JEQ Journal of Natural Resources and Life Sciences Education
HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS
QUICK SEARCH:   [advanced]


     

Published online 5 July 2005
Published in J Environ Qual 34:1422-1434 (2005)
DOI: 10.2134/jeq2004.0353
© 2005 American Society of Agronomy, Crop Science Society of America, and Soil Science Society of America
677 S. Segoe Rd., Madison, WI 53711 USA
This Article
Right arrow Abstract Freely available
Right arrow Figures Only
Right arrow Full Text (PDF) Free
Right arrow Alert me when this article is cited
Right arrow Alert me if a correction is posted
Services
Right arrow Similar articles in this journal
Right arrow Similar articles in ISI Web of Science
Right arrow Similar articles in PubMed
Right arrow Alert me to new issues of the journal
Right arrow Download to citation manager
Citing Articles
Right arrow Citing Articles via ISI Web of Science (3)
Right arrow Citing Articles via Google Scholar
Google Scholar
Right arrow Articles by Cohen, M. J.
Right arrow Articles by DeBusk, W. F.
Right arrow Search for Related Content
PubMed
Right arrow PubMed Citation
Right arrow Articles by Cohen, M. J.
Right arrow Articles by DeBusk, W. F.
Agricola
Right arrow Articles by Cohen, M. J.
Right arrow Articles by DeBusk, W. F.
Related Collections
Right arrow Wetland Soils
Right arrow Wetlands and Aquatic Processes
Right arrow Soil Analysis
Right arrow Soil Methods/Instrumentation
Right arrow Soil Biology

TECHNICAL REPORTS

Wetlands and Aquatic Processes

Visible-Near Infrared Reflectance Spectroscopy for Rapid, Nondestructive Assessment of Wetland Soil Quality

Matthew J. Cohen*, Joseph P. Prenger and William F. DeBusk

Wetland Biogeochemistry Laboratory, Department of Soil and Water Science, University of Florida, P.O. Box 110510, Gainesville, FL 32611-0510

* Corresponding author (mjc{at}ufl.edu)

Received for publication September 15, 2004.

   ABSTRACT
TOP
 ABSTRACT
INTRODUCTION
 MATERIALS AND METHODS
 RESULTS
 DISCUSSION
 REFERENCES
 
Recent evidence supports using visible-near infrared reflectance spectroscopy (VNIRS) for sensing soil quality; advantages include low-cost, nondestructive, rapid analysis that retains high analytical accuracy for numerous soil performance measures. Research has primarily targeted agricultural applications (precision agriculture, performance diagnostics), but implications for assessing ecological systems are equally significant. Our objective was to extend chemometrics for sensing soil quality to wetlands. Hydric soils posed two challenges. First, wetland soils exhibit a wider range of organic matter concentrations, particularly in riparian areas where levels range from <1% in sedimentation zones to >90% in backwater floodplains; this may mute spectral responses from other soil fractions. Second, spectral inference of cation concentrations in terrestrial soils is for oxidized species; under reducing conditions in wetlands, oxidation state variability is observed, which strongly affects chroma. Riparian soils (n = 273) from western Florida exhibiting substantial target parameter variability were compiled. After minimal pre-processing, soils were scanned under artificial illumination using a laboratory spectrometer. A multivariate data mining technique (regression trees) was used to relate post-processed reflectance spectra to laboratory observations (pH, organic content, cation concentrations, total N, C, and P, extracellular enzyme activity). High validation accuracy was generally observed (r2validation > 0.8, RPD > 2.0, where RPD is the ratio of the standard deviation of an attribute to the observed standard error of validation); where accuracy was lower, categorical models (classification trees) successfully screened samples based on diagnostic functional thresholds (validation odds ratio > 10). Graphical models verified significant association between predictions and observations for all parameters, conditioning on biogeochemical covariates. Visible-near infrared reflectance spectroscopy offers both cost and statistical power advantages; hydric conditions do not appear to constrain application.

Abbreviations: VNIRS, visible-near infrared reflectance spectroscopy


   INTRODUCTION
TOP
 ABSTRACT
 INTRODUCTION
MATERIALS AND METHODS
 RESULTS
 DISCUSSION
 REFERENCES
 
THERE IS INCREASING need for analytical tools that aid in quantifying and predicting ecosystem condition over space and time. Some attributes (e.g., vegetation cover) may be monitored remotely, but for mapping soil functional attributes, ground sampling remains the preferred method. Significant costs ensue where spatial variability confounds generalization of sparse point observations of critical indicators (Dent and Young, 1981; Burrough and Frank, 1996). Specific examples of management decision support and risk-based methods that are increasingly spatially and temporally data intensive include precision agriculture (Van Alphen and Stoorvogel, 2000), soil condition mapping (Bruland and Richardson, 2004; McBratney et al., 2003), pedogenetic modeling (Park and Vlek, 2002; Hoosbeek and Bryant, 1992), and large area monitoring (Schroder et al., 2003). The sampling intensity required for effective spatial and temporal characterization and observational control of confounding factors results in significant analytical costs, and frequently less than desirable sample density and analyte diversity.

