Tomato plant drought stress detection method based on hyperspectral imaging

Abstract

The application particularly relates to a kind of tomato plant drought stress detection methods based on hyperspectral imaging, comprising the following steps: collecting the hyperspectral image of the tomato leaf to be identified, and extracting the reflectance spectrum data of leaf according to the hyperspectral image;Characteristic wavelength is screened using genetic algorithm, and the best reflectance image set is determined according to the correlation between the reflectance images corresponding to characteristic wavelength;Deep image features of the best reflectance image set are extracted using convolutional neural network;After the fusion of the spectrum and image features of leaf, it is input into the trained plant drought stress identification model to be identified and the drought stress grade of the tomato to be identified is obtained.The combination of spectrum and image is used to select reflectance image, reduce data redundancy, realize the maximum use of information;Through convolutional neural network, image features are automatically extracted, which is simple and effective, and complex mathematical calculation is avoided;Fusion of spectral and image features improves model recognition effect, provides complementary information, and avoids information loss.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of tomato plant drought stress detection, and particularly relates to a tomato plant drought stress detection method based on hyperspectral imaging. BACKGROUND

[0002] Drought stress is a main factor affecting the growth and development of tomato plants. Water deficiency leads to a decrease in stem diameter, a decrease in photosynthesis, an increase in water potential in the plant body, stomatal closure, an increase in stomatal resistance, and a significant decrease in dry matter accumulation. Water stress also causes an increase in the incidence of tomato collar rot, which in turn leads to a decrease in yield and quality. Therefore, a method for rapidly detecting drought stress in tomato plants is needed. In the prior art, traditional methods for detecting plant drought stress, such as manual visual detection, canopy temperature, thermal imaging, chlorophyll fluorescence, and RGB imaging technology, have long detection periods, complex detection methods, and relatively low test accuracy, which cannot meet the detection requirements in actual applications. Therefore, it is of great significance to develop a rapid, non-destructive, accurate, and highly sensitive tomato plant drought stress detection device and method. SUMMARY

[0003] The present application relates to the technical field of tomato plant drought stress detection, and particularly relates to a tomato plant drought stress detection method based on hyperspectral imaging.

[0004] To achieve the above object, the technical solution adopted by the present application is as follows: a tomato plant drought stress detection method based on hyperspectral imaging, comprising the following steps: collecting a hyperspectral image of a tomato leaf to be identified, extracting reflectance spectrum data of the leaf from the hyperspectral image; using a genetic algorithm to select characteristic wavelengths, determining the best set of reflectance images based on the correlation between the reflectance images corresponding to the characteristic wavelengths; using a convolutional neural network to extract deep image features of the best set of reflectance images; and inputting the fused spectral and image features of the leaf into a trained plant drought stress recognition model for recognition and obtaining the drought stress level of the tomato to be identified.

[0005] Compared with the prior art, the present application has the following technical effects: the combination of spectrum and image is used to select reflectance images, which reduces data redundancy and maximizes information utilization; the convolutional neural network is used to automatically extract image features, which is simple and effective and avoids complex mathematical calculations; the fusion of spectral and image features improves the model recognition effect, provides complementary information, and avoids information loss; the fusion of sample features of different growth types of leaves is used to comprehensively evaluate the drought stress state of the plant, and the sub-sample fusion provides difference information of various types of samples and reduces the generalization error. BRIEF DESCRIPTION OF DRAWINGS

[0006] Figure 1 is a flowchart of the present application.

[0007] Figure 2 This is a diagram of the plant drought stress identification model in this invention;

[0008] Figure 3 yes Figure 2 A schematic diagram of the ECA module in the diagram;

[0009] Figure 4 This is a schematic diagram of the acquisition device structure in this invention;

[0010] Figure 5 This is a flowchart of the data acquisition device operation. Detailed Implementation

[0011] The following is combined Figures 1 to 5 The present invention will be described in further detail below.

