阅读说明：本技术 基于高光谱和叶绿素荧光成像融合的柑橘黄龙病检测方法 (Citrus huanglongbing detection method based on hyperspectral and chlorophyll fluorescence imaging fusion ) 是由 聂鹏程 蔺磊 瞿芳芳 于 2019-09-11 设计创作，主要内容包括：本发明公开了一种基于高光谱和叶绿素荧光融合的柑橘黄龙病检测方法,以建模集中的柑橘叶片在基于脉冲调制式的叶绿素荧光测试程序中获得的的叶绿素荧光参数和机遇高光谱图像的特征波段的平均光谱反射率的融合作为LS-SVM判别模型的输入,建立LS-SVM判别模型。本发明提供的基于高光谱成像技术和叶绿素荧光成像技术融合,对于不同季节、不同果园和不同染病阶段的柑橘黄龙病均具有良好的识别效果。(The invention discloses a citrus greening disease detection method based on hyperspectral and chlorophyll fluorescence fusion, which is characterized in that fusion of chlorophyll fluorescence parameters obtained by modeling concentrated citrus leaves in a chlorophyll fluorescence test program based on a pulse modulation mode and average spectral reflectivity of a characteristic waveband of an opportunistic hyperspectral image is used as input of an LS-SVM discrimination model, and the LS-SVM discrimination model is established. The method provided by the invention has a good identification effect on the citrus greening disease in different seasons, different orchards and different infection stages based on the fusion of a hyperspectral imaging technology and a chlorophyll fluorescence imaging technology.)

1. A citrus greening disease detection method based on hyperspectral and chlorophyll fluorescence fusion comprises the following steps:

(a) Collecting citrus leaves, wherein the citrus leaves comprise healthy leaves and infected leaves in different collection times and different growth environments, and the infected leaves comprise leaves in different infection states;

(b) Dividing the citrus leaves in the step (a) into a prediction set and a modeling set according to a certain proportion;

(c) Starting a chlorophyll fluorescence test program based on a pulse modulation type, obtaining chlorophyll fluorescence parameters of citrus leaves concentrated in modeling under different test programs, and if the value of the chlorophyll fluorescence parameters does not fall between 0 and 2, dividing the chlorophyll fluorescence parameters by (Fm-Fo) to enable the range of data to fall between 0 and 2;

(d) Collecting hyperspectral images of the citrus leaves in a wavelength range of 380-1024nm by using a hyperspectral imaging system, and obtaining the average spectral reflectivity of the region of interest in each wave band in the hyperspectral images; deleting a part of wave bands with low signal-to-noise ratio in the wave bands, and performing smooth filtering on the reflectivity of the rest wave bands by adopting a Savitsky-Golay convolution smoothing method to reduce the interference of noise on signals; then, screening out a wave band which is most sensitive to the citrus greening disease from the wave bands subjected to smooth filtering processing by using a continuous projection algorithm; secondly, calculating pairwise correlation coefficients of all sensitive wave bands selected by the SPA by utilizing correlation analysis, judging that strong correlation exists between the spectral reflectivities of the two sensitive wave bands when the correlation coefficient between the spectral reflectivities of the two sensitive wave bands reaches over 0.9, removing one of the sensitive wave bands, and finally determining a characteristic wave band for constructing a discrimination model of the citrus greening disease according to the principle; extracting the average spectral reflectivity of the determined characteristic wave band;

(e) Combining the average spectral reflectivity of the characteristic wave band obtained in the step (d) and the chlorophyll fluorescence parameter obtained in the step (c) together to obtain a fused characteristic, inputting the fused characteristic into an LS-SVM discrimination model, and establishing discrimination models under different test programs;

(f) Identifying the citrus leaves with different collection times in the prediction set by using the discrimination models under different test programs obtained in the step (e), and determining the optimal test program according to the identification result, wherein the model obtained by the test program is the optimal LS-SVM discrimination model;

(g) And (d) obtaining the average spectral reflectivity of the characteristic wave band of the citrus leaf to be tested and the chlorophyll fluorescence parameter under the optimal test program according to the steps (c) and (d), combining the average spectral reflectivity and the chlorophyll fluorescence parameter to obtain fused characteristics, and inputting the fused characteristics into the optimal LS-SVM model obtained in the step (f), so as to judge whether the citrus plant is infected with the yellow dragon disease.

