Glyphosate is widely used in row crop weed control programs of glyphosate-resistant (GR) crops. With the accumulation of glyphosate use, several weeds have evolved resistance to glyphosate. In order to control GR weeds for profitable crop production, it is critical to first identify them in crop fields. Conventional method for identifying GR weeds is destructive, tedious and labor-intensive. This study developed hyperspectral imaging for rapid sensing of Italian ryegrass ( <i> Lolium perenne </i> ssp. <i> multiflorum </i> ) plants to determine if each plant is GR or glyphosate-susceptible (GS). In image analysis, a set of sensitive spectral bands was determined using a forward selection algorithm by optimizing the area under the receiver operating characteristic between GR and GS plants. Then, the dimensionality of selected bands was reduced using linear discriminant analysis. At the end the maximum likelihood classification was conducted for plant sample differentiation of GR Italian ryegrass from GS ones. The results indicated that the overall classification accuracy is between 75% and 80%. Although the accuracy is lower than the classification of Palmer amaranth ( <i> Amaranthus palmeri </i> S. Wats.) in our previous study, this study provides a rapid, non-destructive approach to differentiate between GR and GS Italian ryegrass for improved site-specific weed management.
Glyphosate [N-(phosphonomethyl)-glycine] is the most widely used herbicide with its increasing use in transgenic, glyphosate-resistant (GR) cropping systems [
Italian ryegrass (Lolium perenne L. ssp. Multiflorum (Lam.) Husnot) is one of the GR weeds that was first reported in Chile in 2001 in fruit and orchard farming systems, followed by Brazil, Argentina, Spain, Italy, Switzerland, New Zealand and Japan in different agronomic situations [
The conventional method for differentiating GR and GS plants is destructive and assesses the plant physiological and biochemical changes before and after glyphosate treatment. During such an assessment, glyphosate is treated on the whole plant, single leaves or leaf disks [
With glyphosate treatment, plants will indicate changes in physiological, biochemical and cell structures [
In order to develop a rapid, non-destructive method for differentiating GR and GS weeds, Reddy et al. [
With our success in previous research with GR and GS Palmer amaranth differentiation, we intend to further study and test if the same approach could be transferred to other GR weed species. Therefore, the objective of this research was to develop a hyperspectral imaging method for rapid, non-destructive measurement, processing and analysis of spectral data for differentiation between GR and GS Italian ryegrass.
For this study, 226 Italian ryegrass plants were grown from seeds, transplanted (2 weeks after emergence) in 6 cm by 6 cm by 6 cm black plastic pots containing a commercial potting mix in a greenhouse at the United States Department of Agriculture, Agricultural Research Service, Crop Production Systems Research Unit at Stoneville, Mississippi. The 226 plants consisted of 119 GR and 107 GS Italian ryegrass.
The plants were imaged using a Resonon Pika II VNIR (Visible + Near Infrared) hyperspectral imaging system (Resonon, Bozeman, MT). The Pika II camera was mounted approximately 1 m above the sample stage, which (when combined with the lens magnification) resulted in sub-millimeter pixel sizes. The camera has 2.1 nm spectral range and 12 bit dynamic range. The spectral range of the camera is 400 - 900 nm, with 240 narrow wavelength bands evenly spaced in the range. The high spatial resolution of the imaging settings ensured that a very large number of pure plant pixels were present in the images. During imaging, the sample stage was covered by black felt cloth, which has almost no spectral reflection, to provide good contrast between the plants and the background in the acquired images. The good contrast simplified segmentation of the plant pixels from the non-plant pixels in the images as well.
The plants were first imaged approximately 3 weeks after transplanting, and then 4 and 5 weeks after transplanting. At the first imaging, each plant had an average of about 7 tillers, and at the second imaging, the number of tillers were 14. At the third, the number of tillers was over 20 on average. The SpectrononPro software (Resonon, Bozeman, MT) developed for the Pika II hyperspectral imaging system combines the white reference and dark current measurements with the images of the plants to calculate spectral reflectance for each pixel of the images.