[0012] See Figure 1 This invention discloses a method for detecting drought stress in tomato plants based on hyperspectral imaging, comprising the following steps: acquiring hyperspectral images of tomato leaves to be identified; extracting reflectance spectral data of the leaves from the hyperspectral images; using a genetic algorithm to screen feature wavelengths and determining the optimal reflectance image set based on the correlation between reflectance images corresponding to the feature wavelengths; using a convolutional neural network to extract deeper image features from the optimal reflectance image set; and fusing the spectral and image features of the leaves and inputting them into a trained plant drought stress identification model to identify the drought stress level of the tomato plant. The method combines spectral and image analysis to select reflectance images, reducing data redundancy and maximizing information utilization; the automatic extraction of image features via a convolutional neural network is simple and effective, avoiding complex mathematical calculations; and the fusion of spectral and image features improves the model's recognition performance, provides complementary information, and avoids information loss.

[0013] Furthermore, in the steps of acquiring hyperspectral images of tomato leaves to be identified and fusing spectral and image features of characteristic wavelengths, the acquired and fused spectral and image features are those of two types of leaves from the same plant: young leaves and mature leaves. By fusing leaf sample features from different growth types to comprehensively assess the drought stress status of the plant, subsample fusion can provide differential information on multiple sample types, reducing generalization error.

[0014] Further, in the step of screening characteristic wavelengths using a genetic algorithm, the support vector machine is used as an evaluator during screening, the initial population, the number of iterations, the crossover probability and the variance probability are set to 100, 400, 0.5 and 0.1 respectively, and 5-fold cross-validation is used to find the global optimal solution in the offspring, which specifically includes the following steps: first apply the genetic algorithm to obtain the accuracy rate of the test set as a reference value; execute GA in a loop, and set the number of loops to 1000; if the running result is greater than the reference value, designate the result as a new reference value, exit the loop, and return to the previous step; if no better result is obtained after 1000 consecutive executions, the previously obtained characteristic wavelength is taken as the best characteristic wavelength. In this way, the best characteristic wavelength can be screened out. The best characteristic wavelength mentioned here refers to a strongly correlated characteristic wavelength that can improve the accuracy rate of the test set, so as to avoid data redundancy caused by excessive band data in hyperspectral data.

[0015] Further, the step of determining the best reflectance image set according to the correlation between the reflectance images corresponding to the characteristic wavelengths includes the following steps: calculating the weight value of the best characteristic wavelength according to the RelifF method; taking the reflectance image corresponding to the characteristic wavelength with the maximum weight value as a reference image; placing the reference image into a defined container; setting a correlation threshold according to the Pearson correlation, and the correlation threshold range is [-0.3, 0.3]; sequentially analyzing the correlation between the reflectance images of other characteristic wavelengths and all images in the container, and adding the reflectance images that meet the threshold condition to the container, finally obtaining a set of mutually unrelated reflectance images; calculating the correlation between the reference image and other reflectance images in the container; according to the correlation ranking, sequentially increasing the number of images to obtain different reflectance image combinations; and determining the best reflectance image set through classifier modeling analysis. The characteristic wavelength itself can be used as a feature. In order to further improve the feature quantity and improve the accuracy of identification, the best reflectance image set is determined according to the screened best characteristic wavelength, so that the spectral and image features of the characteristic wavelength can be fused in the subsequent steps, thereby obtaining more accurate identification results.

[0016] The method used by the plant drought stress identification model herein is to construct a dense convolutional neural network (DACNN) with efficient channel attention (ECA) by using modules such as convolution blocks, dense connection blocks and attention mechanisms. The convolutional neural network (CNN) is generally composed of convolution layers, pooling layers and fully connected layers. The convolution layer continuously learns different characteristics of the input data, and has the characteristics of local connection and weight sharing. The pooling layer retains the most important features while reducing the feature dimension to avoid overfitting. The fully connected layer maps the feature map output by the previous layer into a feature vector and generates a probability vector belonging to different categories.

[0017] Referring to Figure 2 , in particular, the plant drought stress recognition model comprises: using a two-dimensional convolution method to perform dimensionality operation on the fused one-dimensional features to obtain feature Figure 1 In this embodiment, a two-dimensional convolution method is used, the convolution kernel size is 3*3, the step is 1, the padding is 1, the convolution kernel moves on the input feature map, and point multiplication operation is performed. The feature Figure 1 is subjected to three times of dense connection operation to obtain feature Figure 2 An ECA module is added in the dense connection operation to perform cross-channel interaction, and the three times of dense connection is also called a dense connection block. The dense connection block transmits the outputs of all previous layers to the next layer, strengthens feature transmission, realizes feature reuse, and effectively combines low-level and deep features. The convolution layer is used to extract the main features of the feature Figure 2 and eliminate unnecessary features; after sequentially passing through the pooling layer, the flattening layer and the full connection layer, three neurons are obtained, and the three neurons correspond to the mild, moderate and severe drought stress grades of the tomato, respectively. The recognition model of this structure can reliably and accurately identify the drought stress grade of the tomato plant.

