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


     

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 Alert me to new issues of the journal
Right arrow Download to citation manager
Right arrow reprints & permissions
Citing Articles
Right arrow Citing Articles via HighWire
Right arrow Citing Articles via ISI Web of Science (21)
Right arrow Citing Articles via Google Scholar
Google Scholar
Right arrow Articles by Shepherd, K. D.
Right arrow Articles by Vanlauwe, B.
Right arrow Search for Related Content
PubMed
Right arrow Articles by Shepherd, K. D.
Right arrow Articles by Vanlauwe, B.
Agricola
Right arrow Articles by Shepherd, K. D.
Right arrow Articles by Vanlauwe, B.
Related Collections
Right arrow Nutrient Management
Right arrow Plant Analysis
Right arrow Soil Fertility and Productivity
Right arrow Tropical Soil Management
Published in Agron. J. 95:1314-1322 (2003).
© American Society of Agronomy
677 S. Segoe Rd., Madison, WI 53711 USA

ORGANIC RESOURCES

Rapid Characterization of Organic Resource Quality for Soil and Livestock Management in Tropical Agroecosystems Using Near-Infrared Spectroscopy

Keith D. Shepherd*,a, Cheryl A. Palmb,c, Catherine N. Gachengob and Bernard Vanlauweb

a World Agroforestry Cent. (ICRAF), P.O. Box 30677-00100, Nairobi, Kenya
b Trop. Soil Biol. and Fertil. Inst. of CIAT (TSBF-CIAT), P.O. Box 30677-00100, Nairobi, Kenya
c The Earth Inst. at Columbia Univ., P.O. Box 1000, Palisades, NY 10964-8000

* Corresponding author (k.shepherd{at}cgiar.org).

Received for publication November 22, 2002.

   ABSTRACT
TOP
 ABSTRACT
INTRODUCTION
 MATERIALS AND METHODS
 RESULTS AND DISCUSSION
 CONCLUSIONS
 REFERENCES
 
Organic resources constitute a major source of nutrient inputs to both soils and livestock in smallholder tropical production systems. Determination of resource quality attributes using current laboratory methods is both timely and costly. This study tested visible and near-infrared (wavelengths from 0.35–2.50 µm) reflectance spectroscopy (NIRS) for rapid prediction of quality attributes for a diverse range of organic resources. A spectral library was constructed for 319 samples of oven-dried, ground plant material originating from green leaf (186 samples), litter (33), root (25), and stem (21) samples from 83 species including tropical crops and trees used for agroforestry and manure samples (39). Organic resource attributes were calibrated to first-derivative reflectance using regression trees with stochastic gradient boosting, and screening tests were developed for separating various organic resource quality classes using classification trees. Validation r2 values for actual vs. predicted values using a 25% holdout sample were 0.91 for N, 0.90 for total soluble polyphenol, and 0.64 for lignin concentration. Screening tests gave validation prediction efficiencies of 96% for detecting samples with high N concentration, 91% for low total soluble polyphenol, and 86% for low lignin concentration. The spectral screening tests were robust even at small (n = 48) calibrations sample sizes. Screening tests for detecting samples with low or high levels of P, K, Ca, and Mg gave prediction efficiencies of 74 to 92%. Near-infrared reflectance spectroscopy can be used to rapidly screen organic resource quality. Global spectral calibration libraries should be established for a range of resource quality attributes.

