Spectral Characteristics of Citrus Black symptoms, but the most distinguish- ing one is called hard spot (Fig. 1), Spot Disease which is a circular lesion (with 3– 10 mm diameter) with gray necrotic 1 1,4 2 fatal tissue at the center embraced by Alireza Pourreza , Won Suk Lee , Mark A. Ritenour , a black margin. Hard spot becomes and Pamela Roberts3 apparent at the time of fruit coloring just before harvesting (Dewdney et al., 2014). Fruit that have CBS ADDITIONAL INDEX WORDS. CBS, fungus, spectrometry, diagnosis, machine lesions are nonvaluable for the fresh learning, band selection, classification markets. Premature fruit drop is also SUMMARY. Citrus black spot (CBS) is a fungal disease caused by Phyllosticta a consequence of CBS disease in citricarpa (synonym Guignardia citricarpa). CBS causes fruit lesions and signifi- severe conditions. There is a long cant yield loss in all citrus (Citrus) species. The most distinguishing CBS symptom is latent stage in the disease cycle in called hard spot, which is a circular lesion with gray tissue at the center surrounded which the CBS symptoms are not by a black margin. The spectral characteristic of CBS lesions was investigated and apparent for several months after in- compared with the spectral signature of healthy fruit tissue to determine the best fection. Symptoms usually appear at distinguishing wave band. Healthy and CBS-affected samples presented similar the time of fruit ripening (Dewdney reflectance below 500 nm and above 900 nm. However, healthy samples reflected more light between 500 and 900 nm, especially within the visible band. Also, et al., 2014). spectral reflectance of the same symptomatic lesion was acquired six times over To reduce the spread of the a 2-month period to determine the variation of symptom’s spectral signatures over disease, CBS must be efficiently con- time after being harvested. A two-sample t test was employed to compare each pair trolled in the grove. In addition, of consecutive repetitions. The results showed that the spectral signature of the CBS manual detection of CBS symptoms lesion did not change significantly over 2 months. The wavelengths between 587 at the packaging process is extremely and 589 nm were identified as the distinguishing band to develop a monochrome difficult because they may be con- vision–based sensor for CBS diagnosis. A support vector machine (SVM) classifier fused with blemishes caused by other was trained using the spectral reflectance data at the selected bands to identify CBS- disorders and the process is time affected samples in each repetition. The overall CBS detection accuracies varied consuming. Therefore, a rapid and between 93.3% and 94.6%. accurate CBS identification technique can expedite the quality control lorida citrus industry has been 2007). Again after a few years, process and help growers for bet- threatened by several, previ- HLB disease was found in all citrus ter disease management. Computer Fously exotic citrus diseases dur- producing counties in Florida vision–based sensors have been widely ing the past few years. Citrus canker and made a huge impact on the used for plant disease identification (caused by Xanthomonas axonopodis $9 billion citrus industry in Florida (Pourreza et al., 2015; Qin et al., pv. citri), a bacterial disease, was (Pourreza et al., 2014). In Mar. 2012; Sankaran et al., 2010). How- discovered in 1995 near the Miami 2010, CBS was diagnosed on fruit ever, using vision sensors for CBS International Airport (Gottwald in some groves near Immokalee, FL identification has not been thor- et al., 2002) and despite all the (Schubert et al., 2013). CBS is oughly investigated. Bulanon et al. efforts to eradicate the disease, it a fungal disease caused by P. citri- (2013) analyzed hyperspectral images became widespread in Florida carpa (Er et al., 2013). The disease of CBS symptoms in the range of mainly because of the 2004–05 was found in Taiwan, South Africa, 480–950 nm (spectral resolution of hurricanes (Stover