[0018] For the application of such a neural network, generally, after the model is built, the model parameters are initialized, and then the sample set is used to train to obtain a truly usable recognition model. Further in the present application, the plant drought stress recognition model is trained by the following steps: samples are made and manually labeled with their true drought stress grades; the sample set is divided into a training set and a test set according to a certain proportion; the sample set is processed according to the foregoing steps to obtain predicted drought stress grades; the parameters in the plant drought stress recognition model are optimized according to the true drought stress grades and the predicted drought stress grades of the training set; the accuracy of the model prediction is calculated according to the true drought stress grades and the predicted drought stress grades of the test set, if the accuracy is greater than a set threshold, the current parameters are saved to obtain a trained plant drought stress recognition model; otherwise, return to step one to make new samples. According to these steps, a plant drought stress recognition model that meets the requirements can be trained more conveniently and quickly.

[0019] Further preferably, the three times of dense connection operation formula is as follows:

[0020]

[0021] wherein [X0, X1, … X l-1 ] represents that the output feature maps of 0 to (l-1) layers are merged in the channel, H l represents an operable composite function, for example, the operable composite function corresponding to the ECA module mentioned above, X0 is the feature Figure 1 , and X3 is the featureFigure 2 .

[0022] Referring to Figure 3 , further preferably in the present application, the ECA module processes the feature map as follows: after using non-reduced dimension global average pooling to aggregate the convolution features, the coverage of cross-channel interaction is adaptively adjusted by controlling the kernel size; one-dimensional convolution is used instead of the full connection operation in SENet; the Sigmoid function is used to learn the channel attention and give the features importance.

[0023] Referring to Figure 4 , as a preferred scheme of the present application, in the step of collecting the hyperspectral image of the tomato leaf to be identified, a collection device is used to collect the hyperspectral image of the tomato leaf to be identified; the collection device comprises a mobile measurement platform 10, a lifting rod 20, a lifting platform 40, a conveying unit 50, an illumination unit 30 and a spectrometer 70.

[0024] The mobile measurement platform 10 is the main body carrying the entire detection device, which comprises a square base for supporting all detection components, is used to move the entire device to a designated position and perform measurement, and rollers are arranged below the square base of the mobile measurement platform 10 for convenient movement. Specifically in the present application, the size of the square base is 0.45m*0.45m, the square base is made of aluminum alloy material and the surface is black oxide treated, which provides a purer background and is more conducive to subsequent spectrum collection.

[0025] The lifting rod 20, the lifting platform 40 and the conveying unit 50 are arranged on the center line of the square base in sequence. The lifting rod 20 is arranged on one side of the center line of the mobile measurement platform 10 close to one of the short sides, and the length direction of the lifting rod 20 is perpendicular to the square base, the stroke is greater than or equal to 1 m, and the lifting rod 20 is provided with a scale display. The up-down transmission adopts a rack and pinion structure. The spectrometer 70 is fixed on the lifting rod 20 by means of a screw knob, and the height of the spectrometer 70 can be adjusted at will by the knob handle. The spectrometer 70 adopts a Headwall Nano-Hyperspec push-broom sensor, which provides 272 measurement bands and 640 spatial channels, and has the advantages of fast acquisition speed and low power consumption. The spectrometer 70 contains a network interface 801, which is convenient for connecting the display unit 80. The infrared distance sensor 60 is arranged on the spectrometer 70 for detecting the height of the spectrometer 70. The infrared distance sensor 60 uses a GP2Y0A21YK0F sensor module, and the detection distance range is 10 cm-80 cm, and the power voltage is 4.5 V-5.5 V. The lifting platform 40 is used for supporting the tomato plant to be identified. The lifting platform 40 can drive the tomato plant to move in the vertical direction in an electric manner, so that suitable leaves can be found for imaging. The workbench of the lifting platform 40 has a size of 0.15 m x 0.15 m, can be electrically lifted, has a total power of 60 W, a maximum lifting height of 40 cm, a maximum load of 20 kg, and can be completely fitted with the square base of the mobile measurement platform 10 when not lifted, so that the materials and space are effectively utilized. The conveying unit 50 extends to the outside of the mobile measurement platform 10 for moving the tomato plant, reduces the labor cost, and improves the detection efficiency. Specifically, a 4020 aluminum profile is used to build a frame, and a PVC black belt is used, the belt width is 0.2 m, the length is 0.4 m, and the speed and running direction can be adjusted.