The calculated reflectance values were normalized to remove artifacts caused by
height variation in the plants. The normalization was accomplished by dividing the reflectance of each pixel at a wavelength by its vector magnitude across all the wavelengths. The equation for the normalization is as follows:
R i j k = r i j k ∑ k = 1 n r i j k 2 (1)
where the reflectance r i j k is the original reflectance of the pixel at (i, j) for the kth wavelength in the total n wavelengths.
After reflectance normalization, images were segmented to separate the pixels of the plants from the pixels of the background. The segmentation was conducted by assuming that the plant was the only green object in each image. To segment the green object out of each image, a three-dimensional vector was formed using the reflectance values at 450 nm (Blue), 550 nm (Green), and 650 nm (Red) as:
v i j = 〈 r i j , 450 nm , r i j , 550 nm , r i j , 650 nm 〉 (2)
The vector was then normalized similarly as:
V i j = v i j r i j , 450 nm 2 + r i j , 550 nm 2 + r i j , 650 nm 2 (3)
The dot product between the normalized vector V i j and the unit vector 〈 0 , 1 , 0 〉 was, then, thresholded to identify green pixels in the images. Through trial and error, the pixels with values over the threshold of 0.8 were found to be appropriate for identifying plant pixels. In using this technique, when the vector v i j has a small magnitude, there is risk to label dark pixels as green in addition to labeling plant pixels. Therefore, a correct program was coded and implemented to prevent dark pixels from being labeled as green pixels by labeling the pixels with the magnitude of v i j less than 0.2 as background.
Finally, the normalized reflectance R i j k values for the pixels labeled as plant were averaged to obtain the mean reflectance for the plant in the image:
R ¯ k = 1 Number of Plant Pixels ∑ i * j * R i j k ( k = 1 , 2 , ⋯ , n ) (4)
where i* and j* are the locations of plant pixels in the image.
Data analysis was conducted on the mean reflectance for the plants as shown in Equation (4). The data set was randomly divided into training (89 GR and 80 GS) and testing (30 GR and 27 GS) subsets. A classification model was calibrated with the training data set, and then tested using the testing data set. The model was calibrated using a forward selection algorithm to select characteristic wavelengths. Then, Fisher’s Linear Discriminant Analysis (FLDA) was used to reduce the dimensionality of the set of selected wavelength bands. At last, the maximum likelihood classifier was used for classification between GR and GS Italian ryegrass.
Due to the large number of spectral bands and the limited number of training samples, it was necessary to use only some of the available bands to avoid diminishing results. Diminishing results are caused when there are not enough training samples to estimate joint probability distributions in high dimensional feature spaces [
A problem using ROC is that it can only be computed on a one dimensional feature vector. This means that when the dimensionality of the selected bands set is greater than one, all the information contained within those selected bands has to be summarized by a one dimensional vector (that is, reduce the dimensionality to one). One way to reduce the dimensionality is to compute a weighted sum of all the bands. FLDA does this by computing the weights that optimize the ratio of between class scatter divided by within class scatter [
After the spectral bands were selected and dimensionality reduced, the training data set was used to train the maximum likelihood classifier [
The performance of the classifier was determined in two different ways. The first was to compute a calibration accuracy using leave-one-out validation on the training data [
The results of both performance metrics were organized into confusion matrices. To enhance the performance evaluation, a permutation study was conducted by repeating the classifications 100 times and computing the mean and standard deviation of the confusion matrices. Since the training and testing data sets were chosen randomly, each permutation had different training and testing data sets to determine the overall performance of the process.