Visible-near infrared reflectance spectroscopy (VNIRS) offers significant potential for improving this situation; efforts to relate reflectance spectra to a wide array of soil properties have been repeatedly successful (Ben-Dor and Banin, 1995; Kooistra et al., 1997; Foley et al., 1998; Gillon et al., 1999; Reeves et al., 1999; Chang et al., 2001; Shepherd and Walsh, 2002). Visible-near infrared reflectance spectroscopy is well established for qualitative evaluation of manufactured materials, geologic samples, and agricultural products (Ghosh, 1978; Hunt, 1982; Osborne, 1983; Clark, 1999). Only recently, however, has the technique been matched with improved instrumentation and analytical methods capable of providing quantitative inference from complex spectral signatures of heterogeneous media. Performance measures for which inferential models have been developed include soil physical and hydraulic attributes [texture, infiltration capacity (Shepherd and Walsh, 2002; Cohen, 2003)], chemical attributes [cation concentrations, clay activity, phosphorus, nitrogen, organic and total carbon, pH (Ben Dor and Banin, 1995; Dunn et al., 2002; Shepherd and Walsh, 2002)], proximate estimates of biological activity [enzymes (Reeves et al., 2000), C and N mineralization rates (Bouchard et al., 2002; Fystro, 2002)], toxicity and/or contamination (Kooistra et al., 1997), and indicators of ecosystem hydrologic stress (Bacchus et al., 2003). Janik et al. (1998) discuss the potential to supplant standard laboratory methods with VNIRS for all analyses except chemometric development because of high precision observed between replicates (typically <2% RMSE), and low-cost and expedient analysis.

The premise of VNIRS [following from Beer's Law (Johnston and Aochi, 1996)] is that high-resolution diffuse reflectance spectra (1-nm bands) in the visible (0.35–0.75 µm) and near infrared (0.75–2.5 µm) regions of the electromagnetic spectrum contain interpretable information about the composition, structure, and concentration of various sample attributes. While fundamental harmonic oscillations of organic bonds (e.g., C–H, N–H, and O–H) and mineral functional groups (metal oxides, minerals, ions) are primarily found in the mid-infrared (MIR) (Johnston and Aochi, 1996), the VNIR range contains multiple overtones that have proven as useful for chemometric development. While MIR sensors are available, and effective chemometrics have been derived from MIR spectra (Reeves et al., 2001), sample pre-processing requirements, thermal noise, and instrument cost limit widespread application. The size, portability, low cost, high precision, and increasing availability of VNIRS sensors, in addition to their concurrence with multi- and hyper-spectral satellite sensor spectra, make them ideal for routine spectral assessment. These advantages, plus simple sample preparation and limited use of chemical reagents (Cozzolino and Moron, 2003), make VNIRS desirable for ecosystem studies where significant spatial and temporal variability make large sample throughput informative.

Despite the array of soil, plant, and animal tissue attributes for which chemometrics have been developed, and the diversity of regions or habitats for which spectral training libraries have been compiled (Cozzolino and Moron, 2003; Shepherd and Walsh, 2002; Dunn et al., 2002), there are few studies that demonstrate general method efficacy in wetland systems [though see Bouchard et al. (2002) for specialized application in salt marshes, and Bacchus et al. (2003) for hydrologic stress inference at the ecosystem scale]. The soils that form under hydric conditions [i.e., annually inundated and/or saturated soils (Mitsch and Gosselink, 2000)] range widely in physical, chemical, and biological attributes in response to sediment inputs, organic matter mineralization rates, nutrient loads, and geologic substrate. Riparian wetland systems were chosen as the target habitat because the soils that form along rivers and streams are among the most spatially and temporally diverse in many landscapes (Lyons et al., 1998; Johnston et al., 2001; Ettema and Wardle, 2002).

Two factors may limit VNIRS efficacy in wetland soils: (i) wetland soils have a greater range of organic matter content than terrestrial soils (e.g., Histosols may have organic matter contents approaching 100% while mineral wetland soils [e.g., Fluvaquents] often contain less than 10% organic matter in surface horizons); and (ii) the mosaic of oxidized and reduced conditions that prevail in wetland ecosystems controls the valence state of iron and manganese, which has well-known effects on soil chroma (Faulkner and Patrick, 1992; Mitsch and Gosselink, 2000). The challenge arises because extraction methods typically convert cations to one form, but spectral scans are collected on bulk soils. It should be noted that total phosphorus exhibits the same condition; that is, P is a component of numerous organic and inorganic compounds in soils, which could confound spectral inference of total concentrations. The numerous instances of successful chemometric development provide evidence that multivariate calibration methods used with spectral data may be insensitive to such speciation.

Of additional interest are various indicators of microbial community condition that have been shown effective for diagnosing environmental stressors. Soil microbes allocate energy to production of nutrient acquisition enzymes based on presence or absence of readily available nutrients (Sinsabaugh and Moorhead, 1994). For example, ß-glucosidase activity (part of the cellulolytic pathway) is an indication of soil carbon dynamics. Similarly, reduced relative activity of peptidase and acid or alkaline phosphatase may indicate nutrient enrichment. Dehydrogenase, a measure of overall microbial activity, may be a general eutrophication indicator, and ratios of various soil enzyme activities can provide insight into nutrient limitation (Prenger and Reddy, 2004).

The advantage of microbial enzymatic indicators is that they integrate the soil environment; this is valuable because of the difficulty in defining strictly physical–chemical thresholds for system impairment due to confounding interactions between and natural variability within the suite of available biogeochemical indicators (e.g., Al, P, Ca, N, organic matter). While the microbial community responds strongly to integrated nutrient, cation, and organic matter quality gradients (Chrost, 1991; White and Reddy, 2000; Wright and Reddy, 2001a, 2001b; Prenger and Reddy, 2004), their measurement is costly and time consuming. Limited research has addressed the feasibility of predicting microbial enzymatic indicators and nitrification potential from VNIR spectra [in agricultural soils (Reeves et al., 2000)], with promising results for moderate accuracy applications (e.g., indicator mapping). While diffuse reflectance spectra from heterogeneous media are likely not sensitive enough to directly observe microbial enzymes, spectra can provide an integrated simultaneous view of soil functionality that may approximate the various gradients to which the microbial community responds. As a result, pedotransfer functions for complex soil performance measures may be possible.