et al., 2014). and China in 1919, 1920, and 2.8 nm) with the objective of deter- In 2005, another destructive bac- 1936, respectively (Kotze, 1981; mining the potential bands to develop terial disease called Huanglongb- Wang et al., 2012). Later in the a multispectral imaging sensor. They ing (HLB) or citrus greening 1980s, CBS was officially found in defined four wavelengths including caused by Candidatus Liberibacter coastal humid regions of Australia 781, 713, 629, and 493 nm as the asiaticus and vectored by the asian and caused huge yield losses for selected bands for a multispectral im- citrus psyllid (Diaphorina citri) several years (Kotze, 1981). age acquisition system. They achieved was discovered in Florida City, CBS causes fruit lesions and sub- the overall accuracy of 96% using the south of Miami (Gottwald et al., stantial yield loss in all citrus species. four selected wavelengths and nor- Sweet orange (Citrus sinensis)va- malized difference vegetation index We would like to thank the Florida Specialty Crop rieties such as the ‘Valencia’ are band ratio of 781 nm [near-infrared Block Grant Program for funding this research. We extremely susceptible to this dis- (NIR)] and 713 nm (red) as the input would also like to express our sincere appreciation to Ce Yang, Chuanqi Xie, Daeun Choi, and Chuanyuan ease. CBS causes a wide range of features. Zhao for their assistance in this study. 1Department of Agricultural and Biological Engineer- ing, University of Florida, Gainesville, FL 32611 Units 2Indian River Research and Education Center, Uni- To convert U.S. to SI, To convert SI to U.S., versity of Florida, Fort Pierce, FL 34945 multiply by U.S. unit SI unit multiply by 3Southwest Florida Research & Education Center, 25.4 inch(es) mm 0.0394 University of Florida, Immokalee, FL 34142 1 micron(s) mm1 4Corresponding author. E-mail: wslee@ufl.edu. ( F – 32) O 1.8 F C( C · 1.8) + 32
254 • June 2016 26(3) significant difference between the spectral signatures of different rep- etitions. The averages of spectral reflec- tance and their first derivatives were compared between every two consecu- tive repetitions and also between the first and last repetitions for each class. All data analyses were conducted using MATLAB (version R2011a; MathWorks, Natick, MA). Fig. 1. Images of two citrus black spot lesions (CBS positive) and two CBS- BAND SELECTION. To determine negative spots on peels of ‘Valencia’ sweet orange fruit. the most important wavelengths in CBS diagnosis, five feature ranking methods including t test, Kullback– The main goal of this research (QR600-7-SR-125F; Ocean Optics) Leibler distance, Chernoff bound, was to investigate the spectral signa- with a core diameter of 600 mm was receiver operating characteristic, and tures of CBS hard spot lesions using used to measure the spectral reflec- Wilcoxon tests were employed (Liu high spectral resolution data and de- tance of the target spots. A reflection and Motoda, 1998; Theodoridis and termine the best wavelength for CBS probe holder (RPH-1; Ocean Optics) Koutroumbas, 2009). The results of identification. The specific objectives was used to keep the same probe each method included a vector of were to: 1) investigate the progress of positioning to all the target spots coefficients corresponding to the fea- the lesion development on fruit over at 45 in all measurements. The tures (wavelengths). These coefficients time after harvest, 2) select the im- light source (LS-1 Tungsten Halogen represent the level of relevance of each portant wave bands for designing Light Source; Ocean Optics) of the wavelength for the classification pro- a customized and image acquisition spectrometer was turned on 45 min cess. The coefficient ranges for each system to CBS symptoms, and 3) before measurement to ensure it method varied since different factors evaluate the classification accuracy reached its stable status. The optical were considered for different feature using the selected band as the input reference standard