Abbreviations: NIRS, near-infrared spectroscopy


   INTRODUCTION
TOP
 ABSTRACT
 INTRODUCTION
MATERIALS AND METHODS
 RESULTS AND DISCUSSION
 CONCLUSIONS
 REFERENCES
 
ORGANIC RESOURCES CONSTITUTE a major source of nutrient inputs to both soils and livestock in smallholder tropical production systems. The quality of organic resources regulates the potential rate of decomposition and availability of those nutrients, both in the soil and in the rumen. Although the actual rate and degree of decomposition are moderated by the local activity of the decomposer organisms and the environmental conditions, plant litter quality is the factor most amenable to management in agricultural systems (Giller and Cadisch, 1997; Heal et al., 1997). A review of predictive understanding of effects of resource quality on decomposition and nutrient release in soils indicates a hierarchical set of N, lignin, and polyphenol concentration for predicting N release patterns from organic materials (Palm et al., 2001). Other resource quality parameters have been highlighted as important modifiers, including tannin, soluble C, and fiber-bound N (Palm and Rowland, 1997). However, the importance prescribed to individual constituents often results from the differences in the range of materials tested (Giller and Cadisch, 1997). Recently, efforts have been undertaken to compile global information on decomposition and resource quality attributes as a basis for more systematic experimentation and development of predictive models (Palm et al., 2001). Problems noted with finding patterns and trends with this global (tropical) database include the wide variety and lack of standardization of methods by which resource quality is analyzed. The size of the database is also limited by the tedious and costly laboratory methods for analyzing many of the resource quality parameters, including macronutrients, celluloses, and lignin.

Practical use of resource quality information will rely on ability to rapidly determine quality attributes on large numbers of samples of diverse composition. Situations that can benefit from rapid assessments include monitoring surveys of organic resource inputs by farmers, farm advisory services on organic resource quality for fodder or soil amelioration, and selection of materials for experiments on decomposition in relation to resource quality. Near-infrared spectroscopy has shown promise as a rapid, nondestructive method for determining a number of plant constituents and is now widely used in a number of agricultural and food applications (e.g., Burns and Ciurczak, 2001; Davies and Cho, 2002). The use of NIRS for protein determination in forage, feed, and grain analysis is well established (Shenk et al., 2001). Near-infrared spectroscopy has been used to determine total N, C, lignin, cellulose, ash content, acid detergent fiber, and acid detergent lignin contents of tree and shrub species (e.g., McLellan et al., 1991; Joffre et al., 1992; Meuret et al., 1993). Measures of decomposition in leaf litter such as C/N ratios, lignin/N ratios (Gillon et al., 1993), and in vitro gas production in forages (Herrero et al., 1996; Goodchild et al., 1998) have also been predicted.

Capability of NIRS for analysis of livestock manure and compost quality has been demonstrated for moisture, organic matter of solids, ammonium N, organic N, total N, and potentially mineralizable N (Reeves and Van Kessel, 2000; Qafoku et al., 2001; Reeves, 2001; Suehara et al., 2001; Malley, 2002). Near-infrared spectroscopy has in some cases been shown to be more accurate in predicting animal response to diet than any of the current reference methods (Abrams et al., 1987). Although there are no absorption bands for mineral species in the near-infrared region, there has been moderate success in their determination by NIRS (Clark et al., 1987), probably as a result of their association with protein, fiber, and organic acids (Shenk et al., 2001). However, previous studies on use of NIRS for organic resource characterization have typically been restricted to a limited range of materials and often limited numbers of samples. Global calibrations that are designed to analyze diverse types of organic resources are needed for practical application in organic resource management in the tropics, which typically includes an array and mix of organic resources. The objectives of this study were to test the robustness of NIRS for rapid prediction of organic resource quality across a diverse range of organic materials and to develop spectral screening tests for organic resource quality.


   MATERIALS AND METHODS
TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
RESULTS AND DISCUSSION
 CONCLUSIONS
 REFERENCES
 
Organic Resource Library
We identified all available archived plant samples from the Organic Resource Database, developed by the Tropical Soil Biology and Fertility Institute of the Centro Internacional de Agricultura Tropical (TSBF-CIAT) and Imperial College Wye (Palm et al., 2001). The Organic Resource Database was initiated in January 1997 with the aim to collate existing data on plant quality characteristics, primarily chemical attributes including macronutrients, lignin, total soluble polyphenol concentrations, decomposition behavior in soils, and animal feed value. The available archived samples, all of which had been analyzed by standard methods, consisted of 319 samples with complete attribute data and 514 samples with data for N concentration only. The samples were oven-dried at 30°C, ground to pass a 1-mm sieve, and then stored in closed plastic bottles at room temperature in Nairobi, Kenya. All chemical analyses were performed by TSBF at the Kenya Agricultural Research Laboratories in Muguga, Kenya, using standard methods (Anderson and Ingram, 1993). Repeated analysis of the samples suggested little change in the plant quality variables during storage.