Our objective was to develop chemometrics for rapidly sensing soil quality in riparian wetland systems in support of research on linkages between agricultural–urban land uses and riparian soil quality. The attributes include those widely used for functional assessment (e.g., total C, N, and P, organic C, pH, various cation and P extractions, extracellular enzyme activity) (Faulkner and Patrick, 1992). We focused first on continuous prediction; where insufficient prediction accuracy was observed for a given attribute, we explored categorical discriminant models. An additional objective was to enumerate the specificity of spectral inference by examining partial correlations between predicted and observed soil characteristics conditioned on values of those attributes that may act as covariates.


   MATERIALS AND METHODS
TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
RESULTS
 DISCUSSION
 REFERENCES
 
Study Region and Site Selection
This work is part of a study examining land use effects on riparian soil quality. River systems in western Florida comprise the study region (Fig. 1) , chosen because of the variety of agricultural, urban, and managed forest land use activities and an increasing regulatory focus on water quality and freshwater quantity in Pensacola Bay. To characterize impact intensity and modality (urban or agricultural effects vs. reference conditions), we sampled along riparian corridors in 21 subbasins corresponding with an existing database of stream habitat quality [BioRecon/Stream Condition Index (Barbour et al., 1996)]. Only cursory further discussion of land use effects is included; however, sampling the gradient of anthropogenic influence was expected to allow development of a regionally comprehensive spectral reflectance library.



View larger version (67K):
[in this window]
[in a new window]
  		Fig. 1. Study area in western Florida including Pensacola Bay, Santa Rosa County and Escambia County. Shown are Florida county boundaries with (A) regional hydrography and transportation networks and (B) topography. Both show soil sample cluster locations.

 
Soil Sampling
At each sampling location, two 20-m transects were established parallel to stream flow, one near the stream channel (<10 m from the bank) and the other 5 m from the wetland–upland boundary. Every 5 m along each transect, three 5-cm soil cores were collected using a 6.5-cm-diameter corer. Litter and roots were removed and a single composite sample (approximately 50 g) of the three cores was sealed in a watertight plastic bag and placed immediately on ice; samples were stored at 4°C until analyses were performed. Overall 273 soils were collected, analyzed, and scanned.

Chemical Analyses
Total C and N was determined on dried, ground soil samples by dry combustion (Nelson and Sommers, 1996) using a Carlo-Erba NA-1500 CNS Analyzer (Haak-Buchler Instruments, Saddlebrook, NJ). Organic matter content (%) was determined from residue after ashing at 550°C (Anderson, 1976). Total P was determined by combusting approximately 0.2 to 0.5 g oven-dried, finely ground soil at 550°C for 4 h, digesting the ash with 6 M HCl and continuous heating on a hot plate, and filtering through a no. 41 Whatman (Maidstone, UK) filter (Anderson, 1976), followed by analysis of P by automated ascorbic acid method (Method 365.1; USEPA, 1993). Soil pH was determined by pH meter on 20 g wet soil after equilibrating with 10 mL of deionized water. The HCl-extractable cations were determined by extraction of 0.5 g dry soil in 25 mL 1 M HCl with shaking for 3 h, filtration through a 0.45-µm membrane filter, and analysis by inductively coupled plasma (ICP) (USEPA Method 200.7). Water-extractable carbon was determined by extraction of the wet soil equivalent of 2.5 g soil dry weight in 25 mL of distilled deionized water with shaking for 1 h, followed by filtration through a 0.45-µm membrane filter (Kuo, 1996) and analysis on a Dohrmann (Mason, OH) DC-190 TOC Analyzer. Mehlich-I extractable ions were determined by extraction of 5 g soil in 20 mL 0.0125 M H2SO4, 0.05 M HCl with 5 min shaking, followed by filtration through no. 42 Whatman filter paper (Amacher, 1996) and analysis by ICP (Method 200.7; USEPA, 1991).

Enzyme Analyses
Enzyme activities were assayed using the fluorescent model substrate 4-methylumbelliferone (MUF) (Chrost and Krambeck, 1986; Hoppe, 1993; Sinsabaugh et al., 1997) at approximately ambient pH (6.0). All soil enzyme analyses were performed on well-mixed fresh material (within 2 weeks of sampling) from which all visible roots and living plant material had been removed. Soil samples (approximately 1 g) were placed in approximately 9 mL distilled water, and clumps were broken up by brief agitation with a Tissue Tearor Model 398 (Biospec Products, Bartlesville, OK). Immediately before enzyme assays, a 1/100 or 1/200 dilution of soil or detritus was prepared in water by serial dilution. Two hundred microliters of well-suspended soil slurry was transferred by pipette into 8 wells of a 96-well microtiter plate, and 50 µL of substrate solution added to 4 wells (with 4 blanks). Samples were incubated (2 h for phosphatase, 24 h for all others) in the dark at room temperature except for dehydrogenase, which was incubated at 30°C. Phosphatase and ß-glucosidase assays were stopped by addition of 10 µL 0.1 M NaOH. Substrate was added to blanks and immediately read on a Model FL600 fluorometric plate reader (Bio-Tek Instruments, Winooski, VT). Dehydrogenase assays were stopped with 50 µL acetone, incubated an additional 2 h, and read. Substrate solutions were as follows: for acid phosphatase, 500 µM methyl-umbelliferyl (MUF)-phosphate in 5 mM MES pH 6.0; for ß-glucosidase, 500 µM MUF-glucoside in 5 mM MES pH 6.0; for dehydrogenase, 500 µM 5-cyano-2,3-ditolyl tetrazolium chloride (CTC) in 100 mM Tris pH 7.8. Concentrations were calculated from a standard curve of MUF or CTC-Formazan. Excitation (ext) and emission (em) spectra for the two fluoro-chromes were: ext 360 ± 40, em 460 ± 40 (MUF-P, MUF-G); and ext 530 ± 25, em 645 ± 40 (CTC). Enzyme activities were calculated as µg product per g dry soil per h.

Spectral Scanning
Before scanning, soils were air-dried, ground to remove any macro-structure, and passed through a 2-mm sieve. A portion of the soils was hand-ground (n = 126); the remainder was machine-ground (n = 147). Screening models (classification tree models) applied to grinding method as the target variable exhibited minimal discriminatory power (95% confidence interval of the odds ratio = [1.01–1.08]) indicating that the signal is effectively independent of preparatory approach.

Scanning was performed with a FieldSpec FR Pro diffuse reflectance spectroradiometer (Analytical Spectral Devices, Boulder, CO) using Spectralon (Labsphere, Hutton, NH) as a white reference. This instrument uses a high temperature (3000 K) tungsten filament bulb for sample illumination, and collects reflected light in 1-nm bandwidths between 350 and 2500 nm using three internal spectrometers (350–1000, 1000–1800, and 1800–2500 nm). The optical setup consisted of a portable contact probe with a 2-cm field-of-view placed in direct contact with the soil surface.

Four replicate scans of each soil sample (each integrating 10 observations) were taken after adjusting the location and rotation of the contact probe. The means across sample replicates for each waveband and variance (relative error) between replicates were computed; replicate spectra exhibiting relative error > 2% from the sample mean were discarded (n = 13) and the sample mean recomputed.

Data Processing
Raw reflectance spectra were further processed to expedite statistical analysis and control for power source and ambient light variability. First, spectral resampling reduced spectral resolution to 10 nm; previous work (Shepherd and Walsh, 2002) showed no significant loss of information due to resampling. Eleven resampled spectra were selected to reflect the range of albedo in the sample population (Fig. 2) , with data reported in relative reflectance units. Second, first-derivative transformation over a 20-nm gap with second order Savitzky–Golay smoothing (Fearn, 2000) was performed to minimize variance between samples caused by grinding and optical setup. Finally, regions of low signal to noise, due to both splicing between internal spectrometers and inherent power limitations, were omitted (0.35–0.40, 0.97–1.01, 1.78–1.92, and 2.46–2.50 µm); the result is 193 waveband predictors between 400 and 2460 nm. While additional data processing [log reciprocal reflectance transformation, scatter correction, detrending, standard variate normalization (Barnes et al., 1989)] is frequently used to improve model effectiveness, no effect was observed on classification or regression tree models, likely due to the nonparametric, threshold-based algorithms used in this modeling technique (Breiman et al., 1984).



View larger version (29K):
[in this window]
[in a new window]
  		Fig. 2. Eleven raw relative reflectance spectra, sampled as mean albedo end-members from the existing spectral reflectance library (n = 321).

 
Replicate consistency was evaluated using standard deviation, both in raw reflectance units and scaled to mean reflectance in each bandwidth. Results indicate significant noise in the blue region of the spectrum, but SD less than 2% of mean reflectance throughout most of the spectral range (Fig. 3) .



View larger version (20K):
[in this window]
[in a new window]
  		Fig. 3. Mean replicate (n = 4 per soil sample) variance across the spectrum for the entire data set reported as % of raw reflectance and also scaled to the mean. This indicates moderate spectral response stability with the optical setup described; the blue region of the spectrum (350–400 nm) exhibited low signal-to-noise and was excluded from chemometric development.

 
Outlier Detection
Principal components analysis (PCA) was applied to the entire soil archive to ascertain that all samples were from the same population and to visualize spectral variance. Three principal components axes were selected that represent 78% of the spectral variability. Examination of the PC score correlations with soil attribute data indicates that Axis 1 is loaded by total phosphorus (correlation = –0.53), pH (–0.36), iron (–0.5), and organic matter (–0.44), while Axis 2 is loaded by organic matter (–0.54), aluminum (–0.44 for Mehlich extraction), and total nitrogen (–0.35). Axis 1 also exhibits moderate associations with all four extracellular enzymes.

The resulting biplot (Fig. 4 , showing PC Axes 1 and 2) illustrates the utility of data visualization techniques to determine if new samples fall within the spectral bounds of the current archive. Two in-stream sediment samples were included in the original reference library (otherwise all riparian soils); these two outliers (Fig. 4) appear to be sufficiently spectrally distinct to warrant exclusion from model development and were subsequently removed (chemometric development would require additional stream sediment samples for the reflectance library). Further, spectral variance captured by Axis 2 (Fig. 4) indicates that approximately 10 samples are distant from the data cloud, though not statistically outliers. These samples, collected from an area with significant sediment deposition, were retained in further analyses. In practice, data visualization allows new soils that exceed the calibration spectral bounds to be identified and subsequently included in the calibration model after the complement of laboratory tests has been performed [see Shepherd and Walsh (2002) for a thorough treatment of this adaptive spectral library approach].



View larger version (20K):
[in this window]
[in a new window]
  		Fig. 4. Biplot of Axes 1 and 2 from principal components analysis (PCA) of the spectral reflectance library (SRL). Also given are axis correlations with specific soil properties. Two spectral outliers were identified using PCA; these two samples (in-stream sediment samples) were omitted from further analyses.

 
Visible-Near Infrared Reflectance Spectroscopy Calibrations: Continuous and Categorical
Strong spectral autocorrelation necessitates methods that can efficiently deal with highly colinear data sets. There is substantial evidence that pairwise correlations between observed soil attribute levels and individual wavelength responses are strong (Fig. 5) . However, developing models that are not overfit and can be calibrated with relatively small sample sizes (nsamples ~ npredictors) has necessitated use of advanced statistical learning tools.



View larger version (28K):
[in this window]
[in a new window]
  		Fig. 5. Bivariate Pearson correlation coefficients between selected soil attributes and spectral reflectance across the entire visible-near infrared reflectance spectroscopy (VNIRS) spectral range.

 
Typically, inference from spectra is done using partial least squares (PLS) or principal components (PC) regression. Despite the widespread success of the PLS (Dunn et al., 2002; Chang et al., 2001), other methods (e.g., regression trees, multivariate regression splines) have been shown (Shepherd and Walsh, 2002) to provide improved model stability, increasing prediction accuracy. For this work, spectral calibration models were developed using classification and regression trees (Breiman et al., 1984) in the CART software environment (Steinberg and Colla, 1997). Committees of exploratory (i.e., unpruned) regression trees (n = 200) were grown on adaptively resampled (adaptive resampling and combination [ARCing] parameter = 4.0) data sets using a portion (67%) of the data; validation was performed using the remaining data (33%), applying the topology of each committee tree to sample spectra. The committee approach, wherein the predicted value is the mean (or mode for categorical data) prediction across all trees in a committee, addresses implicit limitations in the recursive partitioning algorithm that render individual trees sensitive to nonglobally optimal model topology (Breiman et al., 1984). Committee mean predictions were compared with observed values for both calibration and validation for each soil parameter.

To evaluate model fit, several indicators were employed. First, the coefficient of determination between predicted and observed values provides a common measure of variance reduction. The slope of the fitted line (specifically, deviation from unit slope) provides evidence of systematic bias in the calibrations (Fystro, 2002). The standard error of validation (SEV; computed from validation fit residuals) provides further indication of model effectiveness that can screen models based on required application accuracy. Finally, the RPD statistic, defined as the ratio of the standard deviation of an attribute to the observed standard error of validation, provided a metric of model fit that can be compared across soil attributes. Dunn et al. (2002) and Chang et al. (2001) offer qualitative interpretation of this metric: RPD values less than 1.5 were considered insufficient for most applications, while models with values greater than 2.0 were considered excellent. Models with values of >1.5 but <2.0 were judged useful contingent on accuracy requirements of the desired application.

Where RPD values were less than 2, we explored the potential of categorical models to diagnose soils above and below a functional threshold. Thresholds were defined based on soil property patterns in this data set, corroborated with literature estimates where possible. Specifically, samples in this study were collected from subbasins that were classified along a quantitative gradient of anthropogenic impairment (Brown and Vivas, 2005). Pairwise contrasts between riparian soils from minimally impacted basins and soils from impacted basins were used to identify levels of soil attributes characteristic of each condition. Thresholds were defined as two standard errors from the computed mean across minimally impacted sites in the direction typical of disturbance; cases (1) refer to soil samples that exceed the threshold and appear to be from impacted subbasins, while controls (0) identify soils within the bounds of minimally impacted sites.

Categorical models were developed using classification trees. Trees were overfit and then pruned until a cost:complexity measure that integrates goals of classification accuracy and model parsimony was minimized. Rather than partitioning the data into separate calibration and validation data sets because of the small sample sizes, a V-fold cross-validation procedure was used to estimate the error rate expected for model prediction (cross validation efficiency reports were not available for regression trees used for continuous prediction so separate calibration and validation data sets were used). For each model, we report overall accuracy (% correct), specificity (% cases correctly predicted), sensitivity (% controls correctly predicted), and predictive odds ratio for both calibration and cross-validation. The odds ratio gives an assessment of model fit that incorporates information about case prevalence and discrimination accuracy; low prevalence of either positives or negatives can cause other measures of accuracy to be misleading. Typically, an odds ratio of >10 is considered to provide useful predictive power for diagnostic models in the medical literature (Fischer et al., 2003).

Partial Correlation Analysis
Specific physical–chemical mechanisms by which complex soil attributes are spectrally inferred are absent from most VNIRS studies, including this work; statistical "black-boxes" effectively link spectra and soil properties. This reliance on statistical inference in the absence of direct mechanistic interpretation generates operational uncertainty when chemometrics are applied to new soils. High correlations between certain soil properties make it reasonable to contend that spectral inference of more complex soil attributes (e.g., enzyme activity) that are known to respond to basic soil properties (e.g., organic matter, total phosphorus) is spurious. That is, observed chemometric efficiency for those complex parameters may be due primarily to correlation with other parameters easily inferred spectrally rather than specific spectral response characteristics. It is important to verify that observed chemometric accuracy levels are not spurious.

To explore correlations between predicted and observed values conditioned on the levels of all other variables, we used graphical models—an extension of hierarchical log-linear models (Agresti, 1990) that allows depiction of variable associations using graphs of arcs connecting nodes that are significantly conditionally associated (Edwards, 1995a)—implemented in MIM 3.2 (Edwards, 1995b). The resulting graphical depictions of model statistical form, where arcs connect conditionally associated nodes, allow direct interpretation of partial correlation significance. Starting with saturated models, wherein arcs connect all nodes with each other, various stepwise deletion algorithms allow iterative removal of those arcs for which conditional association is nonsignificant. Typical deletion criteria include the Akaike information criterion (AIC), residual deviance (statistical significance assessed using the likelihood ratio test), and standard F tests. We selected residual deviance as our deletion criteria because it allows the most control over the threshold significance that delimits conditional association; we used a significance threshold ({alpha}) of 0.001.

This method offers important insight about associations that cannot be provided by standard correlations. Pairwise correlation is frequently a spurious indicator of influence within multivariate data because of between-factor correlations; partial correlations provided by this graphical modeling method provide better insight into conditional association patterns. Where partial correlation is nonsignificantly different from zero we infer that the observed predictive power (measured as pairwise correlation) is strictly a function of the associations with other soil parameters. Where conditional association between predicted and observed is significant, spectra provide information about soil attributes that is conditionally independent of what could be inferred by correlations with the suite of soil properties examined in this study. While we cannot infer that the spectrometer acquires reflectance information specific to, for example, enzymatic activity, we can infer that spectra integrate the physical and chemical environments that control that attribute. Other measurable soil properties (e.g., soil C quality, nutrient mobility) might be equally effective predictors for pedotransfer functions, but spectra reduce sample characterization requirements and expedite pedotransfer inference.


   RESULTS
TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS
DISCUSSION
 REFERENCES
 
Summary statistics for the soil attribute data (Table 1) indicate soils with a wide range of chemical and biological characteristics; these standard deviation observations are used in Table 2 to compute the RPD statistic. The biogeochemical bounds of the spectral reflectance library (SRL) implied by these data were considered sufficient to proceed with spectral analysis.


View this table:
[in this window]
[in a new window]
  		Table 1. Population statistics for the spectral reflectance library (SRL).

 

View this table:
[in this window]
[in a new window]
  		Table 2. Calibration and validation efficiency for chemometrics between spectra and soil chemical and biological attributes. Also given are standard errors observed for holdout validation data and the RPD (i.e., the ratio of the standard deviation of an attribute to the observed standard error of validation) metric of model efficiency.

 
Spectral correlations with selected observed soil attributes (Fig. 5; organic matter, pH, total P, cation concentrations, and enzyme activity) are plotted to illustrate spectral sensitivity for model development; soil attributes and spectral response are highly correlated (|r| > 0.6) in numerous spectral regions, particularly regions of the near infrared (750–1100, 1800–2000 nm). Strong spectral autocorrelation typical of high-resolution NIR spectra is also evident from these graphs (e.g., 500–1000 nm).

Chemometric effectiveness is summarized in Table 2. The coefficient of determination (r2) in calibration and validation is given, along with several additional statistical parameters that summarize model fit. The standard error of validation is a useful indicator of value for specific applications, but offers limited insight as a comparative metric of model fit across soil attributes. The RPD statistic offers a clearer indication of soil parameters for which effective models were obtained. Dunn et al. (2002) suggest that RPD values of <1.6 indicate poor models; values of >2 are considered to indicate highly efficient models while the adequacy of values between should be determined by end use.

In several cases, resulting models were poor (H2O-extractable P, HCl-extractable calcium); for others, including several cation extractions (Mehlich Ca and P, HCl-extractable Mg, H2O-extractable total organic C), RPD values suggest that spectral inference may be insufficient for applications requiring high accuracy. However, chemometrics were effective for all other soil attributes. Models for total carbon and organic matter (RPD 5.74 and 5.89, respectively) were particularly effective. Models for inferring microbial enzyme activity were effective, with RPD values exceeding 2 for all except dehydrogenase.

Threshold values for binary case-reference definitions are given in Table 3. Categorical screening models to discriminate case and reference soils for all attributes are summarized in Table 4; we developed screening models for all soil properties, not just those for which the RPD suggested a need. The results demonstrate high degrees of model screening accuracy, with only four parameters (H2O-extractable P, HCl-extractable Fe, HCl-extractable Al, and acid phosphatase activity) providing validation odds-ratios less than 10.


View this table:
[in this window]
[in a new window]
  		Table 3. Development of functional thresholds based on study area reference and impacted site soil attributes. Categorical screening models are developed to discriminate soils above and below thresholds using spectral reflectance. Cases are defined to correspond with impacted site values; controls correspond with reference sites.

 

View this table:
[in this window]
[in a new window]
  		Table 4. Summary of binary screening models for soil parameters given thresholds in Table 3. Given are validation and calibration model summaries.

