Search

• Proximal sensors were used to determine calcic horizons in-situ. • PXRF in isolation was adept at differentiating chernozems and phaeozems. • Combined VisNIR-PXRF models modestly increased predictive model stability. The Transylvanian Plain (TP), Romania is widely used for agronomic production. The Chernisol soil class covers a vast majority of the TP as defined by the Sistemul Roman De Taxonomie A Solurilor (SRTS). Chernisols are fertile, dark soils similar to Mollisols in the United States. Chernisols have four key soil types, two of them occur in the TP: chernozems and phaeozems. While these two classifications appear similar to those found in the World Reference Base (WRB) for Soil Resources , only the Romanian system (SRTS) is applied in this paper. Chernozems require a chroma of ≤2 in the Am horizon when moist and a calcic horizon or secondary CaCO 3 within 125 cm. Phaeozems require chroma of ≤3.5 when wet and a calcic horizon or secondary CaCO 3 deeper than 125 cm. Traditionally, morphological assessment in combination with laboratory data has been used to assess the depth of secondary CaCO 3 , thus establishing the taxonomic classification. Herein, the efficacy of portable X-ray fluorescence (PXRF) and visible near infrared spectroscopy (VisNIR) were evaluated to make such determinations in lieu of laboratory data on 25 soil cores collected across five toposequences. Cores were scanned on-site with both sensors at 10 cm increments to determine depth to CaCO 3 accumulation. Comparing Ca percentages from only PXRF with traditional laboratory pressure calcimetry via simple linear regression (SLR), the following model validation data (based on whole core splitting, Core val) were obtained: R2 = 0.92; root mean squared error (RMSE) = 0.81%; residual prediction deviation (RPD) = 3.27: ratio of performance to interquartile range (RPIQ): 5.34. Thus, most of the Ca in soils of the TP is associated with secondary CaCO 3. Using the three most prominent latent variables, VisNIR spectra (smoothed to 10 nm bands) were combined with PXRF data via partial least squares regression (PLSR) to determine if any improvements could be achieved by the combined approach. Combined Core val models produced the following: R2 = 0.89; RMSE = 0.98%; RPD = 2.73; RPIQ = 4.46. Boosted regression tree Core val modeling produced similar results (R2 = 0.90; RMSE = 0.89%; RPD = 2.99; RPIQ = 4.87). With deference to the law of parsimony, use of PXRF data without VisNIR for calcic horizon identification and quantification in conjunction with morphological assessment and interpretation appears preferable for most Pedological applications given its robust, strong performance. Minimal differences were observed using two different sample splitting schemes (whole core vs. full sample set) relative to PXRF data predictive models for CaCO 3 prediction, especially for PXRF with no VisNIR contribution. Of the sites investigated, PXRF identified six phaeozems (P) and nineteen chernozems (C). Specifically, the following were identified on different slope profiles: summits (1P/4C), shoulders (1P/2C), backslopes (3P/7C), footslopes (0P/4C), and toeslopes (1P/2C). Localized landslides and erosion precluded the identification of a common landscape model differentiating P/C on landscapes. Nonetheless, proximal sensors were adept at informing soil properties needed for taxonomic classification on-site, with minimal need for traditional pressure calcimetry. [ABSTRACT FROM AUTHOR]

Abstract In this study, elemental data from portable X-ray fluorescence (PXRF) spectrometry was used to test the efficiency of four machine learning techniques (random forest; linear and nonlinear support vector machine; classification and regression tree) for distinguishing three land use types in India based upon scans of mineral surface (0–20 cm) soil. Results showed similar performance among the four tested algorithms, with classification accuracy of a randomly selected validation set ranging from 83% to 91%. The classification and regression tree was favored based upon simple "IF AND THEN" rules which make classification of the data simple. In sum, PXRF data was shown highly effective at differentiating land use types in India. Future work should focus on a larger number of land use classification types and possible combination of PXRF data with complimentary proximal sensing datasets (e.g., visible near infrared spectroscopy). Highlights • Elemental data from PXRF can be used for land classification. • Classification accuracy was between 83% and 91%. • Classification and regression tree analysis gave optimal land classification. [ABSTRACT FROM AUTHOR]