Plant material was analyzed for concentrations of N, lignin, total soluble polyphenol, P, K, Ca, and Mg. Samples were digested with H2SO4 and H2O2 by micro-Kjeldahl and extracts analyzed for total N, P, and K (Parkinson and Allen, 1975). Calcium and Mg were determined by atomic absorption spectrophotometry, K by flame photometry, and N by steam distillation followed by titration with HCl. Lignin was determined by the acid detergent fiber method (Van Soest and Wine, 1968) using a fiber analyzer (ANKOM Technol., Macedon, NY) and total soluble polyphenol by a revised Folin–Denis method (King and Heath, 1967; Constantinides and Fownes, 1994). Soluble polyphenol was extracted in 50% (v/v) aqueous methanol for 1 h in a water bath (80°C) with a plant extract ratio of 0.1 g per 50 mL instead of 0.75 g per 50 mL (Constantinides and Fownes, 1994). Total soluble polyphenol concentration was analyzed colorimetrically against a tannic acid standard and expressed as tannic acid equivalents.

Reflectance Measurements
Diffuse reflectance spectra were recorded for each archived sample using a FieldSpecTM FR spectroradiometer (Analytical Spectral Devices, Boulder, CO) at wavelengths from 0.35 to 2.5 µm with a spectral sampling interval of 0.001 µm. Enough plant sample was placed into 7.4-cm-diam. Duran glass Petri dishes to give a sample thickness of about 1 cm. The samples were scanned through the bottom of the Petri dishes (Fig. 1) using a high-intensity source probe (Analytical Spectral Devices, Boulder, CO). The probe illuminates the sample (4.5 W halogen lamp giving a correlated color temperature of 3000 K; WelchAllyn, Skaneatles Falls, NY) and collects the reflected light from a 3.5-cm-diam. sapphire window through a fiber-optic cable.



View larger version (126K):
[in this window]
[in a new window]
  		Fig. 1. Portable spectrometer with high-intensity reflectance probe used for reflectance measurements. Samples are illuminated through the bottom of a glass Petri dish, and reflected light is captured from a 35-mm-diam. window through a fiber-optic cable.

 
To sample within-dish variation, reflectance spectra were recorded at two positions, successively rotating the sample dish through 90° between readings. The average of 25 spectra (the manufacturer's default value) was recorded at each position to minimize instrument noise. Before reading each sample, 10 white reference spectra were recorded using calibrated spectralon (Labsphere, Sutton, NH) placed in a glass Petri dish. Reflectance readings for each wavelength band were expressed relative to the average of the white reference readings. With this method, a single operator can comfortably scan 500 samples a day.

Statistical Methods
Multivariate relationships among resource quality attributes were analyzed to establish to what degree good spectral calibrations could have been due to interdependencies among resource quality variables. Conditional independence assumptions among quality attributes were tested using graphical linear-modeling approaches (Edwards, 2000). Graphical modeling is a way of examining the multivariate dependencies in a sample of data from an observations study. The models describe the structure of dependencies among the variables, allowing the conditional as well as the marginal associations to be studied. By considering the conditional associations among variables, the approach helps to identify spurious associations that can occur when studying only marginal pairwise associations among variables. Where necessary, Box–Cox transformations (Box and Cox, 1964) were applied before analysis to obtain approximately multivariate normally distributed values.

The raw spectral reflectance data were preprocessed before statistical analysis as follows. Relative reflectance spectra were resampled by selecting every 100th micrometer value from 0.35 to 2.5 µm. This was done to reduce the volume of data for analysis and to match it more closely to the spectral resolution of the instrument (0.003–0.01 µm). The reflectance values were then transformed with first-derivative processing (differentiation with second-order polynomial smoothing with a window width of 0.02 µm) using a Savitzky–Golay filter, as described by Fearn (2000). Derivative transformation is known to minimize variation among samples caused by variation in grinding and optical setup (Martens and Naes, 1989). Multiplicative scatter correction (used to compensate for additive and/or multiplicative effects in spectral data) and normalization (sample-wise scaling) of the reflectance data (both described in Vandeginste et al., 1998) did not improve calibrations and so were not used. Wavebands in regions of low signal/noise ratio or displaying noise due to splicing between the individual spectrometers (Analytical Spectral Devices, 1997) were omitted, leaving 205 wavebands for analysis. The omitted bands were 0.35 to 0.39, 0.97 to 1.01, and 2.50 µm.

