Crop Science
HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS
QUICK SEARCH:   [advanced]


     

Published online 31 January 2005
Published in Crop Sci 45:477-485 (2005)
© 2005 Crop 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 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 Web of Science (4)
Right arrow Citing Articles via Google Scholar
Google Scholar
Right arrow Articles by Girma, K.
Right arrow Articles by Stone, M. L.
Right arrow Search for Related Content
PubMed
Right arrow Articles by Girma, K.
Right arrow Articles by Stone, M. L.
Agricola
Right arrow Articles by Girma, K.
Right arrow Articles by Stone, M. L.
Related Collections
Right arrow Weed Management
Right arrow Wheat
Right arrow Remote Sensing

Identification of Optical Spectral Signatures for Detecting Cheat and Ryegrass in Winter Wheat

Kefyalew Girmaa, J. Mosalia, W. R. Raun*, K. W. Freemana, K. L. Martina, J. B. Solieb and M. L. Stoneb

a Dep. of Plant and Soil Sciences, Oklahoma State Univ., Stillwater, OK 74078
b Dep. of Biosystems and Agric. Engineering, Oklahoma State Univ., Stillwater, OK 74078

* Corresponding author (wrr{at}mail.pss.okstate.edu)


   ABSTRACT
TOP
 NOTES
 ABSTRACT
INTRODUCTION
 MATERIALS AND METHODS
 RESULTS
 DISCUSSION
 CONCLUSIONS
 REFERENCES
 
Precision weed management technology has immense potential for treating weed species at a small scale. To this end, however, crop and weeds must be recognized. One approach to this involves identification of reflectance signatures of crops and weeds that differ in the visible and near-infrared (NIR) wavelength region. Reflectance spectra were used for the detection of cheat (Bromus secalinus L.), ryegrass (Lolium multiflorum Lam.), and winter wheat (Triticum aestivum L.) under greenhouse conditions. A total of three experiments (two in December 2002 and one in February 2003) were conducted at the Agronomy Research Station, Stillwater, OK. The three species and two N levels were arranged in a completely randomized design with three replications. Spectral readings were taken at Feekes 3 and 5 winter wheat growth stages with a spectrometer. Two spectral measurements were obtained from each pot. The spectral measurements from the three experiments were combined by the two growth stages because preliminary analysis revealed that date of measurement and N levels were not significant. The spectral readings were measured at 1-nm intervals and averaged into 10-nm bandwidths for the wavelengths from 400 to 865 nm. Data were analyzed using a discriminant analysis procedure. The discriminant function with the band combinations 515/675, 555/675, and 805/815 resulted in the best overall correct classification (94%) of observations at Feekes 3, while for spectral data at Feekes 5 the discriminant function with the band combinations 755 and 855/675 resulted in 66.7% overall correct classification of observations. In several instances, ryegrass was classified as either cheat or winter wheat, while cheat was classified as rye. Cheat was not classified as winter wheat in most instances. This suggests that it is possible to identify cheat in winter wheat using wavelength ratios developed from spectral readings in 10-nm bands between 500 and 860 nm.

Abbreviations: NIR, near-infrared


   INTRODUCTION
TOP
 NOTES
 ABSTRACT
 INTRODUCTION
MATERIALS AND METHODS
 RESULTS
 DISCUSSION
 CONCLUSIONS
 REFERENCES
 
TODAY SITE SPECIFIC application technology or precision farming is becoming an integral part of agriculture. One aspect of site specific application technology involves weed and nutrient management. Precision weed and nutrient management improves weed control efficiency, thereby reducing adverse effects on the environment while maintaining acceptable economic returns (Sawyer, 1994; Brown et al., 1994; Zwiggelaar, 1998; Thompson et al., 1991; Wibawa et al., 1993; Shaw, 2000).

To take full advantage of site specific weed management systems, accurate detection of the location of weeds within crop fields is necessary (Thompson et al., 1991). Cost-effective use of electronically controlled injection sprayers, chemical spot treatment (Stafford and Miller, 1993; Pérez et al., 1997), variable rate sprayers, and chemical mixture delivery systems all require accurate weed distribution records in a field in a form usable by the precision application equipment (Franz et al., 1991).