As mentioned above, the Italian ryegrass plants were imaged 3, 4 and 5 weeks after transplanting. With each time of data 15 most sensitive wavelength bands were selected (
WAT | Wavelength Bands (nm) | ||||||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
3 | 402.7 | 592.0 | 596.2 | 619.3 | 621.4 | 627.7 | 631.9 | 640.4 | 663.5 | 669.8 | 741.3 | 753.9 | 764.4 | 880.1 | 892.7 |
4 | 417.4 | 507.9 | 514.2 | 577.3 | 589.9 | 608.8 | 657.2 | 739.2 | 785.5 | 793.9 | 806.5 | 819.1 | 829.6 | 835.9 | 854.9 |
5 | 396.4 | 413.2 | 423.7 | 425.8 | 440.6 | 470.0 | 669.8 | 686.6 | 735.0 | 739.2 | 741.3 | 747.6 | 762.3 | 768.6 | 791.8 |
Class Label Actual Class | GR | GS | Accuracy |
---|---|---|---|
GR | 23.1 (2.2) | 6.9 (2.2) | 76.9% |
GS | 7.2 (2.6) | 19.7 (2.6) | 73.2% |
Overall Accuracy | 75.1% |
Class Label Actual Class | GR | GS | Accuracy |
---|---|---|---|
GR | 24.3 (2.3) | 5.7 (2.3) | 81.0% |
GS | 5.6 (2.1) | 21.4 (2.1) | 79.3% |
Overall Accuracy | 80.2% |
Class Label Actual Class | GR | GS | Accuracy |
---|---|---|---|
GR | 19.8 (2.4) | 6.2 (2.4) | 76.2% |
GS | 5.9 (2.3) | 17.1 (2.3) | 74.4% |
Overall Accuracy | 75.4% |
Our previous research on Palmer amaranth using similar data analysis techniques resulted in over 90% classification accuracy to differentiate GR and GS plants [
with much narrower leaves. Although the performance for detection of Palmer amaranth and Italian ryegrass is different using hyperspectral imaging, the results of this study are still encouraging in illustrating that the approach developed for differentiating GR and GS Palmer amaranth could be transferred to Italian ryegrass with fairly high classification accuracy. In practice, a number of GR and GS weeds may coexist in a crop field, including Palmer amaranth and Italian ryegrass. In this case, it is possible to use hyperspectral imaging with similar data analysis techniques to detect GR Palmer amaranth and Italian ryegrass simultaneously with high success rate. For the time being, we are working on other GR weed species, such as Johnsongrass (Sorghum halepense), to see if the techniques with GR Palmer amaranth and Italian ryegrass could be further transferred. Furthermore, the analysis techniques used in this study might be limited to handle data complexity and nonlinearity. In this way more advanced methods will be evaluated and developed, such as machine learning and nonlinear analysis, when more training data are available with more experiments to produce better improved performance for detection of GR weeds for crop production management.
The difference in GR and GS classifications between images collected at three, four and five weeks after transplanting may indicate the difference in leaf epicuticular wax content between the different plant growth periods. A further research is needed to test the hypothesis. In addition plant imaging could be conducted at more time periods such as one or two weeks after transplanting at earlier growth stages and six or seven weeks after transplanting at later growth period for a broader comparison. Perhaps, less number of plants could be imaged at each growth period to balance the equipment and labor resources spent on imaging at more growth periods.
One drawback with the analysis used for the Italian ryegrass in this study was that the sensitive bands in each permutation varied because the training set was different with each permutation. In order to determine what the common sensitive bands are, a meta analysis is needed with more sample plants to be imaged and analyzed. With more data in such an approach, more accurate probability distributions of GR and GS plants could be estimated to more reliably determine the wavelength bands that are sensitive to be useful for analysis.
This study developed the methods of hyperspectral image analysis and data classification for GR and GS Italian ryegrass differentiation by adopting the methods we developed previously for GR and GS Palmer amaranth differentiation. The results indicated that the overall classification accuracy is between 75% and 80%. Although the accuracy varies from the classification of Palmer amaranth in our previous study, this study proves that the success of the methods on one GR species can be transferred to another species and provides a rapid, non-destructive approach to differentiate between GR and GS Italian ryegrass for improved site-specific weed management.
Mention of trade names or commercial products in this publication is solely for the purpose of providing specific information and does not imply recommendation or endorsement by the US Department of Agriculture.
The authors thank Jason Corbitt, Efren Ford, and Earl Gordon for technical assistance with growing the Italian ryegrass and assisting with data collection efforts.
Huang, Y., Lee, M.A., Nandula, V.K. and Reddy, K.N. (2018) Hyperspectral Imaging for Differentiating Glyphosate-Resistant and Glyphosate-Susceptible Italian Ryegrass. American Journal of Plant Sciences, 9, 1467-1477. https://doi.org/10.4236/ajps.2018.97107