Individual plant quality attributes from the chemical analyses were then calibrated against the reflectance wavebands. Calibrations were compared using (i) all 205 reflectance wavebands and (ii) the 148 reflectance wavebands in the near-infrared range only from 1.02 to 2.49 µm. Calibrations were done using regression models implemented by TreeNet stochastic gradient boosting (Steinberg et al., 2002). TreeNet is a pattern recognition technique for predicting a response variable from a set of independent variables using decision trees (tree-structured classifiers; Ripley, 1996). With regression trees, the construction process splits the observations into subsets, according to whether or not they are less than a particular value of one of the independent variables. The aim is to form subsets that have similar values for the target variable. The predicted value of the response variable for each node (branch) of the tree is the mean of its value for the subset of observations at that node. The best predictor is chosen using a variety of impurity or diversity measures. TreeNet constructs a model by constructing and combining many small trees, each typically no larger than two to eight terminal nodes. During this training process, each tree improves on its predecessors through boosting, a method for improving the accuracy of the learning algorithm. Gradient boosting constructs additive regression models by sequentially fitting a simple parameterized function to current residuals by least squares at each iteration (Friedman, 1999a, 1999b). Randomness is incorporated into the procedure to improve the execution speed and model robustness.

TreeNet has advantages over more conventional approaches in that: (i) no prior variable selection or data reduction is required, (ii) data do not need to be rescaled or transformed, (iii) the procedure is resistant to outliers in predictors or the target variable, (iv) it is generally robust to partially inaccurate data, (v) it is fast, and (vi) it is resistant to overtraining. TreeNet was found to give the best average prediction performance on holdout validation samples compared with several alternative calibration methods. These alternative methods included partial least-squares regression; classification and regression trees (CART; Brieman et al., 1984; Steinberg and Colla, 1997), including the use of bootstrap aggregation and adaptive resampling and combining (i.e., averaging of a large number of trees generated by resampling and replacement from the original training data); and multivariate adaptive regression splines (MARS; Friedman, 1991; Steinberg et al., 2001).

For each plant quality attribute, calibration models were developed on a random sample of three-quarters of the plant samples in the entire library. The calibrations were tested by predicting the plant variables on validation data sets composed of the remaining one-quarter of the samples. No samples were omitted from the analysis in either the calibration or validation data sets. Prediction success was evaluated on predicted and actual observations using the coefficient of determination (r2), root mean square error, and bias.

To test the ability of NIRS to classify organic resource samples into quality categories, a number of screening tests were defined for threshold values of selected organic resource quality variables. The aim of screening tests is to evaluate how successful a particular test is in diagnosing abnormal as distinct from normal cases (e.g., Jones and Payne, 1997). Samples were classified either as abnormal or normal based on a cutoff value defined by the critical limits for predicting N release rates established by Palm et al. (2001): N > 25 mg kg-1, lignin < 150 mg kg-1, and total soluble polyphenol < 40 mg kg-1. Critical limits for other nutrients were defined as the 25th and 75th percentiles of the actual data. Classification models, implemented in TreeNet, were used to develop calibrations for each screening test with the 148 reflectance wavebands as dependent variables. One-quarter of the data set was held out at random as a validation data set. Again, no samples were omitted from the analysis in either the calibration or validation data sets. Predictive performance was assessed using the sensitivity (percentage of abnormal cases correctly predicted), specificity (percentage of normal cases correctly predicted), and the positive likelihood ratio [percentage sensitivity/[100 - percentage specificity)], which indicates the value of the test for increasing certainty about a positive diagnosis (Jones and Payne, 1997).

To evaluate the stability of the classification trees using a small number of samples, the size of the data set used for training was varied systematically from 10 to 75% of the total number of samples, with the remainder of the samples used for validation. The evaluation was repeated three times, using a different seed value to generate the random samples each time, and the results averaged.