2. The method of claim 1, wherein: in step (a), samples are taken from four orientations of each citrus tree, south-east-west-north.

3. The method of claim 1, wherein: in the step (a), collecting citrus leaves of an orchard 1 and an orchard 2, wherein the growing environments of the orchard 1 and the orchard 2 are different, the citrus leaves comprise healthy leaves and infected leaves, the infected leaves in the orchard 1 are infected and disease-showing leaves, and the infected leaves in the orchard 2 are infected and disease-showing leaves.

4. The method of claim 1, wherein: in the step (b), the data sets of different acquisition times of the citrus leaves are divided into modeling sets and prediction sets by using a Kennard-Stone algorithm, and then the modeling sets of different acquisition times are combined to obtain modeling sets capable of reflecting different seasons for establishing a discrimination model.

5. The method of claim 1, wherein: the chlorophyll fluorescence parameter at least comprises one of the following parameters: fluorescence intensity emitted at steady state, Rfd, Fo, Fv/Fm, Fv/Fo, phi PSII, phi NO.

6. The method of claim 1, wherein: in the step (c), the method for preprocessing the chlorophyll fluorescence image comprises the following steps: and judging whether all pixel points in the region of interest have abnormal values by adopting a 3 sigma criterion, and filling by using adjacent pixel points when judging that the ROI has the abnormal pixel points.

7. The method of claim 1 or 6, wherein: the region of interest is the entire leaf.

8. The method of claim 1, wherein: deleting the first 43 wave bands with low signal-to-noise ratio in the 512 wave bands in the step (d), and smoothing and filtering the average reflectivity of the remaining 469 wave bands by adopting an SG convolution smoothing method; and then selecting sensitive wave band subsets capable of reflecting the Huanglongbing characteristics of the respective acquisition time from different acquisition times by using a continuous projection algorithm, and combining the subsets of all the acquisition times to form sensitive wave bands capable of reflecting the Huanglongbing characteristics of the citrus in different seasons.

9. The method of claim 1, wherein: the characteristic wave band of the step (d) is as follows: 493nm, 515nm, 665nm, 716nm and 739 nm.

10. the method of claim 1, wherein: and (e) taking the radial basis kernel function as the kernel function of the LS-SVM model.

(I) technical field

The invention relates to a citrus greening disease detection method based on fusion of a hyperspectral imaging technology and a chlorophyll fluorescence imaging technology.

(II) background of the invention

The optical technology has the advantages of no damage and high speed for detecting plant diseases and insect pests, thereby having good application prospect. At present, optical instruments commonly used in the field of plant disease and insect pest detection mainly comprise visible-near infrared spectrums, chlorophyll fluorescence, thermal infrared and the like, and imaging equipment and non-imaging equipment are used. The imaging equipment can provide spatial two-dimensional information and can reflect the spatial heterogeneity of disease information in the leaves; with respect to the imaging instrument device, the non-imaging instrument can only acquire information at a limited point in the blade space, but the device structure is relatively simple. Different forms of instrumentation have the respective advantage of being able to detect information relating to the physiological state of the plant from different angles. The visible-near infrared spectrum is related to leaf cell structure and biochemical component information, while chlorophyll fluorescence can reflect actinic light quenching and non-actinic light quenching components in the leaf photosynthesis process and is related to the photosynthesis capacity of plants.