For weed detection in cultivated crops, two interrelated general approaches have typically been used (Thompson et al., 1990; Guyer et al., 1986, 1993; Woebbecke et al., 1995; Zhang and Chaisattapagon, 1995). The first is to detect certain morphological differences between the crop and weeds, such as leaf shape or plant structure. Franz et al. (1995) used the morphological characteristics of plant species, like hairiness, shininess, and shape, which affect the absorption and reflection bands of plants to detect weeds.

Guyer et al. (1986) studied the feasibility of using leaf shape for plant identification of three crops and five weed species. According to their report, the differences between vegetation and soil reflectance in the NIR region proved to be successful for detecting plants from a soil background. This was true since plant reflectance in the NIR region, which covers the spectra between 720 and 800 nm, is substantially greater than soil (a magnitude of 25% more energy reflection than soil) (Guyer et al., 1993).

Likewise, Woebbecke et al. (1995) used the shape feature analysis for discriminating between monocots and dicots. They tried to identify 10 common weed species in corn (Zea mays L.) and soybean [Glycine max (L.) Merr.] using roundness, aspect, perimeter/thickness, and elongatedness shape features. Aspect enabled them to correctly distinguish 60 to 90% of dicots from the monocots. Their work, however, was restricted to individual plants and not canopy.

Zhang and Chaisattapagon (1995) studied three different approaches to identify weeds in winter wheat fields using machine vision: color, shape, and texture analysis. They used black and white digital images with various color filters under laboratory conditions for color analysis. The red and green filters were effective in detecting reddish stems of some weed species. They also showed that shape parameters such as eccentricity and compactness were effective in distinguishing the broadleaf weed species redroot pigweed (Amaranthus retroflexus L.), wild buckwheat (Polygonum convolvulus L.), and kochia (Kochia scoparia L.) from wheat. From texture analysis that used fineness index, species such as kochia were distinguished from other species with coarse texture. Using a color feature, Humphries and Simonton (1993) identified geranium (Geranium maculatum L.) plant parts, with a success rate of 97, 95, and 93% for leaf, petiole, and stem, respectively.

The second general approach is based on differences in spectral reflectance (Feyaerts et al., 1998; Lass and Thill, 2000). The visible and infrared portion of the electromagnetic spectrum captures the most discriminating information (Richards and Kelly, 1984). Combination of visible and NIR with thermal IR spectra could permit effective use of existing indices, such as greenness (Price, 1987)

Feyaerts et al. (1998) developed a sensor based on reflectance in visible and NIR spectra, which can detect weeds in corn and sugarbeet (Beta vulgaris L.) with a success rate of 80%. Lass and Thill (2000) tried to measure differences in reflectance for different weed species with a hand-held spectroradiometer, recording the full reflectance spectrum at 2-nm increments. However, species identification often was not easy with the remotely sensed aerial multispectral data. Despite the different problems encountered thus far in detecting weeds, some researchers argue that the spectral characteristics of plants are sufficient to differentiate plant species without introducing geometric complexities (Price, 1987; Gutman, 1991).

Hatfield and Pinter (1993) reviewed the potential of remote sensing techniques for crop protection in the field and suggested that one way to distinguish between weeds and crops was by examining the temporal patterns of vegetation indices throughout the growing season. This was also supported by Brown et al. (1994), who reported the potential for distinguishing weeds from agricultural crops on the basis of their relative spectral reflectance characteristics through time. However, they had reservations about identifying individual species and suggested the necessity to group weeds on the basis of some well-defined criteria in real agricultural environments rather than looking for individual weeds. According to Price (1994), unique discrimination of species would be possible using high spectral resolution.

Plant canopy architecture also has a significant effect on canopy reflectance. Moran et al. (1989) found that alfalfa (Medicago sativa L.) has a more erectophile (vertical) leaf architecture when under water stress. The plants also tended to have a lower NIR reflectance when under stress that tended to support the result found in modeling winter wheat (Hatfield and Pinter, 1993). However, in practice, canopy architecture might be more useful in detecting genetically distant species such as cereal crops and broadleaf weeds.