   RESULTS AND DISCUSSION
TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS AND DISCUSSION
CONCLUSIONS
 REFERENCES
 
Organic Resource Attributes
The organic resource spectral library for which there was complete data on N, lignin, and total soluble polyphenol (n = 319) contained a diverse range of plant materials (Table 1). The samples included 83 species from 16 families, with a total number of samples of leaf, 186; litter, 33; root, 25; and stem, 21. Additional samples for which N concentration data was available (n = 195) were made up of samples of maize (Zea mays L.), 83; manure, 42; Lablab sp. (tropical grain legumes), 41; and unidentified weeds, 15.


View this table:
[in this window]
[in a new window]
  		Table 1. Numbers of species and number of samples of a particular plant part in the organic resource spectral library (n = 319).

 
Although not all analyses were available for all organic resource samples, there was wide variation in individual attributes in the library (Table 2). The ranges in N, P, lignin, and total soluble polyphenol concentrations are similar to the ranges reported by Palm et al. (2001) for the entire Organic Resource Database of 1929 entries. The 39 manure samples also showed a wide range of attributes with 2.5th to 97.5th percentile ranges of 5 to 52 mg kg-1 for N and 61 to 190 mg kg-1 for lignin although total soluble polyphenol values for manure were in the low range at 0.5 to 20 mg kg-1. The correlation and partial correlation coefficients among N, lignin, and total soluble polyphenol concentrations (Table 3) were low enough to indicate a reasonable spread of these three attributes in the multivariate data space. Phosphorus concentration was significantly correlated with N concentration (r = 0.48; n = 178), but there was no significant correlation among any other pairs of quality variables.


View this table:
[in this window]
[in a new window]
  		Table 2. Number of plant samples and percentile{dagger} values for each plant property in the organic resource spectral library.

 

View this table:
[in this window]
[in a new window]
  		Table 3. Correlation coefficients (italic type) and partial correlation coefficients (roman type) for lignin, N, and total soluble polyphenol concentrations (n = 319).

 
Spectral Features
Reflectance spectra of different types of organic resource materials (Fig. 2) displayed increasing absorbance with decreasing wavelength in the visible range (0.4–0.7 µm), which is the edge of a large absorption feature mainly caused by lignin, centered at 0.28 µm (Schubert, 1965). The leaf materials showed distinctive reflectance features, typical of green leaf material in the visible region (Elvidge, 1990), characterized by intense pigment absorptions in the blue, a well-pronounced chlorophyll a absorption near 0.68 µm, and steep rise in reflectance at the red edge between 0.68 and 0.70 µm. The chlorophyll absorption features were absent in the manure and sawdust samples and are typical of originally dry plant materials. The litter sample retained a weak absorption at 0.68 µm, similar to spectra of chlorotic leaf material (Elvidge, 1990). These patterns are consistent with a general deterioration of the absorption wing in the 0.4- to 0.9-µm region as plant decay progresses and litter quality declines.



View larger version (40K):
[in this window]
[in a new window]
  		Fig. 2. Reflectance spectra for selected contrasting materials from the organic resource spectral library.

 
There were relatively strong absorption features in the near infrared centered at 1.47, 1.74, 1.94, and 2.10 µm and weaker features at 1.59, 1.73, 1.78, 2.28, 2.32, 2.34, 2.36, and 2.38 µm. Other workers have reported similar absorption features in green and dry plant materials. Elvidge (1990) found that holocellulose spectral features tended to dominate in brown wood samples with well-developed absorptions at 2.10, 2.28, and 2.34 µm, whereas lignin spectral features tended to predominate in decayed plant materials, with a broad absorption between 2.05 and 2.14 µm, a sharply defined absorption at 2.27 µm, and distinct absorptions at 2.33 and 2.38 µm. McClure et al. (2002) reported major protein absorption bands centered at 2.06 and 2.27 µm. Shenk et al. (2001) reported basic characterizing wavelengths of 2.18 µm for protein, 2.27 µm for lignin, and 2.34 µm for cellulose. They also give tentative band assignments for lignin at 1.17, 1.41, 1.42, 1.44, and 1.68 µm. A strong lignin signature was particularly noticeable in the sawdust and litter samples and was weak in the manure sample. During plant decay, there is generally loss of holocellular spectral features and an increase in lignin spectral features (Elvidge, 1990). However, many different plant substances also have absorption features at similar wavelengths so that such interpretations may have limited utility. For example, D-ribulose 1-5-diphosphate carboxylase, the most abundant N-bearing compound in green leaves, also has corresponding features at 1.50, 1.74, 1.94, 2.05, and 2.17 µm (Elvidge, 1990). Based on literature values tabulated by Shenk et al. (2001), there are 32 protein absorption bands (rounded to nearest 0.01 µm) distributed across the near infrared from 1.03 to 2.35 µm.