At present, researches on the detection of citrus greening disease based on a hyperspectral imaging technology or a chlorophyll fluorescence imaging technology are reported, and the researches have some effects. However, before the research results are shifted from the laboratory to the practical application, a series of problems need to be solved. For example, the problem of high data latitude generally exists, which is not beneficial to the development of portable instruments and needs to simplify data dimensions; the established citrus huanglongbing model has applicability to infected leaves in different infection stages (from non-obvious disease to obvious disease); the established citrus greening disease discrimination model needs to be suitable for different seasons and different orchards because the season and the orchard environment have large influence on citrus plants; in order to meet the requirements of field use, the defects of high technical cost, long measuring time and the like need to be overcome. Therefore, the method tries to detect the citrus greening disease in different seasons, growth environments and infection degrees by combining the visible-near infrared hyperspectral imaging technology and the chlorophyll fluorescence imaging technology, and lays the foundation for the subsequent development of a handheld citrus greening disease detector.

Disclosure of the invention

the invention aims to provide a citrus greening disease detection method based on the fusion of a hyperspectral imaging technology and a chlorophyll fluorescence imaging technology, which has a good identification effect on citrus greening diseases in different seasons, different growth environments and different infection stages, and has the advantages of low cost and short measurement time.

In order to solve the technical problems, the invention adopts the following technical scheme:

A citrus greening disease detection method based on fusion of a hyperspectral imaging technology and a chlorophyll fluorescence imaging technology comprises the following steps:

(a) collecting citrus leaves, wherein the citrus leaves comprise healthy leaves and infected leaves in different collection times and different growth environments, and the infected leaves comprise leaves in different infection states;

(b) Dividing the citrus leaves in the step (a) into a prediction set and a modeling set according to a certain proportion;

(c) Starting a chlorophyll fluorescence test program based on a pulse modulation type, obtaining chlorophyll fluorescence parameters of citrus leaves concentrated in modeling under different test programs, and if the value of the chlorophyll fluorescence parameters does not fall between 0 and 2, dividing the chlorophyll fluorescence parameters by (Fm-Fo) to enable the range of data to fall between 0 and 2;

(d) Collecting hyperspectral images of the citrus leaves in a wavelength range of 380-1024nm by using a hyperspectral imaging system, and obtaining the average spectral reflectivity of the region of interest in each wave band in the hyperspectral images; deleting a part of wave bands with low signal-to-noise ratio in the wave bands, and performing smooth filtering on the reflectivity of the rest wave bands by adopting a Savitsky-Golay (SG) convolution smoothing method to reduce the interference of noise on signals; then utilizing a continuous projection algorithm (SPA) to screen out a wave band which is most sensitive to the citrus greening disease from the wave bands subjected to smooth filtering processing; secondly, calculating pairwise correlation coefficients of all sensitive wave bands selected by the SPA by utilizing correlation analysis, judging that strong correlation exists between the spectral reflectivities of the two sensitive wave bands when the correlation coefficient between the spectral reflectivities of the two sensitive wave bands reaches over 0.9, removing one of the sensitive wave bands, and finally determining a characteristic wave band for constructing a discrimination model of the citrus greening disease according to the principle; extracting the average spectral reflectivity of the determined characteristic wave band;

(e) combining the average spectral reflectivity of the characteristic wave band obtained in the step (d) and the chlorophyll fluorescence parameter obtained in the step (c) together to obtain a fused characteristic, inputting the fused characteristic into an LS-SVM (least square vector machine) discrimination model, and establishing discrimination models under different test programs;

(f) Identifying the citrus leaves collected at different times in the prediction set by using the discrimination models under different test programs obtained in the step (e), and determining the optimal test program according to the identification result, wherein the model obtained by the test program is the optimal LS-SVM discrimination model;

(g) and (d) obtaining the average spectral reflectivity of the characteristic wave band of the citrus leaf to be tested and the chlorophyll fluorescence parameter under the optimal test program according to the steps (c) and (d), combining the average spectral reflectivity and the chlorophyll fluorescence parameter to obtain fused characteristics, and inputting the fused characteristics into the optimal LS-SVM model obtained in the step (f), so as to judge whether the citrus plant is infected with the yellow dragon disease.

in step (a) of the present invention, since citrus yellow shoot pathogens are not uniformly distributed in the host, it is preferable to sample from four locations, south, east and west, of each citrus tree.

