| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
a Plant Research International, P.O. Box 16, 6700 AA, Wageningen, the Netherlands
b Animal Sciences Group, Applied Research, Wageningen Univ. and Research Centre, P.O. Box 65, 8200 AB Lelystad, the Netherlands (A.G.T. Schut, present address: Dep. of Spatial Sciences, Curtin Univ. of Technology, GPO Box U1987, Perth, Australia)
* Corresponding author (tschut{at}agric.wa.gov.au)
Received for publication August 5, 2005.
 |
ABSTRACT |
|---|
 TOP ABSTRACT INTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSION AND CONCLUSIONSREFERENCES |
|---|
Abbreviations: CCD, charge coupled device • DM, dry matter • PLS, partial least squares • RMSEP, root mean squared error of prediction • RMSECV, root mean squared error of cross validation
|
INTRODUCTION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSION AND CONCLUSIONSREFERENCES |
|---|
Spectroscopy may provide the means to measure grass quantity or quality, specifically NIRS (near-infrared spectroscopy) in the laboratory and remote sensing in the field. Near-infrared spectroscopy is based on diffuse reflectance of ground samples and is widely used for laboratory measurement of the concentration of nutrients and feeding value in dried and fresh crop material (Berardo, 1997; Park et al., 1998; Paul and Häusler, 2002; Ruano-Ramos et al., 1999). Remote sensing has been used extensively to measure biomass in crops and a limited number of quality aspects (e.g., protein content, lignin, and cellulose; Bausch et al., 1998; Blackburn, 1998; Moran et al., 1997; Zagolski et al., 1996). Although suitable for measuring large areas, the usefulness of remotely sensed reflectance of visual and near-infrared light is limited because soil and dead material confound reflectance when ground coverage of the canopy is incomplete and the atmosphere is nontransparent in specific wavelength regions.
Based on literature, it is expected that quality aspects such as energy content, digestibility, and fiber content and composition can only be detected with high-resolution and continuous spectra (Curran, 1989; Curran et al., 1992; Jacquemoud et al., 1995). Imaging spectroscopy provides the means to measure the reflectance of light with an array of detectors, combining a high spectral and spatial resolution. Imaging spectroscopy brings the concept of NIRS one step further, as it measures the in situ leaf reflectance with high spectral and high spatial resolution in the visible and near-infrared parts of the spectrum. In theory, yield and quality aspects can be measured with sensors from any available platform (on the ground, airborne, or in space), although the availability of hyperspectral sensors on air- or space-borne platforms is severely limited. In our work, we have concentrated on an imaging spectroscopy system that measures reflectance at a distance of 1.3 m from the soil surface. This system has the advantage that the size of a pixel can be as small as 1 mm2. This makes it possible to easily differentiate between pixels that represent soil, dead plant material, and green plant material. This system makes it possible to use artificial light and limits the disturbing influence of the atmosphere. Finally, a novel design of light sources can be used to provide additional information on crop density and canopy geometry (Schut et al., 2002).
In previous work, this concept of close-range imaging spectroscopy was tested on mini-swards under semicontrolled conditions with a semistationary system. It was concluded that imaging spectroscopy is an accurate technique for measurement of yield, N deficiency, drought stress, nutrient content, and feeding value (see Schut et al., 2005, and references therein). To test this methodology under field conditions, the Imspector Mobile was built that records reflection from a distance of 1.3 m from the soil surface in two-dimensional images and hyperspectral image lines while driving (Molema et al., 2003). This system is designed for the assessment of experimental and practical fields and uses high-intensity flash lights to maintain illumination levels while minimizing shutter opening to control image blurring. Based on this work, we anticipate that sophisticated sensors will be developed for practical applications as an integral part of farm or lawn machinery or as a stand-alone application for extension services, either using on-the-ground or air-borne platforms. The objective of this study was to evaluate the accuracy of on-farm measurements of DM yield, nutrient content, and feeding value, using imaging spectroscopy. To this end, the Imspector Mobile was calibrated and validated using measurements in several experimental fields, after which yield and quality were measured on two farms.
|
MATERIALS AND METHODS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSION AND CONCLUSIONSREFERENCES |
|---|
The Imspector Mobile can record one image per second for each sensor. This recording frequency is limited by the time required for recharging the flash light capacitors. Driving speed during recording was maintained between 0.3 and 0.5 m s–1, resulting in 8 to 15 recorded images per sensor per plot, depending on plot size.
