR), Cubist, and memory‐based learning (MBL) were employed to estimate 12 soil properties from NIR data.Cubist outperformed PLSR and MBL, with 10 of 12 properties having R2 > 0.5.The five NeoSpectra scanners yielded comparable NIR spectral data and PLSR models.Cross‐device assessment suggested that calibration transfer is needed to maintain higher accuracy of NIR models. [ABSTRACT FROM AUTHOR]