A close investigation of leaf structure gives more insight about reflectance characteristics of vegetation. The upper and lower epidermis of leaves have a protective function with regard to the interaction with electromagnetic radiation, the mesophyll region being the most important part (Jordan, 1969; Lawrence and Ripple, 1998; Richardson and Wiegand, 1977). Accordingly, the range between 400 and 700 nm (visible band) is characterized by very low reflectance due to intense absorption of the incident radiation by pigments in the plant, mainly chlorophyll. All pigments absorb at 430 to 450 nm (blue), and chlorophyll has an additional absorption band at about 650 nm (red). A small reflectance peak also exists at about 550 nm (green). The range between 700 and 1300 nm is characterized by very little absorption and high reflectance. The high reflectance peak in this range is caused by the mesophyll structure, which causes multiple reflection of NIR radiation on the cell walls (Broge, 2003; Gates et al., 1965; Gausman, 1985). The range between 1300 and 2600 nm is characterized by a pronounced minimum reflectance. Wavelengths between 580 and 680 nm (red) and between 725 and 1100 nm (NIR) are high reflectance bands for vegetation.

Despite the importance of detecting multiple species in a mixture of crop and weeds, the task remains challenging. This task is particularly complex when attempting to detect grass weeds in grass crops like wheat. To date, no study has fully achieved a sound method to detect cheat and ryegrass in winter wheat. In our study, we intended to bridge this gap by developing a procedure that could be integrated into sensors to detect cheat and ryegrass in winter wheat.

The objectives of this study were to detect spectral signatures for cheat, ryegrass, and winter wheat and to develop functions that can be encoded into on-the-go variable rate weed control systems for winter wheat.

Meeting the objective will provide new information necessary to identify cheat and ryegrass in winter wheat and later to integrate the information into variable rate technologies developed to manage weeds.


   MATERIALS AND METHODS
TOP
 NOTES
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
RESULTS
 DISCUSSION
 CONCLUSIONS
 REFERENCES
 
Experimental Design and Treatment Structure
Two experiments were conducted at the Agronomy Research Station, Stillwater, OK in December 2002 and one experiment in February 2003. A completely randomized experimental design with three replications was employed.

Cheat, ryegrass, and winter wheat were planted in separate pots (height 20 cm) filled with manure-rich soil with N rates of 0 and 50 kg ha–1 and placed into a greenhouse. Emergence difference of species was accommodated by performing a preliminary study of planting to emergence date of the two weed species with respect to winter wheat in identical growing conditions under the actual experimental conditions. Species population densities after emergence were 250 plants m2 for winter wheat, while the density varied (350–400 plants m2) for cheat and ryegrass to obtain comparable stands when taking measurements. Nitrogen was applied to each pot as urea (46% N). A flat rate of 100 kg ha–1 triple super phosphate (46% P2O5) was applied to each pot. The winter wheat cultivar used for both experiments was Jaggar. The seed for the two weed species was obtained from Weed Science Research Program, Oklahoma State University. Germination tests were performed for both species and were more than 90%. The greenhouse temperature was maintained at 25.5°C with a 12-h day length. Any other species except the target were eliminated on emergence throughout the experiment.

