Comparing Spectral Soil Inference Systems and Mid‐Infrared Spectroscopic Predictions of Soil Moisture Retention

Abstract

Mid‐infrared spectroscopy has been proposed as a cheap yet accurate alternative to a number of laboratory methods for measuring soil properties. While accurate predictions of a number of basic soil constituents have been reported, properties associated with soil structure have received far less attention. In this study, we looked at the efficacy of mid‐infrared reflectance spectroscopy in predicting moisture retention and whether better predictions can be achieved using pedotransfer functions using spectroscopic predictions of basic soil constituents as inputs. Three methods were used to predict volumetric moisture retention: (i) mid‐infrared (MIR) spectra and partial least squares regression, (ii) a neural network pedotransfer function (PTF) using laboratory particle‐size distribution and bulk density data, and (iii) a pedotransfer function with MIR‐predicted particle‐size distribution and bulk density as inputs. We used Lin's concordance correlation coefficient as a goodness‐of‐fit measure. Predictions of volumetric moisture retention on intact structured soils were generally poor, particularly at the wet end. Improved predictions were observed at dry‐end matric potentials, where moisture retention was more correlated with particle‐size distribution than soil structure. The neural network PTF was found to have better goodness of fit for all matric potentials; however, predictions at larger matric potentials (wet end) were still poor due to poor bulk density predictions. In light of this work, we propose that while MIR spectroscopy may be a valuable predictor of fundamental soil constituents such as particle‐size fractions and organic C, predictions of soil properties dependant on soil structure, such as volumetric moisture retention, may prove difficult. Mid‐infrared spectroscopy in combination with PTFs should provide improvements to moisture retention predictions through improved representation of the influential processes, namely soil structure and adsorptive forces.

Soil Chemical Insights Provided through Vibrational Spectroscopy

“…Recent work shows that we can potentially calibrate for other parameters in soil such as metals, carbonates (inorganic C), enzymes, potential nitrification, and pH (Du et al, 2013b;McCarty et al, 2002;Mimmo et al, 2002;Reeves et al, 2001;Siebielec et al, 2004). Because DRIFTS contains information related to organic and inorganic components that relate to texture and particle size distribution, calibrations can also be developed for soil physical attributes including moisture retention, bulk density, or hydraulic conductivity (Janik et al, 2007;Minasny et al, 2008;Tranter et al, 2008). For example, PLS regression coefficients show that absorbance at the 3630 cm −1 clay band, and at the 1640 cm −1 amide I or phenyl C-C band are important for calibrations for soil water retention .…”

Diffuse Reflectance Spectroscopy for High-Resolution Soil Sensing

“…Brown et al, 2006), and regression rules (e.g. Minasny and McBratney, 2008); however, the outcome often only has limited advantages over PLSR. Viscarra Rossel (2007) combined PLSR with bootstrap aggregation (bagging-PLSR) to improve the robustness of the PLSR models and produce predictions with a measure of their uncertainty.…”

Application of FTIR Spectroscopy to Agricultural Soils Analysis

“…However, the study of Bertrand et al (2002) clearly shows that KBr dilution of the samples or high CaCO 3 concentration deteriorate the results. Although this has not been specifically verified, the very poor results obtained by Tranter et al (2008) on Dataset #1 were most probably due to the fact that the spectra were recorded from intact soil cores rather than from grounded samples as in all the other studies. Regarding moisture of air-dried and oven-dried soil, the few results available are very good, unless the soil has a high CaCO 3 concentration.…”

“…A summary of the main results obtained with mid-IR DRIFT spectroscopy is presented in Table 2. With the exception of Dataset #1 in Tranter et al (2008), mid-IR spectroscopy was capable of estimating to some extent water retention at -10kPa and -1500kPa, with R 2 values ranging from 0.64 to 0.81 and from 0.66 to 0.89, respectively. However, the study of Bertrand et al (2002) clearly shows that KBr dilution of the samples or high CaCO 3 concentration deteriorate the results.…”

“…An important feature of this approach is that the determination errors are propagated through the various estimation stages so that a confidence interval can be associated with each prediction (Figure 9). Tranter et al (2008) showed how such an approach improved considerably the estimation of soil water retention, which was very poorly estimated directly from the mid-IR spectra (See Table 2). Minasny&McBratney (2008) also reported more parsimonious models and greater accuracy when using an inference system then when using straightforward PLS regression on the mid-IR spectra.…”