was collected be- ranking methods. Therefore, the co- feature. fore each measurement using a certi- efficients in each method were nor- fied 99% white reflectance standard malized by the maximum obtained Materials and methods (SRS-99-020; Labsphere, North Sut- coefficient that belonged to the top DATA COLLECTION. Healthy and ton, NH). wavelength ranked with the corre- CBS symptomatic citrus fruit of The measurement was repeated sponding method. To select the most ‘Valencia’ sweet orange were col- six times on 11 Apr., 18 Apr., 2 May, relevant bands, a combination of the lected from a citrus grove near Immo- 15 May, 29 May, and 13 June 2014. entire feature ranking results was used kalee, FL, in Apr. 2014. To conduct The CBS-negative samples were added to appoint the best set of nominated the spectral measurement, the citrus to the dataset from the second repeti- wavelengths, which were voted by all samples were transferred to the post- tion (18 Apr.). At the forth repetition methods. For this purpose, all the harvest laboratory in the Indian River of spectral measurement, a few sam- wavelengths that had the coefficients Research and Education Center, Uni- ples began to decay. These decayed greater than 0.99 in all methods were versity of Florida, Fort Pierce. A total samples were removed before the fifth selected. of 134 citrus fruit including 104 and sixth repetitions of spectral mea- CLASSIFICATION. The selected CBS-positive and 30 CBS-negative surement. Color images of the citrus bands were used to train a SVM bi- samples were selected randomly for samples and the selected spots were nary classifier with two classes of CBS spectral measurement. Two or three taken as a reference of the sample positive and CBS negative (Bishop, CBS symptomatic hard spot lesions status at the time of the spectral mea- 2006). Classification was conducted were selected and marked randomly surement. The citrus samples were for every repetition as well as the on each of the CBS-positive samples kept at air temperature of 1 Cand entire dataset. A 10-fold cross valida- to create the CBS-positive class. Also, humidity of 70% during the experi- tion method was used in the classifi- two CBS asymptomatic spots on each ment period. The spectral range of the cation process in which the dataset of the CBS-positive samples and spectrometer was between 340 and was randomly divided into 10 folds; two or three spots on CBS-negative 1030 nm; however, the measured re- 9 folds were used for training while fruit were selected to create the CBS- flectance below 400 nm was very the other fold was used for validation. negative class. noisy. Therefore, the analysis was per- The classification was repeated 10 S PECTRAL REFLECTANCE formed within the entire visible and times until all the samples were in- MEASUREMENT. A portable spectrom- part of NIR bands. In total, 1914 cluded at least once in the validation eter (USB2000+; Ocean Optics, wavelengths between 400 and 1030 set. Dunedin, FL) with spectral resolu- nm were available for the analysis tion range of 0.28 to 0.38 nm and process. Results an optical resolution of 0.1 nm was C OMPARISON BETWEEN DATASET PREPARATION. In total, used to measure the spectral reflec- REPETITIONS. Two-sample t test 3102 spectral reflectances were mea- tance of the target spots on the citrus (Snedecor and Cochran 1989) was used sured during the six repetitions of the fruit. A fiber optic reflection probe to determine whether there is any experiment. Two CBS-positive and
• June 2016 26(3) 255 RESEARCH REPORTS two CBS-negative spots were shown The average spectral reflectance of both CBS-negative classes created on the citrus samples (Fig. 1). The two CBS lesions (advanced and early very similar spectral signatures. It dataset included spectral data of 1553 stages) and two healthy (healthy spot can be concluded that the nonsymp- CBS positive (458 advanced stage and on CBS-positive fruit and healthy tomatic parts on CBS-infected fruit 1095 early stage) and 1460 CBS spot on CBS-negative fruit) classes had similar spectral characteristics to negative (1026 healthy spot on were shown (Fig. 2). The variation the healthy fruit. All classes had sim- CBS-positive fruit and 434 healthy (SD) within each class and at each ilar reflectance at the wavelengths spot on CBS-negative fruit). Early wavelength was illustrated (Fig. 3). below 500 nm with small standard and advanced stages of CBS lesions The CBS-positive classes had less re- deviations. The spectral signatures of were confirmed by several observa- flectance than healthy classes at the the four classes at the wavelengths tions conducted by plant patholo- range of 500 to 820 nm, which is over 930 nm also generated very gists. Early stage of a CBS hard spot more distinctive in the spectral re- similar patterns. Since the spectrom- lesion is a small and round spot in flectance curve of the advanced stage eter accuracy at the lower and upper dark gray color, whereas the advanced of the CBS-positive class. However, bands of the detector range was not as stage of lesion shows a tan center the CBS-negative spots reflected accurate as the middle band, and the surrounded by a brick-red border. more at the visible band. In addition, four classes generated similar spectral reflectance at the two ends of the spectral range, It can be concluded that the reflectance information ac- quired from the lower and upper bands of the spectrometer range were not useful for CBS identification. The first derivatives of spectral reflectance for each class were shown (Fig. 4). Also the variation within the first derivatives for each class was demonstrated (Fig. 5). The spectral signatures of CBS-positive and CBS- negative classes created different slopes in the 490- to 590- and 630- to 730-nm bands, which are clearly highlighted (Fig. 4). The first deriva- tive graphs for all four classes over- lapped between 590 and 630 nm, which suggested that their spectral Fig. 2. Average spectral reflectance of citrus black spot (CBS) lesions in two classes reflectance at this band changed in (advanced and early stages of hard spot) and healthy spots in two CBS-negative a uniform manner. Between 637 and classes (healthy spot on CBS-positive fruit and healthy spot on CBS-negative fruit) 670 nm, the first derivatives for the on peels of ‘Valencia’ sweet orange fruit within the visible and near-infrared healthy classes and the early stage of spectral ranges. CBS had negative values; however, advanced stage of CBS generated positive first derivative values at the same band. The first derivatives of early stage of CBS and the two CBS- negative classes also overlapped and picked up to 0.6 within the wave band of 675 to 715 nm; while the first derivatives of the advanced stage of CBS never reached over 0.2 at the same band. The first derivatives for advanced stage of CBS had the small- est standard deviations compared with other classes (Fig. 5) that sug- gested more within class spectral sig- nature similarity for this class. At 690 nm, the standard deviations of the first derivatives for the healthy classes as well as the early stage of Fig. 3. Average standard deviations of spectral reflectance of samples in two citrus CBS picked up to 0.35 while t black spot positive (CBS-positive) classes (advanced and early stages of hard spot) remained under 0.14 for the ad- and two CBS-negative classes (healthy spot on CBS-positive fruit and healthy spot vanced stage of CBS at the same band. on CBS-negative fruit) on peels of ‘Valencia’ sweet orange fruit within the visible C OMPARISON BETWEEN and near-infrared spectral ranges. REPETITIONS. The probability values