Spectral Readings
Two spectral measurements were made at each of Feekes 3 and 5 winter wheat growth stages for each experiment from each pot using a SD2000 portable fiber optic spectrometer (Ocean Optics Inc., Dunedin, FL) that operates in the visible and NIR region of the spectrum (350–1000 nm) with a resolution of 1 nm (for 50-µm slit) full width half maximum. A 2-m-long glass fiber (Ocean Optics) with a diameter of 200 nm was mounted at 80 cm above the top of the sample in a specially designed lighting system (Fig. 1) and connected to the spectrometer. The lighting system was built as wooden box frame in a rectangular pyramid shape and had two compartments. The bottom compartment (50 by 50 by 20 cm) housed the electrical line and lamps. The top pyramid shape compartment (height 1 m) was used to place samples. The top compartment was painted white inside. The lamps were installed to light upwards along the wall of the pyramid box through circular openings (diameter slightly larger than that of the lamp) at the top of the bottom compartment of the lighting system. Six TRU-AIM-R16 tungsten halogen lamps (Osram Sylvania, Danvers, MA), each 50 W and 12 V with a beam angle of 40° and diameter of 51 mm, were installed. The tungsten halogen lamps were suitable for taking light measurements from samples in the visible and infrared electromagnetic spectra while suppressing the ultraviolet light. The field of view at the sample pot was 10.2 cm in radius. The fiber optic spectrometer was attached to a SAD500 serial A/D (Ocean Optics), which converts analog data to digital data (Fig. 1). The SAD500 serial A/D was connected to a laptop computer that had Ocean Optics OIBase software to record the light intensity for separate wavelengths. Before readings of actual samples were made, reference and dark intensity readings were taken. Reference intensity count was determined by placing a BaCO4–coated metal plate (20 by 30 cm) in the light system, while the dark intensity count was made by blocking the fiber completely with black smooth rubber. Spectral reading of intensity for each sample used the same lighting, temperature, and integration time of 125 ms for each measurement. Spectra were measured with an interval of 1 nm. Reflectance was then calculated as the ratio of reflected light intensity (from the sample plants) to the incident count. Reflectance data were partitioned into 10-nm bandwidths. From the resulting averages, wavelength ratios were determined. The denominator wavelengths selected were 555 nm (the green peak, where the visible spectrum maximum reflectance occurs), 675 (the red minimum where the maximum absorption occurs), and 815 nm (highest point on the NIR plateau). The selection of these wavelengths was based on the reflectance pattern of the three species at 450 to 850 nm (Fig. 2 and Fig. 3) and previous research findings (Vrindts et al., 1999; Borregaard et al., 2000; Hahn and Muir, 1993).


Figure 1
View larger version (27K):
[in this window]
[in a new window]

  		Fig. 1. Spectrometer and lighting system used for collecting spectral data on cheat, ryegrass, and winter wheat.

 

Figure 2
View larger version (18K):
[in this window]
[in a new window]

  		Fig. 2. Reflection patterns of cheat (ch), ryegrass (ry), and wheat (wh) in the spectra between 450 and 865 nm for measurements made at Feekes 3 wheat growth stage.

 

Figure 3
View larger version (19K):
[in this window]
[in a new window]

  		Fig. 3. Reflection patterns of cheat (ch), ryegrass (ry), and wheat (wh) in the spectra between 450 and 865 nm for measurements made at Feekes 5 wheat growth stage.

 
Data Analyses and Classification Methods
Spectral data were analyzed using three discriminant analysis procedures in SAS software (SAS Institute, 2001): Stepwise discriminant analysis (STEPDISC), discriminant analysis (DISCRIM), and canonical discriminant analysis (CANDISC). The STEPDISC procedure was used to identify sets of suitable wavelengths and wavelength ratios. The procedure performs a stepwise discriminant analysis by stepwise selection of variables important in discriminating species. In using the procedure, it was assumed that the spectral data represent a sample from a multivariate normal distribution and that the variance–covariance matrices of variables were homogeneous across species. Decision on selection was based on the respective F to enter and remove reflection data at a specific average wavelength or wavelength ratio. Further analysis of the data was based on the wavelengths and wavelength ratios selected by this procedure.

Once the relevant variables were selected, discriminant functions were developed using DISCRIM procedure. The procedure computed generalized squared distances and various discriminant functions (classification rules) for classifying observations into species. The generalized squared distance between species, otherwise known as Mahalanobis distance, was calculated using the following equation (Mahalanobis, 1936):

Formula 1[1]
where D2ij denotes the Mahalanobis distance between species i and j, cov–1 denotes the inverse covariance matrix, and Av(xi) and Av(xj) denote the mean reflection for species i and j, respectively. The equation assumes the populations from which the groups are derived have common variance. It also takes into account the variances and covariances of the measuring distance. According to Vrindts et al. (1999), the procedure is constrained by multicollinearity. However, the literature (e.g., Klecka, 1980) has shown that the risk is high only when a group is considered against many independent variables.