Data from Experiments
Three data sets were collected in experimental fields in the Netherlands (Table 1). The first data set comprised swards on sand and clay soils that consisted mainly of perennial ryegrass (Lolium perenne L.). This data set included data from experiments where one or more of the following factors were studied: urine and P supply, method and timing of sward renewal, and N supply. Treatments in the urine application experiment received 300 kg N ha–1 yr–1 as fertilizer and no (control) or 400 kg N ha–1 as artificial urine after harvest no. 1, 2, 3, 4, or 5. In the P experiment, P was supplied at 0, 50, or 100% of the advised amount. In the sward renewal trials, swards were not renewed (control) or were reseeded in autumn 2002 without soil tillage, plowed in autumn 2002 and reseeded in spring 2003, and plowed and reseeded in spring 2002, autumn 2002, or spring 2003. Within each treatment, annual N application was 0, 150, 300, or 450 kg N ha–1. Treatments with 300 kg N ha–1 were duplicated and also P supply was varied (normal or no P supply). Images were recorded between June 2003 and October 2004.
|
The third data set comprised swards containing mixtures of L. perenne and white clover (Trifolium repens L.) from an experiment where application of N (0, 20, 40, or 60 kg N yr–1) and application of manure were varied (0–40 t yr–1). Grass and clover were manually separated and weighed. Images were recorded between June 2003 and October 2004.
Net plot sizes in all experiments were 1.5 m wide and either 8 or 10 m long. All plots in all experiments were harvested with a forage harvester, cutting swards 5 cm above the soil surface. All harvested material was collected and weighed. From all harvested material, a subsample was taken and gravimetrically and chemically analyzed in the laboratory for DM content and contents of N, P, K, sugar, crude fiber, and ash. For the peat data set, neutral and acid detergent fiber, acid detergent lignin, and in-vitro digestibility (Tilley and Terry, 1963) also were determined. Samples included in the peat data set were analyzed at the Institute of Animal Husbandry and samples in the sand and clay data set were analyzed at the Dutch Agricultural Laboratory in Oosterbeek, the Netherlands.
Data from Farm Fields
Two data sets were collected on farms in the Netherlands. The first farm was experimental farm De Marke, near Hengelo (Gelderland). Fields 6 and 7 were used. Many weeds were present, mainly dandelion (Taraxacum officinale L.) and shepherd's-purse [Capsella bursa-pastoris (L.) Medik.]; white clover was also present. The fields were harvested on 3 May 2004. Immediately after harvest, manure was applied with a manure injection system. Images were recorded before the harvest of 3 May and on 12 May 2004.
The second farm was the farm of Mr and Mrs Van Wijk in Beuningen. Adjacent fields 8, 9, and 10 were used. There was hardly any white clover present in these fields, but dandelions were abundant. Images were recorded before the second harvest on 4 June 2004.
The fields were recorded in an hourglass pattern. Images were recorded in strips of 10 m length and 0.6 m wide with 15 to 20 images recorded within each strip. A prediction was made for each strip, using the grass–clover data set as the training set for the prediction model (see below) used on data from De Marke and the sand and clay data set as the training set for the prediction model used on data from the Van Wijk farm.
Image Analysis
Image lines from the V9 and N17 sensors were analyzed with a pixel-based classification procedure (Schut and Ketelaars, 2003b). First, for each sensor, pixels were assigned to three major classes: soil, dead plant material, or green plant material. Ground coverage was calculated as the percentage of pixels assigned to the green material class. Within the green material class, a subclass for pixels with leaves with specular reflection was created. All remaining pixels were subdivided into intensity classes. From all pixels within the green material class, excluding pixels in the subclasses with specular reflecting leaves or the lowest reflection intensity, a mean spectral curve was calculated. This mean spectral curve was normalized by dividing the spectra by the mean reflection between 743 and 955 nm for the V9 sensor and between 1070 and 1130 nm for the N17 sensor. A spline function was fitted through the mean normalized spectra (Silverman, 1985). This spline was resampled at intervals of 3 nm for the V9 sensor and 6 nm for the N17 sensor.
Images from the 3CCD sensor were analyzed to determine image-based measures: heterogeneity and image texture measures for ground coverage of green material. The index of reflection intensity, calculated as the percentage of pixels classified as green vegetation with high reflection intensity, was used to characterize the abundance of pixels with high reflectance. Heterogeneity was calculated as the standard deviation of ground coverage estimates within areas of 0.25, 2.25, and 2.4 dm2. The fraction of the image covered by "large" clover leaves was calculated by clustering neighboring pixels above a threshold area with a more or less homogeneous reflection. The fraction of the image covered by these leaf clusters is positively correlated with the white clover content of the sward.
The wavelet entropy is a measure of the energy distribution of wavelet frequencies that are required to describe a signal (Rosso et al., 2001; Schut and Ketelaars, 2003a). The wavelet entropy was calculated for 25 transects of 1024 neighboring pixels in two directions within a 3CCD image. A full description of the wavelet procedures used and the ability to discriminate between grass and clover swards with wavelet entropy can be found in Schut (2003).
Preprocessing
Laboratory and spectroscopic data were screened to remove outliers. Samples with ash content >250 g kg–1 were considered to have been contaminated with sand and were removed. Samples that deviated strongly in DM yield and N content from the other replicates within the same treatment (indicating errors in sample labeling) were removed. Spectra were checked visually and when extremely noisy, indicating a malfunction in the recording process, were removed.
Laboratory and spectroscopic data were normalized across the complete data set to zero mean and unit variance. These normalized data were then used in a PLS (partial least squares) model as described below.
For the grass–clover data set, the PLS procedure was used to predict relative values, i.e., predict the value relative to the mean of a unique combination of experiment and harvest number (a group). This was done by normalizing the data per group instead of normalizing the full data set at once. This procedure is comparable with rank orders, but here the relative distance between ranked observations does vary, in contrast to normal rank orders. Due to this normalization per group, the relative weight of experiments with large variability is decreased and the weight of groups with small variability is increased.
Predictor data derived from the 3CCD sensor comprised ground coverage, index of reflection intensity, wavelet entropy, fraction of clusters, and mean reflection of green pixels at 600, 710, and 800 nm. Predictor data from the V9 and N17 sensors comprised reflection percentages at interpolated wavelengths.
For all analyses, Matlab was used (Matlab, 2000). The following Matlab toolboxes were used: PLS_Toolbox (Wise et al., 2005) for PLS analysis; Wavelab (Donoho et al., 1999) for wavelet analysis; DIPlib and DIPimage (Luengo Hendriks et al., 2005) for mathematical morphology; and DACE (Lophaven et al., 2002) for kriging procedures. For display of predicted values for farm fields, values were interpolated using kriging with a zero-order polynomial and Gaussian correlation. Only interpolated values below 0.975 of the maximum error are shown.
Partial Least Squares Analysis
Partial least squares regression was used to relate spectral measurements to yield and quality parameters measured with destructive sampling. Partial least squares was used because it efficiently combines data compression and information extraction (Geladi and Kowalski, 1986). A PLS model relates the information of predictors (in this case, information derived from the three sensors; X block) to quantitative information of the measured characteristics of interest, such as biomass and N content (the Y block, which can also be a univariate Y vector). The PLS regression method effectively performs a canonical decomposition of the X block in such a way that the resulting set of orthogonal factors are both predictive for the Y block and describe as much as possible of the variance in the X block. In this sense, PLS maximizes the covariance of X and Y rather than the correlation between X and Y (as is true for multiple linear regression) or the variance within X (as is true for principal components analysis); PLS is an iterative process where, in every iteration, scores, weights, loadings, and inner coefficients are calculated. Then the X and Y block residuals are calculated and the entire procedure is repeated for the next factor (commonly called a latent variable in PLS analysis).
The error (expressed as the residual sum of squares) continues to decrease as the number of latent variables increases. The more latent variables are chosen, however, the more liable the model is to overfitting (fitting to the noise). The optimal number of latent variables that is used to create the final PLS model is determined by cross validation: a certain number of objects are excluded from the model calibration and predictions are made for the excluded objects; this process of excluding and predicting a subset of k objects is repeated, until all n objects have been excluded (and predicted) once.
In many instances, the subset size k is 1: leave-one-out cross validation. For a large number of objects (n > 40), we have chosen to let the value of k be equal to the square root of the number of objects (n) divided by 5 and rounded downward: k = 
(n/5).
The total set was randomly permuted and for n/k iterations divided into a calibration set of n – k and a validation set of k (in the last step, the remaining objects are used for cross validation, which is often less than k). Prediction accuracy of the model was evaluated with the RMSECV (root mean squared error of cross validation):
 | [1] |
i is the leave-k-out predictions of the variable of interest. The number of latent variables in the model is determined as the lowest number of latent variables for which the RMSECV is not significantly different (
= 0.10) from the overall minimum RMSECV using a randomization t-test (Van der Voet, 1994). The results of the PLS procedure were rescaled (value x standard deviation + mean) before residuals and indicators of model accuracy were calculated.
Model Evaluation
To obtain insight in the prediction capability of the model, we used the percentage of variation accounted for by the cross-validated model (using the leave-k-out predictions) with regard to the total variation in the data set:
 | [2] |
is the average value of the observed variable of interest across the complete data set. The value of Q2 has a strong resemblance to R2 but can become negative if the prediction of the model is inadequate (e.g., in the case of overfitting). The value of Q2 is (as R2) strongly sensitive to the variation within the data set, and so should be considered with caution.
Influence of Date of Measurement on Predictions
Regular measurements of grassland yield and quality using imaging spectroscopy would be impractical if a new calibration involving biomass sampling and chemical analysis were required for each set of circumstances. We simulated the absence of calibration data for a specific date by the following procedure. The PLS models were trained on all data, excluding all nj observations of a combination of experiment and harvest (a group of data j). This group j was then used to evaluate the accuracy of the predictions of the trained data set. This was repeated for all J groups. Afterward, the RMSEP (root mean squared error of prediction) was calculated according to
 | [3] |
ij is the predicted and yij the observed value of object i in group j. The value of Q2 was calculated as described above, except that now normalized data (i.e., before rescaling) were used to minimize the influence of differences in mean values of the groups.
|
RESULTS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSION AND CONCLUSIONSREFERENCES |
|---|
|
|
In this study, relative errors for N, P, and K contents ranged between 6 and 12%. For sugar concentration, relative errors were between 15 and 16% and for crude fiber, neutral detergent fiber, acid detergent fiber, and digestibility, relative errors were between 3 and 5%. The prediction errors for N, P, and crude fiber content were only slightly smaller but much (nearly a factor of two) smaller for K and sugar content than the prediction errors found in the mini-sward study (Schut et al., 2005).
Influence of Measurement Date on Predictions
The influence of the date of measurement on prediction results was evaluated for the grass–clover data set. In Table 4, RMSEP values are presented for the prediction of absolute and relative values. The RMSEP values for predictions of relative values are much lower than for the predictions of the absolute values. This is illustrated by RMSEP values for DM yield, ranging from 237 to 9212 kg DM ha–1 and from 73 to 292 kg DM ha–1 for the prediction of absolute and relative values, respectively. The largest difference was found for the harvest on 27 Oct. 2004. The mean RMSEP values for the relative predictions are more or less comparable with the RMSECV in Table 3.