In the step (a) of the present invention, the "different time" for the collection of the citrus fruit leaves can be determined according to actual needs, such as monthly or quarterly. The different growing environments mean that at least two orchards have different growing environments. The affected leaves comprise leaves in different affected states, and at least comprise affected leaves and affected leaves. Preferably, the method is used for collecting citrus leaves in an orchard 1 and an orchard 2, the growing environments of the orchard 1 and the orchard 2 are different, the citrus leaves comprise healthy leaves and infected leaves, the infected leaves in the orchard 1 are infected and disease-showing leaves, and the infected leaves in the orchard 2 are infected and disease-non-showing leaves.

in step (b) of the present invention, the citrus fruit leaf preferably utilizes the Kennard-Stone (KS) algorithm to separate the data sets at different acquisition times into a modeling set and a prediction set. And then combining the modeling sets of different acquisition times to obtain a modeling set capable of reflecting different seasons, and finally using the modeling set for establishing a discriminant model.

In step (c) of the present invention, the classic pulse modulation-based chlorophyll fluorescence measurement procedure is roughly as follows:

(1) fully adapting the blade in dark;

(2) Turning on the measuring light, and measuring the initial fluorescence yield Fo under dark adaptation;

(3) After a period of time, turning on the saturated light, and measuring the maximum fluorescence yield Fm under dark adaptation;

(4) Then, opening the actinic light, wherein the intensity of the actinic light can maintain the normal photosynthesis of the leaves until the leaves reach a stable state, and the measured fluorescence is the stable fluorescence Fs;

(5) Applying a saturated light again, and measuring the maximum fluorescence yield Fm' under the light adaptation;

(6) And (3) closing actinic light, enabling the blade to enter a dark relaxation state, opening far red light for 3-5s, and reoxidizing the electron transfer chain. The minimal fluorescence Fo' under photopic conditions was measured. In the case of a light source not equipped with far-red light, an approximation can be obtained by the formula Fo/(Fv/Fm + Fo/Fm').

preferably, the saturation light intensity is set to 1500. mu. mol photons.m ^-2. s ^-1, and the intensity of the actinic light is set to correspond to the average light intensity in the orchard.

preferably, the dotting time mode of the measuring program is shown in fig. 5-3, and is specifically selected from one of the following test programs: only nine measurement procedures are Dark adaptation (Dark), L1 (from Dark adaptation to L1) 32.24s, L2 (from Dark adaptation to L2) 42.24s, L3 (from Dark adaptation to L3) 52.24s, L4 (from Dark adaptation to L4) 72.24s, Lss (from Dark adaptation to Lss) 92.24s, D1 (from Dark adaptation to D1) 122.24s, D2 (from Dark adaptation to D2) 152.24s, and D3 (from Dark adaptation to D3) 184.24 s. In addition to dark adaptation, for each applied saturated light, a set of fluorescence parameters corresponding to the time instant is obtained.

preferably, the chlorophyll fluorescence parameters include at least one of: fluorescence intensity emitted at steady state, Rfd, Fo, Fv/Fm, Fv/Fo, phi PSII, phi NO.

In the step (c), if the values of some pixel points in the obtained chlorophyll fluorescence image obviously deviate from the values of other pixel points in the fluorescence image, the abnormal pixel points need to be preprocessed, so that the final result is more accurate. The pre-processing method recommended by the invention comprises the following steps: judging whether all pixel points of a Region of interest (ROI) have abnormal values by adopting a 3 sigma criterion, and filling by using adjacent pixel points when judging that the ROI has the abnormal pixel points, preferably replacing by using the average value of a plurality of adjacent pixel points. The region of interest described in the present invention is preferably the entire leaf.