256 • June 2016 26(3) for the two-sample t test conducted CBS-negative samples were added to probability value less than 0.05, it for the spectral reflectance of every the dataset at the second repetition so can be concluded that there was no two consecutive repetitions and also healthy spot on healthy sample class significant difference between the between the first and last repeti- did not have a probability value at the repetitions. tions were illustrated (Table 1). The first column. Since there was no The probability value for two- sample t test analyses conducted for the first derivatives of spectral reflec- tance of every two consecutive repeti- tions and also between the first and last repetitions were shown (Table 2). First derivative of a spectral reflectance describes the slopes in the spectral signature. No probability value below 0.05 was achieved for the two-sample t tests. These results confirmed that the spectral signatures of samples in each did not significantly change over the period of 11 Apr. and 13 June 2014. It can be inferred that the spectral reflectance of CBS symptoms did not develop or dis- appear during the experiment period. Fig. 4. Average first derivatives of spectral reflectance of samples in two citrus black BAND SELECTION. The normal- spot positive (CBS-positive) classes (advanced and early stages of hard spot) and ized coefficients of all wavelengths two CBS-negative classes (healthy spot on CBS-positive fruit and healthy spot on obtained by the feature ranking CBS-negative fruit) on peels of ‘Valencia’ sweet orange fruit within the visible and methods were illustrated (Fig. 6). near-infrared spectral ranges. Those wavelengths that had coeffi- cients over 0.99 in all ranking methods were selected as the po- tential wavelengths. As shown in Fig. 6, wavelengths between 587 and 589 nm had coefficients over 0.99 in all methods. Therefore, this band was selected for the classifica- tion purpose. The selected band (587 to 589 nm) represents yellow color in the electromagnetic spectrum. As shown in Fig. 2, the spectral reflectances of CBS-positive and CBS-negative clas- ses were distinctively separated be- tween 587 and 589 nm. CLASSIFICATION. The classifica- tion results for all repetitions are shown in Table 3. The overall accu- Fig. 5. Average standard deviations of first derivatives of spectral reflectance of racies varied in a very narrow range samples in two citrus black spot positive (CBS-positive) classes (advanced and early from 93.3% to 94.6%. It can be stages of hard spot) and two CBS-negative classes (healthy spot on CBS-positive concluded that using the selected fruit and healthy spot on CBS-negative fruit) on peels of ‘Valencia’ sweet orange band as the input feature in the fruit within the visible and near-infrared spectral ranges. SVM classifier resulted in analogues
Table 1. Two-sample t test probability values for pairwise comparison between the spectral reflectance of citrus black spot (CBS) lesions and healthy spots on the peels ‘Valencia’ sweet orange fruit for every pair of consecutive repetitions as well as the first and last repetitions of spectral measurements on 11 Apr., 18 Apr., 2 May, 15 May, 29 May, and 13 June 2014. P value for each pair of repetitions Classes 1 vs. 2 2 vs. 3 3 vs. 4 4 vs. 5 5 vs. 6 1 vs. 6 Advanced stage CBS 0.98 0.98 1.00 0.96 0.89 0.85 Early stage CBS 0.96 0.97 0.98 0.95 0.85 0.87 Healthy spot on CBS sample 0.99 1.00 0.98 0.95 0.96 0.87 Healthy spot on healthy sample — 1.00 1.00 1.00 1.00 1.00
• June 2016 26(3) 257 RESEARCH REPORTS
Table 2. Two-sample t test probability values for pairwise comparison between the first derivatives of the spectral reflectance of citrus black spot (CBS) lesions and healthy spots on the peels of ‘Valencia’ sweet orange fruit for every pair of consecutive repetitions as well as the first and last repetitions of spectral measurements on 11 Apr., 18 Apr., 2 May, 15 May, 29 May, and 13 June 2014. P value for each pair of repetitions Classes 1 vs. 2 2 vs. 3 3 vs. 4 4 vs. 5 5 vs. 6 1 vs. 6 Advanced stage CBS 0.84 0.84 1.00 0.59 0.39 0.76 Early stage CBS 0.91 0.82 0.82 0.70 0.27 0.40 Healthy spot on CBS sample 0.92 0.91 0.71 0.71 0.84 0.45 Healthy spot on healthy sample — 1.00 1.00 1.00 1.00 1.00
lesion appeared on the healthy spots of CBS-infected fruit. The wave band of 587 to 589 nm was selected as the most relevant band for detecting CBS. At this band, early and advanced stages of CBS reflected only 8.9% and 5.7% of light, respec- tively, while healthy spots reflected 30.5% of light on CBS infected and 26.7% of light on healthy fruit. Com- pared with healthy orange peel, black spots had darker color and as a result, they absorb more light. The standard deviation for CBS classes was 4.7% at the selected band, whereas it was 7.4% for healthy classes. This values confirmed more variation within each Fig. 6. Normalized coefficients of all wavelengths between 400 and 1030 nm that of healthy classes at the selected band. illustrated the level of relevance of each wavelength to classification of spots on Onereasonmightbethehigher the peels of ‘Valencia’ sweet orange fruit in the range of zero to one in which the reflection from the healthy spots wavelengths with normalized coefficient of zero had no contribution in the compared with the both CBS classes. classification process, while the wavelengths with normalized coefficient of one All four classes had positive first de- had the maximum contribution in the classification process. The coefficients were rivatives ranging from 0.035 to 0.13 acquired by five feature ranking methods [t test, Kullback–Leibler distance, Chernoff bound, receiver operating characteristic (ROC), and Wilcoxon test] and at the band of 587 to 589 nm that they were normalized by the maximum coefficient. The selected band is pointed by indicated an increasing trend for all a red arrow. spectral reflectance curves. The re- sults also showed that the early and advanced stages of CBS at the band classification accuracies in all repeti- spectral data between 587 and 589 nm. of 637 to 670 nm had opposite first tions. These results confirmed that These results were similar to the derivative values. Therefore, the there was no significant difference classification accuracies achieved within spectral information at this band among the spectral reflectance of each repetition. In most of cases, CBS- might be useful for detecting the different repetitions. False negative negative samples were identified with stage of CBS. ratesvariedfrom6.9%to9.2%for higher accuracies compared with CBS- The spectral data at the selected different repetitions while the false positive samples. Again the false nega- band were used as the only input for positive rates were between 0.5% and tiveratewasequalto9.5%,whichwas classifying CBS and healthy spots in 4.6%. The spectrometer probe has higher than the false positive rate thisstudy,whichresultedinanac- a very narrow core with the diameter (3.6%). ceptable overall accuracy of 93.4% of 600 mm. It is possible that the while Bulanon et al. (2013) employed probe was pointed at the area around four wavelengths (three in visible band the hard spot, which was asymptom- Discussion and one in NIR band) to achieve the atic and as a result the spectral re- Spectral reflectances of CBS le- accuracy of 96% in CBS detection. flectance of an asymptomatic spot sions in two stages of the disease as Although their results showed a higher was measured instead of the hard well as two types of healthy spots were accuracy, their proposed vision sensor spot. This measurement error might measured six times over 2 months in required a multispectral camera, which be one reason for higher false nega- 2014. The results showed that the will be more expensive and more tive rates. spectral signatures did not signifi- difficult to use. The selected band in The results of a classification con- cantly change after the fruit were this research was within the visible ducted for the entire dataset were harvested. It can be derived that band and a monochrome camera can shown (Table 4). An overall accu- neither the stage of the disease be used to develop the vision sensor. racy of 93.4% was achieved using the changed after harvesting nor new Therefore, the final product might be