|
|
|
|
|
DISCUSSION AND CONCLUSIONS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSION AND CONCLUSIONSREFERENCES |
|---|
The potential of imaging spectroscopy for prediction of yields, nutrient concentrations, and nutritive values reported in earlier work (Schut et al., 2005) was confirmed in this study. It was also shown that the concept can be used to visualize patterns of quality aspects in practical fields (Daughtry et al., 2005; Nagler et al., 2003). Various users could profit from this concept of imaging spectroscopy. In research, it enables fast and automated comparison of plots and fields, i.e., for comparison and screening of cultivars in selection trials. It can also be used to measure fields intensively to calibrate remote sensing images for studies at a regional level (Clevers et al., 2005). For practical purposes, a small, simplified version can be used as a tool for extension services. Components of this system can be used to develop specific sensors, standing alone or integrated into farm machinery. Ultimately, this concept can be used for air- or space-borne sensors. It is expected that prediction accuracies will decrease because (i) reflection intensity cannot be used as a measure of canopy height, (ii) soil and plant litter influence the signal, and (iii) atmospheric absorption will play a role (Asner, 1998; Van Leeuwen and Huete, 1996).
There was a large difference between the RMSECV values calculated with the leave-k-out procedure and the RMSEP values calculated with the leave-group-out procedure for prediction of absolute values. This indicates that the covariation within all spectra of a single harvest can correct for the influence of system instability or environmental disturbances (effects of weather and location). In contrast to the prediction of absolute values, the prediction of relative values did result in satisfactory results, with RMSEP values comparable to the RMSECV values. This indicates that the instability of the independent predictions of absolute values may be a matter of finding the proper offset, resulting in a large bias of the prediction results. This is major concern for nearly all indirect methods relying on sensitive equipment. Similar problems of method incompatibility were found by (Givens and Deaville, 1999). They reported that using a near-infrared spectroscopy calibration set by three different consultants resulted in poor predictions. The spectral equation used was very sensitive for differences in sample preparation (Baker and Barnes, 1990). They indicated that this can be resolved by decreasing the sensitivity of the spectral equation to sample preparation or characterizing system changes by recording internal standards frequently.
For laboratory systems, standardized methods are available to correct for drift and differences between systems (Fearn, 2001; Feundale et al., 2002; Wang and Kowalski, 1992). These methods are based on a mathematical description of the differences between instruments or between recording dates that is based on recording a set of reference standards before measurement. System drift may also be corrected with regular recording of standards to characterize system changes, which can then be used to transform spectra to these standards. Within our concept, incorporating system drift in the spectral equation is less convenient as it is unknown how to model this drift; in other words, it is not known how the spectra may change. The influence of equipment maintenance may be large for imaging systems because the projection of spectra on the detector will change with every minor disturbance. Therefore, the influence of decoupling the detector from the spectrograph is large. Probably the best option to deal with this drift or the instability of this system is a duosampling approach, i.e., regular collection of ground truth information (reference samples), simultaneously with collection of reflectance spectra. These can then be used to update the training set, incorporating all changes in the system. Then, temporal extrapolation is severely limited and prediction results would approach the leave-k-out predictions. Other options include more accurate system calibration methods than we have applied up till now.
This study shows that the system is very suitable for measuring large fields. The spatial variation within a field for DM yield and concentration of N, sugar, and crude fiber can be depicted, giving new means for assessing the potential of spatial fine tuning under practical circumstances without large costs for sample analysis.
The occurrence of weeds in the practical fields resulted in unsatisfactory predictions of clover content. This can be resolved by improving discrimination within the images of broad-leaved weeds and white clover, i.e., removal of images (or parts of an image) where the circumference or area of single leaves exceeds a certain threshold.
We conclude that imaging spectroscopy provides accurate measurements of DM yield, nutrient content, and feeding value of standing grass in the field. Cross-validation errors were small for all three data sets. The large errors in the independent predictions of absolute values indicate that a method such as duosampling is needed to achieve the desired accuracy. The number of reference samples required for calibration may eventually be reduced by using a mathematical correction for system changes based on recording standards on a regular basis. Using imaging spectroscopy for ranking purposes and prediction of relative values is less sensitive for system disturbance, and in such situations fewer reference samples would be required.
|
ACKNOWLEDGMENTS |
|---|
|
REFERENCES |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSION AND CONCLUSIONSREFERENCES |
|---|
wavelab/ (accessed 15 May 2006; verified 10 July 2006). Stanford Univ., Stanford, CA.hbn/dace/ (accessed 15 May 2006; verified 10 July 2006). Technical Univ. of Denmark, Kgs. Lyngby.
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| 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 | |||