Spectral Correlation
Nitrogen concentration displayed high positive or negative correlation (absolute value of r > 0.6) with derivative spectra around the green- and red-edge regions of the visible range, centered at 1.49, 1.75, 1.94, 2.07, 2.14, 2.28, 2.32 µm (Fig. 3) . These correspond with the absorption features described above for protein. Lignin showed moderate correlation (absolute value of r > 0.4) peaks in the green and red regions and at 1.47, 1.57, and 1.93 µm in the near infrared (Fig. 3). Interestingly, these near-infrared peaks do not correspond with Elvidge's visual interpretations of lignin absorption features although smaller peaks were evident at 2.07, 2.16, 2.23, 2.28, 2.33, and 2.39 µm, which do correspond with the reported spectral features for lignin. Total soluble polyphenol did not display high correlation (absolute value of r > 0.4) with peaks in the visible range, although the red-edge effect was evident, but had peaks in the near-infrared range at 1.65, 1.77, 2.10, and 2.20 µm (Fig. 3).



View larger version (33K):
[in this window]
[in a new window]
  		Fig. 3. Correlation of N, lignin, and total soluble polyphenol concentrations with first derivatives of relative reflectance at different wavelengths. Lignin and total soluble polyphenol were loge–transformed before processing.

 
Prediction of Organic Resource Attributes Using TreeNet Regression
Calibrations judged using the 25% holdout validation r2 values (Fig. 4) were good for total soluble polyphenol (0.90) and N (0.85); moderate for K (0.68), lignin (0.64), and Ca (0.50); and poor for P (0.35) and Mg (0.35). The extreme outlier in the validation plot for N concentration (Fig. 4) was a manure sample whose actual value (52 mg kg-1 N) was found to be a far outlier (>3.0 interquartile ranges from the upper quartile) within the population of manure samples (n = 81). Removal of this far outlier from the validation scatterplot gave an improved fit with r2 of 0.91, bias of 0.08, and root square mean error of 4.0 g kg-1. The second largest outlier in the scatterplot (actual value of 66 mg kg-1 N) was leaf material of Crotolaria sp. and was a near outlier (>1.5 interquartile ranges from the upper quartile) within this sample population (n = 11). Removing this outlier further reduced the validation root mean square error to 3.5 g kg-1.



View larger version (24K):
[in this window]
[in a new window]
  		Fig. 4. Validation scatter plots of actual against predicted values for (top) N and (bottom) total soluble polyphenol concentrations. Calibration models were developed with TreeNet stochastic gradient boosting using a random selection of 75% of the total number of samples and validated on the remaining 25% holdout sample. No outliers were removed from either calibration or validation data sets. Root mean square errors (RMSE) are also given.

 
The wavebands that contributed most to predictive power in the TreeNet models (Table 4) generally corresponded well with the correlation peaks in Fig. 3. Including wavebands from 0.40 to 0.96 µm in the calibration runs did not improve predictions, despite the presence of high correlation peaks for N and lignin in the visible range. The correlation peaks in the visible range may have been simply overtones of peaks in the near-infrared range. TreeNet regression, compared with the more conventionally used partial least-squares regression technique, gave a large improvement in predictive performance for total soluble polyphenol (36% reduction in root mean square error) and small improvements in predictive performance for N and lignin (5% reduction in root mean square error).


View this table:
[in this window]
[in a new window]
  		Table 4. Wavebands with highest contribution to TreeNet calibration models for N, lignin, and total soluble polyphenol concentrations. The wavebands are sorted in order of decreasing importance.

 
Screening Tests for Organic Resource Attributes Using TreeNet Classification
For many applications, such as organic resource quality assessments, it is often sufficient to classify a resource with respect to a critical test value or quality class, rather than needing a precise estimate of constituent concentrations. Based on the 25% holdout validation data, good predictive performance was achieved (efficiencies of 86–96%) for spectral screening of organic resources into classes of high and low concentrations of N, lignin, and total soluble polyphenol (Table 5). Positive likelihood ratios for these tests were 41, 6.2, and 4.7, respectively, indicating that an abnormal sample would be 5 to 41 times more likely to be abnormal than normal given a positive test result. Even where TreeNet regressions were poor, screening tests gave good results (efficiencies of 74–92%; Table 5). For example, for screening samples for low P concentration (<1.2 mg kg-1), the test gave an efficiency of 89% with a positive likelihood ratio of 10. A screening test to identify samples with high N (>25 mg kg-1) but low P concentration (<2.5 mg kg-1), that might indicate high risk of N leaching from this organic source on P-deficient soils, gave an efficiency of 90% and likelihood ratio of 8. The diagnostic value of a test, however, depends on the prevalence of abnormal cases in the population sampled. For instance, as prevalence of abnormal cases drops from 80 to 20%, the posttest probability of an individual sample being abnormal will fall from 98 to 71% for a likelihood ratio of 10 and from 95 to 56% for a likelihood ratio of 5.


View this table:
[in this window]
[in a new window]
  		Table 5. Prediction of organic resource quality tests from reflectance spectra using TreeNet stochastic gradient boosting for a 25% holdout validation data set.

 
The prediction success in the P screening tests could have been due to the correlation between P and N concentration in the samples. Using the validation data from the screening tests (n = 90), we tested the conditional independence assumptions between actual and predicted values for the N and P with graphical modeling, using a coherent backward selection procedure. The test showed that the relationship between actual and predicted class for N > 25 mg kg-1 and between actual and predicted class for P < 1.2 mg kg-1 displayed conditional dependence (P = 0.05). Therefore NIRS provided additional information on P concentration over and above that provided by prediction of N concentration.

The screening tests for N, lignin, and polyphenol were remarkably robust when tested with small training sample sizes (Fig. 5) . For the high-N test, there was little evidence for decay in sensitivity or specificity as the training sample was decreased from 75 to 10% of the total data set. For the low-lignin and low-polyphenol tests, sensitivity and specificity were >80% when only 15% of the total data was used for training.



View larger version (16K):
[in this window]
[in a new window]
  		Fig. 5. Response of (top) sensitivity and (bottom) specificity to size of training data set in spectral screening tests for high N (<25 mg kg-1), low lignin (<110 mg kg-1), and low total soluble polyphenol (<40 mg kg-1) concentrations. The results are for the proportion of the total data set (n = 319) not used for training. Sensitivity is the percentage of abnormal cases correctly classified, and specificity is the percentage of abnormal cases correctly classified.

 
The screening of samples into organic resource management categories as described by Palm et al. (2001), but using a lignin cutoff value of 110 g kg-1 (median value) instead of 150 g kg-1, gave high predictive performance (Table 6). Much of the confusion in the validation tests was between neighboring classes rather than among extreme classes, and misclassification was greatest for Category 2 (high N with low lignin and polyphenol).


View this table:
[in this window]
[in a new window]
  		Table 6. Prediction of four organic resource management categories from reflectance spectra using TreeNet stochastic gradient boosting for a 25% holdout validation data set.

 
Using classification trees (CART), we reanalyzed previously published data on N release as a function of initial concentrations of N, lignin, and total soluble polyphenol (Palm et al., 2001). The percentage of N released after 8 wk of incubation was classified into four ranges, representing fast N immobilization, slow immobilization, slow mineralization, and fast mineralization (Table 7). Models were tested using 10-fold cross validation, whereby random selections of one-tenth of the data are iteratively used for validation (Brieman et al., 1984). Lignin concentration had no predictive power in classifying N release once N concentration was included in the model. Total soluble polyphenol concentration significantly reduced confusion among classes of slow and fast N mineralization but gave no improvement compared with N concentration alone in identifying materials likely to immobilize N. The splitting rules (Table 7) gave cutoff limits that are very similar to those suggested by Palm et al. (2001) for identifying the resource management categories in Table 6, but the CART analysis suggested that for screening low-quality materials with fast N immobilization rates, N concentration <11 mg kg-1 can be used as a rule instead of including lignin concentration. These patterns correspond well with recent African studies on tropical legumes that found polyphenols, but not lignin, slowed initial N release in materials with high N concentration (Vanlauwe et al., 2002). Because N and total soluble polyphenol concentrations both calibrated well to reflectance spectra, we expect spectral screening tests will accurately predict N release patterns. More samples with high lignin (>200 mg kg-1) should be included in future studies to further test the effect of lignin on decomposition rates and to extend the spectral calibrations.


View this table:
[in this window]
[in a new window]
  		Table 7. Classification tree predictions of published data from 11 incubations studies (Palm et al., 2001) on N mineralized or immobilized after 8 wk of incubation from initial N and total soluble polyphenol concentrations. Prediction success is for 10-fold cross validation.

 

   CONCLUSIONS
TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS AND DISCUSSION
 CONCLUSIONS
REFERENCES
 
Near-infrared spectroscopy can provide rapid and accurate prediction of N and total soluble polyphenol concentrations with a single global calibration across a wide range of organic resources. Because these variables are principal determinants of N mineralization and immobilization rates among organic resources, spectral screening tests can be used to characterize this aspect of resource quality. Spectral screening tests of moderate accuracy can also be developed for diverse organic resource materials that identify samples with high or low concentrations of lignin and minerals. Near-infrared reflectance spectroscopy prediction of N concentration and screening of organic resource quality are of potential value in studies where large numbers of samples are needed and will allow variability to be more adequately sampled than with conventional approaches. Thus, NIRS may be useful in studies on management of organic resource inputs by farmers, in nutrient budget studies, for determination of fodder quality, and for routine testing for advisory purposes. Where more accurate measurements are required, spectral screening may allow reduction in costs of conventional laboratory determinations by providing a means of sample stratification.

We recommend that the Organic Resource Database concept of Palm et al. (2001) and similar work by Jensen et al. (2002) be extended to include a global spectral library with calibrations, as suggested for soil spectral libraries by Shepherd and Walsh (2002). In this way, spectral outliers detected among new samples can be added to the calibration libraries, thereby increasing the predictive value of the spectral library for global use as new samples are added. This would be much more efficient than current practice where data from samples analyzed in different laboratories around the world have no value beyond the immediate study of interest. Future studies should test calibration of decomposition characteristics directly to near-infrared reflectance and also compare NIRS predictive accuracy with standard errors of duplication for standard chemical reference analyses.


   ACKNOWLEDGMENTS
 
We thank Elvis Weullow for technical support on spectroscopy and Andrew Sila for assistance with data management. We gratefully acknowledge the Rockefeller Foundation for financial support.


   REFERENCES
TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS AND DISCUSSION
 CONCLUSIONS
 REFERENCES
 




This article has been cited by other articles:


Home page
 Soil Sci.Home page
B. Vanlauwe, C. Gachengo, K. Shepherd, E. Barrios, G. Cadisch, and C. A. Palm
Laboratory Validation of a Resource Quality-Based Conceptual Framework for Organic Matter Management
Soil Sci. Soc. Am. J., June 2, 2005; 69(4): 1135 - 1145.
[Abstract] [Full Text] [PDF]

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 Alert me to new issues of the journal
Right arrow Download to citation manager
Right arrow reprints & permissions
Citing Articles
Right arrow Citing Articles via HighWire
Right arrow Citing Articles via ISI Web of Science (21)
Right arrow Citing Articles via Google Scholar
Google Scholar
Right arrow Articles by Shepherd, K. D.
Right arrow Articles by Vanlauwe, B.
Right arrow Search for Related Content
PubMed
Right arrow Articles by Shepherd, K. D.
Right arrow Articles by Vanlauwe, B.
Agricola
Right arrow Articles by Shepherd, K. D.
Right arrow Articles by Vanlauwe, B.
Related Collections
Right arrow Nutrient Management
Right arrow Plant Analysis
Right arrow Soil Fertility and Productivity
Right arrow Tropical Soil Management

HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS
The SCI Journals Crop Science Vadose Zone Journal
Journal of Natural Resources
and Life Sciences Education		 Soil Science Society of America Journal
Journal of Plant Registrations Journal of
Environmental Quality		 The Plant Genome