 
A correlation matrix containing pairwise associations for a selection of measured soil attributes is given (lower half of Table 5). Strong pairwise correlations between parameters (e.g., organic matter and total C, total C and total N, enzyme activity, total P and Fe/Al) may suggest spurious models; that is, apparent spectral responsiveness to a specific soil attribute may only be a function of reflectance due to other characteristics. Partial correlations (upper half of Table 5) illustrate the extent to which pairwise correlations between parameters may change dramatically when they are conditioned on levels of confounders. In many cases, strong pairwise associations are nonsignificant when controlling for levels of other parameters (e.g., dehydrogenase and total C, peptidase and dehydrogenase, dehydrogenase and ß-glucosidase); in some cases (e.g., total N and Fe, total P and total C, total P and acid phosphatase) the direction of association reverses. Partial correlation information is summarized graphically in Fig. 6 . Nodes (soil attributes) connected by arcs are significantly conditionally associated (p < 0.001). Association strength and direction can be obtained from Table 5. Notably, soil organic content is conditionally associated with all other parameters, substantiating its central role on soil processes.


View this table:
[in this window]
[in a new window]
  		Table 5. Pairwise (lower) and partial (upper) correlation matrix for selected soil attribute observations. Partial correlations different from 0 are significant at p < 0.001; pairwise correlation significance is not assessed.

 


View larger version (47K):
[in this window]
[in a new window]
  		Fig. 6. Graphical depiction of conditional association pattern for observed soil parameters. Nodes represent selected properties; arcs between nodes are significant at p < 0.001. Partial correlation coefficients are given in Table 5.

 
Also shown (Fig. 6) is an arc (rpartial = 0.75) between predicted and observed pH values. Similar graphical models were developed for each parameter. Partial correlations between predicted and observed values conditioned on levels of other measured parameters (Table 6) enumerate the effect of covariates on the efficacy of chemometric models. In each case, partial correlations between observed and spectrally predicted values are positive and strongly significant indicating that spectral signatures provide information about each soil attribute that cannot be obtained using correlations with the laboratory observations in this study. In each case, the partial correlation coefficient is smaller than the pairwise correlation; the difference indicates the degree to which spectral response due to correlated parameters drives the observed chemometric efficiency.


View this table:
[in this window]
[in a new window]
  		Table 6. Summary of partial correlation analyses. The observed partial correlations between predicted and observed soil attribute levels conditioned on observed values of other soil attributes are shown, along with the p value (likelihood ratio test) assessing arc significance between predicted and observed nodes.

 

   DISCUSSION
TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS
 DISCUSSION
REFERENCES
 
Numerous aspects of ecosystem evaluation can benefit from calibrations developed in this study. Specifically, low costs of sample evaluation allow high spatial and temporal resolution for routine monitoring across large areas, which may greatly reduce management uncertainty and allow increases in spatial sampling densities for mapping applications. Where confounding factors and spatial heterogeneity make generalization from point observations problematic, this technique can increase statistical power to control for confounders and assess spatial–temporal generality. This is particularly true in riparian wetland regions that are highly heterogeneous in space and time due to differential hydrologic regimes, differential vegetative community structure, and anthropogenic impacts. Effective characterization of riparian soil quality for developing linkages with land use decision making will ultimately require tools for dealing with the effects of variability.

Spectra offer a unique integrative view of soil quality. The ability to obtain simultaneous information about mineralogy, soil chemistry, soil organic matter, and soil physical attributes allows pedotransfer functions, widely employed in agronomy for inference of complex soil performance attributes [e.g., hydraulic properties (Minasny and McBratney, 2002)] directly from spectra; the chemometrics presented herein provide strong evidence in support of the contention that VNIRS can predict difficult to measure parameters using an easy to measure set of reflectance attributes.

The strong associations between extracellular enzyme activities and spectra are particularly emblematic of this pedotransfer potential; we consider it highly unlikely that the low concentrations of extracellular enzymes in the soil matrix are influencing reflectance characteristics sufficiently to register a measurable effect, eliminating the possibility for direct mechanistic interpretation. However, the spectra appear to be effectively integrating the environment to which the microbial community is responding (e.g., organic matter quality and quantity, nutrient status and bioavailability), and therefore can provide a relatively reliable (and low-cost, rapid) predictor of a soil parameter that, while informative, is expensive and time consuming to measure. Further exploring this aspect of chemometric development using standard dosing experiments is considered a critical avenue of ongoing research. Our primary interest in integrative microbial indicators of soil quality emerged from previous ambiguity of soil chemical–physical attributes for direct indication of intermediate ecosystem impairment or situations where stressors overlap or have only recently been introduced; in cases of obvious nutrient enrichment or gross sedimentation they continue to work well. The ability to derive other complex pedotransfer functions directly from spectra [hydraulic properties (Cohen, 2003); carbon mineralization rates (Bouchard et al., 2002; Fystro, 2002); toxicity (Kooistra et al., 1997); soil erodibility and erosion and/or degradation condition (Cohen, 2003)] offers an avenue of continuing research and a potentially indispensable diagnostic tool for agricultural, ecological, and hydrologic assessments, particularly where spatial heterogeneity or statistical confounders necessitate large sample throughput.

Ranges for most parameters in this study exceeded those for which chemometrics have been developed (Shepherd and Walsh, 2002; Cozzalino and Moron, 2003; Dunn et al., 2002), primarily due to physical, chemical, and biological processes in wetland systems. Most notably, organic matter content ranged from approximately 0 to 82%; while some Histosols (e.g., those typical of the Everglades) may exhibit higher organic matter content, the range observed in this limited spectral reflectance library (SRL) is a useful test of wider applicability in wetland systems. Similarly, total phosphorus concentrations range from values characteristic of highly oligotrophic systems to values that generally indicate severe enrichment. Observed extracellular enzyme activities also vary over the range typical of riparian and freshwater wetland systems exposed to varying human disturbance (Prenger and Reddy, 2004).