In the step (d), the hyperspectral imaging system needs to be subjected to parameter adjustment before the hyperspectral image is collected, so that the requirement of the working condition set by the test is met.

In the step (d), because the hyperspectral image acquisition system does not perform equidistant sampling when acquiring the image, only 512 pieces of band information can be acquired within the band range of 380 plus 1030 nm. In order to eliminate the error of the reflectivity of each point in the ROI area, the reflectivity of all pixel points in the ROI area at each wave band is averaged, and finally the average reflectivity of the ROI area at 512 wave bands is obtained. Since the spectrum may contain noise, it is preferable that the first 43 bands with low snr of the 512 bands are deleted in step (d), and the average reflectivity of the remaining 469 bands is smoothed by SG convolution smoothing. Preferably, the step of screening the wave bands most sensitive to citrus greening disease from the wave bands subjected to the smoothing filtering by using the SPA comprises the following steps: firstly, SPA is utilized to select sensitive wave band subsets capable of reflecting the characteristics of the citrus greening disease of the respective acquisition time from different acquisition times, and then the subsets of all the acquisition times are combined (union set) to form sensitive wave bands capable of reflecting the characteristics of the citrus greening disease in different seasons.

Preferably, the characteristic bands in step (d) are: 493nm, 515nm, 665nm, 716nm and 739 nm.

Preferably, in the step (e), a Radial Basis Function (RBF) is used as a kernel function of the LS-SVM model. After the RBF kernel function is selected, a Grid searching (Grid searching) algorithm is further applied to optimize the parameter sigma and the regularization parameter gamma of the RBF kernel function, so that the LS-SVM model obtains the best prediction effect.

Compared with the prior art, the invention has the beneficial effects that: the citrus greening disease detection method based on the fusion of the hyperspectral imaging technology and the chlorophyll fluorescence imaging technology has good identification effect on citrus greening diseases in different seasons, different orchards and different infection stages. According to the invention, after two technologies are combined, the reflectivity of five characteristic bands (493nm, 515nm, 665nm, 716nm and 739nm) and the 29 chlorophyll fluorescence parameters obtained by the measurement procedure L2 are fused, so that the difference among three types of samples, namely healthy, infected and deficient samples can be effectively increased, and the detection time of a single sample is shortened (reduced by 58.3%). In the whole experiment period, the overall identification accuracy of the citrus huanglongbing of the orchard 1 with the larger disease degree and the orchard 2 with the lower disease degree is superior to the result obtained by singly using the spectral reflectivity of the characteristic wave band or the model established by the chlorophyll fluorescence parameter obtained by the measuring program L2, more importantly, the missing rate is reduced (11.6 percent and 17.6 percent are respectively reduced in 5 months and 9 months in the orchard 2), and the defects of high cost and long measuring time respectively existing in the singly using a visible-near infrared imaging technology or a chlorophyll hyperspectral fluorescence imaging technology are overcome.

(IV) description of the drawings

FIG. 1 is a chlorophyll fluorescence imaging system.

fig. 2 is a procedure for measuring the chlorophyll fluorescence signal of citrus.

FIG. 3 is the chlorophyll fluorescence image before outliers are removed: (a) and (b) comparative effect graph after treatment.

FIG. 4 is a visible-near infrared hyperspectral imaging system.

Fig. 5 is a graph of the average spectral reflectance of typical healthy (n-1296), non-diseased (n-648), diseased (n-648) and cellulose deficient citrus leaves (n-201).

Fig. 6 shows the sensitivity bands for different months selected based on the sequential projection method.

fig. 7 is a graph of pairwise correlation coefficients between spectral reflectivities of selected 16 sensitive bands projected in succession.

Fig. 8 is an image of leaves in different seasons of the orchard 1 at different wavebands and corresponding second principal components.

(V) detailed description of the preferred embodiments

The technical solution of the present invention is further described below with reference to the accompanying drawings, but the scope of the present invention is not limited thereto: