Search

Use of X‐ray fluorescence (XRF) spectrometry for estimation of soil texture, pH, and cation exchange capacity (CEC) is desirable given the time‐consuming nature of traditional methods. Recent studies have shown promising results; however, further investigation is required to determine the effects of sample preparation and data evaluation techniques on accuracy. Our objective was to compare (I) a simple but well‐founded approach, combining measurement of powder samples and modeling with elemental contents as predictors in stepwise multiple linear regressions (MLR), with alternative approaches including (II) use of partial least squares regression (PLSR), (III) sample preparation as a pressed pellet, and (IV) spectral intensities as predictors (20 kV, 40 kV, and concatenated 20 + 40 kV). A total of 395 loess soils from three arable fields were used with a fivefold random training‐testing approach and a hold‐one‐site‐out training‐testing approach. With random partitioning, clay, silt, and sand accuracy with approach I was excellent (ratio of performance to interquartile distance in validation (RPIQv) = 8.5–12.9), while pH and CEC estimations were satisfactory to excellent (RPIQv = 2.0–2.5 and 2.2–3.3, respectively). Differences between MLR and PLSR were negligible. Increases in accuracy with pellet samples were 1%–13% of RPIQv for 20 kV intensities, but effects were inconsistent for other predictors. The optimal predictor varied by property, and differences ranged from 3% to 13% of RPIQv. Improvements to accuracy from Approach I to the best alternative were largest for texture (10%–15%) but may be superfluous given the excellent accuracy across all approaches. Although the leave‐one‐site‐out training resulted in variable performance, inclusion of soils from the target site in training assured reliable accuracy. Core Ideas: With random partitioning into training and test sets, XRF estimation of soil texture, pH, and cation exchange capacity was successful.Effects of preparation (powder vs. pellet), predictor (elements vs. spectra), and algorithm (multiple linear regression vs. partial least squares regression) were minor.Accuracy improvements with optimized approaches were largest for texture prediction but may be superfluous.Inclusion of soils from the target site in training improved the reliability of models. [ABSTRACT FROM AUTHOR]

Rapid estimation of the amount and stability of soil organic carbon (SOC) by infraredspectroscopy has been proposed as a means to inform land use decision-making for theoptimization of environmental services provided by soils. Our research tests theviability of this strategy by determining the accuracy and robustness of modelspredicting total and labile SOC content using near (NIRS) and middle (MIRS) infraredspectroscopy. Soil samples were collected from the 0-20 cm depth of arable fields inGermany with variable texture and SOC contents. Seven treatments were applied to20 soils, including application of 2, 4 or 8 g C per kg soil in the form of straw orclover residues, as well as a control treatment without residues. Samples were thenincubated for 56 days at 20˚ C and cumulative CO2 emissions were determined.Additionally, SOC, total nitrogen (Nt), NIRS and MIRS were measured on dried, milledsamples before and after incubation, as well as microbial biomass (Cmic) afterincubation. Partial least square regression (PLSR) and support vector machine regression,both with and without mathematical pretreatments, were compared as a means toestimate SOC, Nt, Cmic and labile C, quantified as the residue application rate,cumulative CO2 loss during incubation, or the difference in bulk SOC before and afterincubation. Separate models for straw and clover were created, both calibrated andcross-validated using control and 8 g C kg−1 soil residue treatments (n=40), and validatedusing 2 and 4 g C kg−1 soil residue treatments (n=40). While PLSR validationestimations of bulk SOC and Nt by both MIRS and NIRS can be categorized as good toexcellent according to the ratio of performance to interquartile distance, estimations oflabile C ranged from not usable to approximately quantitative. Neither spectralregion consistently outperformed the other, suggesting there is value in utilizingboth NIRS and MIRS when possible. As a measure of model robustness, modelscalibrated with samples prior to incubation were also used to predict SOC and Ntafter incubation, when substantially less labile SOC was present. Compared tomathematically pretreated models optimized in cross-validation, models using untreateddata for all wavelengths were a more robust means to accurately predict total SOCand Nt when the amount of labile C differed greatly from the calibration dataset. [ABSTRACT FROM AUTHOR]