The linear discriminant function (Fisher, 1936) for a species was given by the formula:

Formula 2[2]

In this formula, the subscript i denotes the respective species; the subscripts 1, 2,..., m denote the m wavelength or wavelength ratio; ci is a constant for the ith species; Cij is the coefficient for the jth wavelength or wavelength ratio in the computation of the classification score for the ith species; xj is the observed value for the jth wavelength or wavelength ratio; and Si is the resultant classification score for a species.

The CANDISC procedure approximates the F statistic, and estimates the probabilities for Mahalanobis distance. It also computes the Hotelling–Lawley trace multivariate statistic (Kleinbaum, 1973) for the wavelength or wavelength ratio under consideration.

One of the two spectral measurements made per pot for each experiment was used for developing the discriminant function, while the other was used to classify new observations. Classification errors were determined by calculating the percentage of wrongly classified spectra for each species.


   RESULTS
TOP
 NOTES
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS
DISCUSSION
 CONCLUSIONS
 REFERENCES
 
In the study the background noise due to differential emergence time and canopy cover of the three species was minimized by synchronizing the emergence times and densities of the three species. Planting to emergence time was 3 d more for cheat and ryegrass compared with winter wheat. Thus, the two species were planted 3 d before winter wheat. Similarly, on average, 50% more plants were placed in a pot for cheat and ryegrass to obtain reasonably uniform soil cover at the time of measurement. The spectral measurements from the three experiments were combined for each growth stage because preliminary analysis of spectral reflectance patterns revealed that date of measurement did not suggest separate analysis. Similarly, the spectral reflectance pattern for the two N levels overlapped, and the two levels were considered replicates in the spectral analysis.

Wavelength Selection
Using STEPDISC procedure for data at both Feekes 3 and 5, individual and wavelength ratios were selected to develop the discriminant functions (Table 1) to separate cheat, ryegrass, and wheat. The wavelengths and ratios obtained were different for measurements made at Feekes 3 and 5. Five categories (functions) of wavelengths were derived for data collected at Feekes 3. The categories include combinations of single average wavelength bands and wavelength ratios with denominators of 555, 675, and 815 nm. All categories except one were highly significant using multivariate statistic (Table 1). Associated r2 values ranged from 0.36 to 0.76. The two functions that resulted in the two largest r2 values contain ratios developed from denominators 675 and 815 nm. For data at Feekes 5, six significant groups of wavelength and wavelength ratios were identified. The r2 values ranged from 0.38 to 0.54. More wavelengths and wavelength ratios were included in most functions for data at Feekes 3 compared with data at Feekes 5 (Table 1).


View this table:
[in this window]
[in a new window]

  		Table 1. Wavelengths selected using STEPDISC procedure to develop discriminant functions that identify cheat, ryegrass, and winter wheat.

 
Data at Feekes 3
Reflection patterns of the three species in the spectra range 450 to 850 nm are presented in Fig. 2. Discriminant function coefficients were determined and presented in Table 2. The larger the absolute value of the coefficient, the better the discriminating power. In general, most of the coefficients in the linear discriminant functions in data at Feekes 3 had good discriminating power of a species, as absolute values of the coefficients were much greater than zero. However, the power of discrimination varied for each function, resulting in a difference in the ability of discrimination of each coefficient for each species. For example, in Table 2 for Function 1, the coefficients c1 for cheat, ryegrass, and wheat were 3756, 3239, and 4625, respectively. Since the c1 coefficient was associated with the wavelength 725 nm in discriminant Function 1-A, winter wheat had the largest coefficient and was more discriminable than the other two species at this wavelength band. Likewise, c2 coefficients, which correspond to the 735-nm wavelength band in this function, indicated that winter wheat had the largest absolute value coefficient and thus the highest discrimination. On the other hand, c3 coefficients corresponding to 745 nm in the same function revealed that cheat and winter wheat were highly discriminable from ryegrass but were similar in magnitude to each other. The power of the function lies in the combined effect of all the wavelengths in the function.


View this table:
[in this window]
[in a new window]

  		Table 2. Parameter coefficients for linear discriminant functions and their wavelengths as selected using STEPDISC procedure for data at Feekes 3 winter wheat growth stage.

 
The squared distance between species was significant between winter wheat and the two weed species for all functions except Function 1-D (Table 3). The greatest discrimination among the three species was due to the Function 1-E with wavelength ratios 515/675, 555/675, and 805/815 nm. In fact, this function also resulted in the highest r2 (Table 1).


View this table:
[in this window]
[in a new window]

  		Table 3. Generalized squared distance between species for spectral measurement data at Feekes 3 winter wheat growth stage.

 
The misclassification of observations from one species into another is given in Table 4. Here, Functions 1-C and 1-E correctly classified all observations to the respective species except winter wheat (16.7% classified as ryegrass). Function 1-B classified all observations from wheat as wheat, and none of the observations from the other two species were classified as wheat. All functions classified at least 83% of observations from winter wheat into winter wheat. Few cheat or ryegrass plants were classified as winter wheat and visa versa in some of the functions. On the other hand, Functions 1-A and 1-D misclassified 33 to 50% of the cheat and ryegrass samples.


View this table:
[in this window]
[in a new window]

  		Table 4. Percentage of observations classified from species to species for spectral measurement data at Feekes 3 winter wheat growth stage. There were 18 observations per species.{dagger}

 
Overall error rates or correct classification rates were determined for all functions assessed (Table 4). The overall error rate was 6% for Functions 1-C and 1-E. Function 1-B (wavelength ratios with denominator 555 nm) attained a low (10%) overall error rate for all species, although it achieved comparably better results for winter wheat (Table 4).

Data at Feekes 5
Overall reflection patterns for data at this growth stage are presented in Fig. 3 for the three species. Feekes 5 results were somewhat different than those of Feekes 3. The coefficient c1 in each linear function was larger than either weed species for winter wheat in three of the linear functions (Table 5). In each function, the wavelength or wavelength ratios associated with c1 enabled the discrimination of winter wheat from the weed species. Similarly, the absolute value of c2 was large for winter wheat in two of the linear functions (Table 5).


View this table:
[in this window]
[in a new window]

  		Table 5. Parameter coefficients for linear discriminant functions and their wavelengths as selected using STEPDISC procedure for data at Feekes 5 winter wheat growth stage.

 
The squared distances between species at Feekes 5 were low and less consistent than those for Feekes 3 across the wavelengths selected (Table 6). Generalized squared distances between winter wheat and ryegrass were large and highly significant for most functions. Likewise, the distance was significant for most functions between winter wheat and cheat. However, cheat and ryegrass were not significantly different in any of the functions for data at Feekes 5. The magnitude of the squared distance difference was largest between winter wheat and the weed species for two functions (Function 2-E with 755 nm, 855/675 nm; Function 2-F with 745 and 755 nm, 855/675 and 685/815 nm).


View this table:
[in this window]
[in a new window]

  		Table 6. Generalized squared distance between the three species for spectral measurement data at Feekes 5 winter wheat growth stage.

 
Misclassification of observations was very high for data at Feekes 5 (Table 7). Two functions (Function 2-E with 755 nm, 855/675 nm; Function 2-F with 745 and 755 nm, 855/675 and 685/815 nm) classified all observations from winter wheat as winter wheat, while 66.7% of observations from cheat were classified as cheat by these functions. Most functions were not effective in classifying observations from ryegrass as ryegrass, except function 2-B developed using wavelength ratio 745/555 nm (83.3%). In most functions, the highest misclassification of species was for cheat classified as rye and rye classified as cheat (Table 7). The order of correct classification for this data set was: winter wheat > cheat > rye across all functions evaluated (Table 7).


View this table:
[in this window]
[in a new window]

  		Table 7. Percentage of observations classified from species to species for spectral measurement data at Feekes 5 winter wheat growth stage. There were 18 observations per species.

 
Error rates for data at Feekes 5 were large, as observations in most functions were misclassified at this later growth stage. The highest overall correct classification was 66.7%, which resulted from Functions 2-E and 2-F (Table 7).


   DISCUSSION
TOP
 NOTES
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS
 DISCUSSION
CONCLUSIONS
 REFERENCES
 
The stepwise discriminant function analysis showed that wavelength bands in the visible and NIR regions of the spectrum were required to discriminate the three species. Several researchers also reported similar results in studies of discrimination of different crop and weed species (Smith and Blackshaw, 2003; Feyaerts et al., 1998; Vrindts and De Baerdemaeker, 1996; Borregaard et al., 2000). The spectral signatures identified from this study strongly suggest that single averages between 720 and 750 nm are suitable for separating the three species. At Feekes 3, ratios with numerators in the green, red, and NIR region and the three denominators were suitable for separating the species. At Feekes 5, however, the numerator wavelengths were in the red and NIR range.

Using the discriminant functions and generalized square distances, the best functions to discriminate the three species were identified for the data at both stages. At Feekes 3, with high r2, significant multivariate statistic, low misclassification of observations, and low error rate, Function 1-C with wavelength ratios 515/675, 545/675, and 555/675 nm and Function 1-E with 515/675, 555/675, and 805/815 nm were the best functions. However, since Function 1-C had slightly lower generalized square distance between species, lower r2, and it misclassified some winter wheat measurements, Function 1-E was the preferred function. Using this function, all observations from cheat and ryegrass were correctly classified (Fig. 4) . Some researchers have successfully discriminated weed species from several crops in cases where the weed species were morphologically very distinct from the crop (Smith and Blackshaw, 2003; Feyaerts et al., 1998). For these data, most functions resulted in excellent discrimination of winter wheat and the two weed species with few exceptions. At the early stage of growth, chlorophyll is not well developed, and other pigments such as carotenoids are found in relatively high abundance. This subsequently caused higher reflectance, in both red and NIR spectral regions, which was different for the three species evaluated. The composition of the best discriminant function for data at Feekes 3 strongly suggest that the green peak, red, and NIR portion of the spectrum are good enough to discriminate cheat, ryegrass, and winter wheat.


Figure 4
View larger version (28K):
[in this window]
[in a new window]

  		Fig. 4. Discrimination of cheat (ch), ryegrass (ry), and wheat (wh) at Feekes 3 wheat growth stage with the discriminant function containing wavelength ratios of 515/675, 555/675, and 805/815 nm. There were 18 observations per species. Note that 17% of observations from wheat were classified into ryegrass.

 
For data at Feekes 5, the performance of most of the functions was poor when discrimination of all species was attempted. However, Functions 2-E (with wavelengths of 755 nm and wavelength ratio of 855/675 nm) and 2-F (with wavelengths of 745 and 755 nm and ratios 855/675 and 685/815 nm) resulted in good discrimination (100%) of winter wheat, but not the weed species. Of the two functions, the first is preferred because it had fewer variables (wavelengths) and adds simplicity. Inclusion of more wavelength bands in a discriminant function would enable more discrimination (Vrindts and De Baerdemaeker, 1997), but, at the same level of precision, the simpler function would help in the use of selected wavelengths. The discrimination of the three species at Feekes 5 with the best selected function (2-E) is presented in Fig. 5 .


Figure 5
View larger version (10K):
[in this window]
[in a new window]

  		Fig. 5. Discrimination of cheat (ch), ryegrass (ry), and wheat (wh) at Feekes 5 wheat growth stage with the discriminant function for the 755-nm wavelength and wavelength ratio of 855/675 nm. There were 18 observations per species.

 
Correct classification and percentage of observations classified from a species to another species was acceptable for data at Feekes 3 but not for data at Feekes 5. Error rates were larger between the two weed species than between the weed species and winter wheat. According to past research (Vrindts et al., 1999, 2000), this is not a significant concern, since the primary objective of the research was to select wavelengths that differentiated the crop from weeds. However, lower error rates are desirable in laboratory tests because of better control of the factors. Total error of only 3% in crop vs. weed classification using a small number of simple ratios of 10-nm wavelength bands in a discriminant function has been reported (Vrindts et al., 1999).

The wavelength and wavelength ratios obtained for each data set were different. This suggests that as plants continue to grow from Feekes 3 to 5 and increase in height and canopy coverage, exposure of the sample and the subsequently measured reflectance pattern are affected (Noble et al., 2002; Wang et al., 2001). Typically, since chlorophyll concentration drastically increases with increase in growth, the reflection pattern in the green region of spectrum decreases. Thus, measurements vary when compared with earlier measurements made on the same plants. At the early growth stage, winter wheat and cheat had distinctive appearances in this study and previous observations (Franz et al., 1991; Cooper, 1964; Jackson and Pinter, 1986). This difference may contribute to the powerful discrimination of the two species by selected functions for data at Feekes 3. On the other hand, for data at Feekes 5, an increase in canopy closure coupled with a decrease in pubescence of leaves emerging at a later growth stage decreased discriminability. The significance of canopy cover in spectral weed detection was discussed in detail (Andreason et al., 1997). This has important consequences when using selected wavelengths to identify wheat from weeds. The lack of consistent results obtained across growth stages requires accurately defining the appropriate growth stage of the species where discrimination and treatment are optimal.


   CONCLUSIONS
TOP
 NOTES
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS
 DISCUSSION
 CONCLUSIONS
REFERENCES
 
The results obtained here showed that spectral measurements differed with growth stage of the plants. A thorough evaluation of change in reflectance pattern for the species under consideration is required to accurately classify crops and weeds. The best overall classification obtained for data at Feekes 3 (94%) and Feekes 5 (66.7%) was attributed to the discriminant functions with 515/675, 555/675, and 805/815 nm and with 755 nm and a ratio of 855/675 nm, respectively. For data at Feekes 3, ryegrass was classified as cheat and visa versa. Cheat was not classified as winter wheat in most instances, except Functions 1-A and 1-D, whereas ryegrass was misclassified as winter wheat only in Function 1-D. For data at Feekes 5, although the magnitude was small, some observations from cheat were classified as winter wheat for most functions. In several instances, ryegrass was classified as either cheat or winter wheat, while cheat was classified as rye.

Overall, cheat was rarely classified as winter wheat. This suggests that it is possible to identify cheat in winter wheat using wavelength ratios developed from spectral readings in the 500 and 860 nm bands. At early growth stages, winter wheat and cheat have distinctive appearances. This difference might have contributed to the powerful discrimination of the two species by selected functions for data at Feekes 3. On the other hand, for data at Feekes 5, an increase in canopy closure coupled with a decrease in pubescence of leaves that emerge at later growth stages decreased the power of discrimination. The discrimination results reported were based on pure stands of each species. The information provided here will guide researchers who would like to advance the discrimination of the two weed species in winter wheat to mixed populations of the species and field studies of precision weed control.


   NOTES
TOP
 NOTES
ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS
 DISCUSSION
 CONCLUSIONS
 REFERENCES
 
Contribution from the Oklahoma Agric. Experiment Station.

Received for publication April 13, 2004.


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




This article has been cited by other articles:


Home page
 jashsHome page
M. T. Gomez-Casero, F. Lopez-Granados, J. M. Pena-Barragan, M. Jurado-Exposito, L. Garcia-Torres, and R. Fernandez-Escobar
Assessing Nitrogen and Potassium Deficiencies in Olive Orchards through Discriminant Analysis of Hyperspectral Data
J. Amer. Soc. Hort. Sci., September 1, 2007; 132(5): 611 - 618.
[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 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 Web of Science (4)
Right arrow Citing Articles via Google Scholar
Google Scholar
Right arrow Articles by Girma, K.
Right arrow Articles by Stone, M. L.
Right arrow Search for Related Content
PubMed
Right arrow Articles by Girma, K.
Right arrow Articles by Stone, M. L.
Agricola
Right arrow Articles by Girma, K.
Right arrow Articles by Stone, M. L.
Related Collections
Right arrow Weed Management
Right arrow Wheat
Right arrow Remote Sensing

HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS
The SCI Journals Agronomy Journal 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