258 • June 2016 26(3) Table 3. Number of spots on the peels of ‘Valencia’ sweet orange fruit that were classified into either citrus black spot positive (CBSD) or CBS-negative (CBSL) classes along with their corresponding classification accuracies and misclassification error for six repetitions of the spectral measurements on 11 Apr., 18 Apr., 2 May, 15 May, 29 May, and 13 June 2014. First repetition (overall accuracy: 94.1%) Second repetition (overall accuracy: 93.3%) Actual class [no. samples (%)] Actual class [no. samples (%)] Prediction CBSD CBSL Sum Prediction CBSD CBSL Sum CBS+ 265 (90.8) 1 (0.5) 266 CBS+ 258 (96.0) 8 (4.0) 266 CBS– 27 (9.2) 184 (99.5) 211 CBS– 20 (8.3) 270 (91.7) 290 Sum 292 185 477 Sum 278 278 556
Third repetition (overall accuracy: 93.8%) Forth repetition (overall accuracy: 94.6%) Actual class [no. samples (%)] Actual class [no. samples (%)] Prediction CBSD CBSL Sum Prediction CBSD CBSL Sum CBS+ 263 (91.3) 10 (3.6) 273 CBS+ 258 (93.1) 10 (3.8) 268 CBS– 25 (8.7) 264 (96.4) 289 CBS– 19 (6.9) 256 (96.2) 275 Sum 288 274 562 Sum 277 266 543
Fifth repetition (overall accuracy: 93.9%) Sixth repetition (overall accuracy: 94.2%) Actual class [no. samples (%)] Actual class [no. samples (%)] Prediction CBSD CBSL Sum Prediction CBSD CBSL Sum CBS+ 217 (90.8) 7 (2.9) 224 CBS+ 166 (92.7) 10 (4.6) 176 CBS– 22 (9.2) 233 (97.1) 255 CBS– 13 (7.3) 207 (95.4) 220 Sum 239 240 479 Sum 179 217 396
Table 4. Number of spots on the peels of ‘Valencia’ sweet orange fruit in the entire dataset of the citrus black spot (CBS) lesions and healthy spots (including Literature cited all six repetitions of the spectral measurements on 11 Apr., 18 Apr., 2 May, 15 Bishop, C.M. 2006. Pattern recognition May, 29 May, and 13 June 2014), which were classified into either CBS-positive and machine learning. 1st ed. Springer, D L (CBS ) or CBS-negative (CBS ) classes along with their corresponding New York, NY. classification accuracies and misclassification error. Bulanon, D.M., T.F. Burks, D.G. Kim, and All datasets (overall accuracy: 93.4%) M.A. Ritenour. 2013. Citrus black spot Actual class [no. samples (%)] detection using hyperspectral image anal- Prediction CBSD CBSL Sum ysis. Agr. Eng. Intl. CIGR J. 15:171–180. CBS+ 1,406 (90.5) 52 (3.6) 1,458 Dewdney, M.M., T.S. Schubert, M.R. Estes, CBS– 147 (9.5) 1,408 (96.4) 1,555 and N.A. Peres. 2014. Florida citrus pest Sum 1,553 1,460 3,013 management guide: Citrus black spot. 10 Aug. 2015.
• June 2016 26(3) 259 RESEARCH REPORTS
Pourreza, A., W.S. Lee, R. Ehsani, J.K. Sankaran, S., A. Mishra, R. Ehsani, and C. Stover, E., R. Driggers, M.L. Richardson, Schueller, and E. Raveh. 2015. An optimum Davis. 2010. A review of advanced techniques D.G. Hall, Y. Duan, and R.F. Lee. 2014. method for real-time in-field detection of for detecting plant diseases. Comput. Elec- Incidence and severity of asiatic citrus Huanglongbing disease using a vision sen- tron. Agr. 72:1–13. canker on diverse citrus and citrus-related sor. Comput. Electron. Agr. 110:221–232. germplasm in a Florida field planting. Schubert, T., M. Dewdney, N. Peres, M. HortScience 49:4–9. Pourreza, A., W.S. Lee, E. Raveh, R. Ehsani, Palm, A. Jeyaprakash, B. Sutton, S. Mondal, and E. Etxeberria. 2014. Citrus Huan- N.-Y. Wang, J. Rascoe, and D. Picton. 2013. Theodoridis, S. and K. Koutroumbas. glongbing detection using narrow-band im- First report of Guignardia citricarpa associ- 2009. Pattern recognition. 4th ed. Elsevier, aging and polarized illumination. Trans. ated with citrus black spot on sweet orange Boston, MA. Amer. Soc. Agr. Biol. Eng. 57:259–272. (Citrus sinensis) in North America. Eur. J. Plant Pathol. 137:635–647. Wang, X., G. Chen, F. Huang, J. Zhang, Qin,J.,T.F.Burks,X.Zhao,N.Niphadkar, K.D. Hyde, and H. Li. 2012. Phyllosticta and M.A. Ritenour. 2012. Development of Snedecor, G.W. and W.G. Cochran. species associated with citrus diseases in a two-band spectral imaging system for 1989. Statistical methods. 8th ed. Iowa China. Fungal Divers. 52:209–224. real-time citrus canker detection. J. Food State Univ. Press, Ames, IA. Eng. 108:87–93.
260 • June 2016 26(3)