In several cases, chemometric model performance improved over efficiencies previously observed (Shepherd and Walsh, 2002; Ben-Dor and Banin, 1995; Chang et al., 2001; Dunn et al., 2002); in particular, improved efficiency for predicting organic content was observed. This is likely due to the greater range in target values. Those attributes designed to characterize specific functional pools (e.g., Mehlich extractions, water-extractable P, and total organic C) performed more poorly than expected, but in all cases except water-extractable P and HCl-extractable calcium, the RPD statistic suggests potential for effective spectral inference, depending on the accuracy requirements of the application.

Categorical screening models, developed based on differences observed between reference and impacted sites, provide a useful extension of spectral inference for poorly predicted parameters or for indicator applications. In general, model screening accuracy was high, with only four parameters (H2O-extractable P, HCl-extractable Fe and Al, and acid phosphatase activity) having validation odds-ratios of less than 10; all models were statistically significant (p < 0.001). Poor performance of models for individual parameters may be due to interactions between some of the observed parameters, which can be significant confounders. For example Fe, Al, and Ca exert a significant influence on P dynamics and on each other (Lockaby and Walbridge, 1998), a phenomenon illustrated by the conditional correlations (Fig. 6, Table 5). As a result, we can effectively predict the aggregate quantity of P in a soil (RPD for total P = 2.40), but perform more poorly for Mehlich P (RPD = 1.70) that represents a functional subset of the total P pool. We predict, however, that the spectral pedotransfer functions for enzyme activity (acid phosphatase RPD = 2.62; ß-glucosidase RPD = 2.64) are likely to be more biologically and/or ecologically meaningful than an operationally defined acid extraction procedure.

The two parameters with the weakest continuous prediction accuracy (Table 2), Ca and Mg, are usually reported as a percentage of cation exchange capacity (CEC) for agronomic purposes, since one of their effects is to saturate the soil's ability to bind nutrients; previous work (Cohen, 2003) showed strong conditional associations between spectral predictions of Ca and Mg and predictions of CEC (r > 0.90). It may follow that where Ca and Mg occur in mineral forms (e.g., carbonates) not associated with soil cation binding sites, as is the case in the study region where low clay content, sandy soils have extremely low CEC values and regional geologic strata (though relatively deep in this area of Florida) consist of limestone and dolomite, spectral predictions are more challenging. In particular, while most sites had relatively low concentrations of these cations, significant variability (both Ca and Mg vary over two orders of magnitude) may overwhelm the predictive ability of our relatively small spectral library. Further work with larger sample sizes and more carefully controlled distinction between CEC bound cations vs. stable mineral forms should provide greater predictive capability.

One of the goals of using VNIRS-based predictive models is to develop an integrative measure of impact. This concept is an extension of efforts to use biological communities or functions as integrators of ecosystem change by including both biotic and abiotic factors. Since microbial activity is dependent on numerous interactive factors, including organic matter and nutrient concentrations, reflectance measurements may provide a higher order of integration than biological activities alone. Given the efficiency of binary discrimination between impaired and reference riparian soils, and the significance of contrasts between case and reference soils (Table 3, p values), VNIRS may offer an extremely useful tool for impact assessment (both mode and intensity) at the watershed scale.

The soil archive used in this study was selected to capture the range of biogeochemical variability typically observed in riparian systems. However, further work on wetland soils is needed to provide a systematic library that can be used in wetland systems throughout the state and ecoregion, ranging from highly mineral soils typical of isolated oligotrophic marshes to histosols typical of the Everglades and deep-water habitats. Soil physical attributes were not evaluated in this work despite their utility for riparian soil characterization; research to develop predictive models for physical properties (penetrability, texture, infiltration capacity, bulk density, water holding capacity) is needed.

The routine application of this technology for soil-based assessments can clearly take advantage of the ability to infer estimates of conventional soil attributes from spectra. However, our experience with ecological evaluations that employ standard biogeochemical attributes is that direct interpretation is confounded by context; specifically, there are interactions among soil attributes, and the effects of hydrology, community, and history (e.g., fire) make direct inference of condition problematic. We propose that ongoing VNIRS research focus on the unique integrative properties of spectral reflectance to develop indicators of condition that control for these confounders. Examples might include indicators of site-specific trends in soil quality or ecological condition related to long-term monitoring studies or evaluation of restoration success, measures of variability among soils from different vegetation communities, and soils responding to natural and induced gradients such as elevation, hydrologic variation, or nutrient and/or contaminant enrichment.


   ACKNOWLEDGMENTS
 
The authors gratefully acknowledge input from Dr. Mark Clark, University of Florida. Funds were provided by a seed grant from the School of Natural Resources and Environment at the University of Florida ("Calibrating Diffuse Reflectance Spectra to Indicators of Soil Quality in Wetlands").


   REFERENCES
TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS
 DISCUSSION
 REFERENCES
 





This Article
Right arrow Abstract Freely available
Right arrow Figures Only
Right arrow Full Text (PDF) Free
Right arrow Alert me when this article is cited
Right arrow Alert me if a correction is posted
Services
Right arrow Similar articles in this journal
Right arrow Similar articles in ISI Web of Science
Right arrow Similar articles in PubMed
Right arrow Alert me to new issues of the journal
Right arrow Download to citation manager
Citing Articles
Right arrow Citing Articles via ISI Web of Science (3)
Right arrow Citing Articles via Google Scholar
Google Scholar
Right arrow Articles by Cohen, M. J.
Right arrow Articles by DeBusk, W. F.