[0026] The lighting unit 30 is arranged on both sides of the center line to provide illumination for the tomato plant on the lifting platform 40. The lighting unit 30 mainly uses halogen lamps 301 as the main radiation and blue lamps 302 as the auxiliary radiation to provide the illumination conditions required for imaging. Specifically, the lighting unit 30 includes two halogen lamps 301 with a wavelength range of 350 nm-2500 nm and two blue lamps 302 with a wavelength range of 450 nm-490 nm. The two halogen lamps 301 are symmetrically arranged on both sides of the tomato plant to be identified, and the two blue lamps 302 are symmetrically arranged about the tomato plant to be identified. Such an arrangement can remove high-frequency noise in the visible region, and can ensure uniform illumination.

[0027] The acquisition device further includes a display unit 80. The display unit 80 displays the spectral curve of the tomato leaf in real time. Specifically, the display unit 80 is a LEDSMD0808 full-color display screen, and has a refresh frequency of 1920 Hz and a visual angle of 110°-120°. The display unit 80 is more conducive to interaction with the operator and can directly display some information.

[0028] The specific operation process of the device is shown in Figure 5 Firstly, the position of the spectrometer 70 in the lifting rod 20, i.e. the height, is determined, and then the exposure time, frame period, scanning speed and other parameters are set; at the same time, the height of the light unit 30 is adjusted and preheated for 30 minutes to prevent the light intensity from changing with time; secondly, by loosening the screw, the focal length is adjusted, the white board and the cover lens are scanned to obtain a black and white reference image; the height of the lifting platform 40 is adjusted by electricity, and the distance between the leaf and the lens is measured by the infrared distance sensor 60 to select the appropriate leaf imaging; then the hyperspectral image of the leaf is shot and the data is saved; finally, different types of leaves of the same plant are selected for imaging, and leaf samples on different plants are obtained by using the conveying belt, and the drought stress data set of the tomato plant or the hyperspectral image of the tomato plant to be detected is obtained to facilitate subsequent detection. The collection device improves the imaging sensitivity by supplementing blue light illumination, removes the high-frequency noise in the visible region, adopts real-time display and automatic analysis method to feedback the drought situation of the plant, which can reduce the human resources and detection time, and the device is simple and easy to operate, and the detection personnel can use and operate after simple learning, which can greatly improve the detection efficiency.

Claims

1. A method for detecting drought stress in tomato plants based on hyperspectral imaging, characterized by: The method comprises the following steps: Collecting hyperspectral images of tomato leaves to be identified, and extracting reflectance spectrum data of the leaves from the hyperspectral images; Screening characteristic wavelengths by using a genetic algorithm, and determining an optimal reflectance image set according to the correlation between reflectance images corresponding to the characteristic wavelengths; Extracting deep image features of the optimal reflectance image set by using a convolutional neural network; Fusing the spectrum and image features of the leaves, and inputting the fused features into a trained plant drought stress identification model to identify the drought stress grade of the tomato to be identified; The step of determining the optimal reflectance image set according to the correlation between reflectance images corresponding to the characteristic wavelengths comprises the following steps: Calculating weight values of the optimal characteristic wavelengths according to a RelifF method; Taking a reflectance image corresponding to a characteristic wavelength with the largest weight value as a reference image; Putting the reference image into a defined container; Setting a correlation threshold according to a Pearson correlation; Analyzing the correlation between the reflectance image of the other characteristic wavelength and all images in the container in sequence, adding the reflectance image meeting the threshold condition to the container, and finally obtaining a set of mutually unrelated reflectance images; Calculating the correlation between the reference image and other reflectance images in the container; According to the correlation ranking, sequentially increasing the number of images, and obtaining different reflectance image combinations; Determining the optimal reflectance image set through classifier modeling analysis; The plant drought stress identification model comprises: Performing dimension upgrading on the fused one-dimensional features by using a two-dimensional convolution method to obtain feature map one; Performing dense connection operation on the feature map one for three times to obtain feature map two, and adding an ECA module for cross-channel interaction in the dense connection operation; Extracting main features of the feature map two and eliminating unnecessary features by using a convolution layer; After sequentially passing through a pooling layer, a flattening layer and a full connection layer, three neurons are obtained, and the three neurons correspond to the mild, moderate and severe drought stress grades of the tomato, respectively.

2. The method for detecting drought stress in tomato plants based on hyperspectral imaging according to claim 1, characterized in that: In the steps of collecting the hyperspectral images of the tomato leaves to be identified and fusing the spectrum and image features of the characteristic wavelengths, the spectrum and image features of two types of leaves on the same plant are collected and fused, and the two types of leaves are young leaves and mature leaves.

3. The method for detecting drought stress in tomato plants based on hyperspectral imaging according to claim 2, characterized in that: In the step of screening the characteristic wavelengths by using the genetic algorithm, a support vector machine is used as an evaluator, the initial population, the number of iterations, the crossover probability and the variance probability are set to 100, 400, 0.5 and 0.1 respectively, 5-fold cross-validation is used to find a global optimal solution in the offspring, and the step specifically comprises the following steps: Firstly, applying the genetic algorithm to obtain the accuracy rate of the test set as a reference value; Executing the GA in a loop, and setting the number of loops to 1000; If the running result is greater than the reference value, the result is specified as a new reference value, the loop is exited, and the previous step is returned; If no better result is obtained by continuously executing the GA for 1000 times, the previously obtained characteristic wavelengths are taken as the optimal characteristic wavelengths.

4. The method for detecting drought stress in tomato plants based on hyperspectral imaging according to claim 1, characterized in that: The plant drought stress identification model is trained through the following steps: Preparing samples and manually marking the real drought stress grades of the samples; Dividing the sample set into a training set and a test set according to a certain proportion; The sample set is processed according to the steps in claim 1 to obtain a predicted drought stress level; According to the real drought stress level and the predicted drought stress level of the training set, the parameters in the plant drought stress identification model are optimized; According to the real drought stress level and the predicted drought stress level of the test set, the accuracy of the model prediction is calculated, if the accuracy is greater than the set threshold, the current parameters are saved to obtain the trained plant drought stress identification model; otherwise, return to step to start to make samples again.

5. The method for detecting drought stress in tomato plants based on hyperspectral imaging as claimed in claim 1, wherein: The continuous three dense connection operation formulas are as follows: wherein, representing the output feature map of the 0 to ) layer is merged in the channel, representing the operable composite function, i.e. feature map one, i.e. feature map two.

6. The method for detecting drought stress in tomato plants based on hyperspectral imaging as claimed in claim 1, wherein: The ECA module processes the feature map according to the following steps: After using the global average pooling to aggregate the convolution features, the coverage range of the cross-channel interaction is adaptively adjusted by controlling the kernel size; A one-dimensional convolution is used instead of the full connection operation in SENet. The Sigmoid function is used to learn the channel attention and give the features importance.

7. The method for detecting drought stress in tomato plants based on hyperspectral imaging as claimed in claim 1, wherein: In the step of collecting the hyperspectral image of the tomato leaf to be identified, the hyperspectral image of the tomato leaf to be identified is collected by using a collection device; the collection device includes a mobile measurement platform (10), a lifting rod (20), a lifting platform (40), a conveying unit (50), an illumination unit (30), and a spectrometer (70); the mobile measurement platform (10) includes a square base and rollers arranged below the base; the lifting rod (20), the lifting platform (40), and the conveying unit (50) are arranged in sequence on the center line of the square base; the spectrometer (70) is installed on the lifting rod (20); the lifting platform (40) is used to support the tomato plant to be identified; the conveying unit (50) extends to the outside of the mobile measurement platform (10) to move the tomato plant; and the illumination unit (30) is arranged on both sides of the center line to provide illumination for the tomato plant on the lifting platform (40).

8. The method for detecting drought stress in tomato plants based on hyperspectral imaging as claimed in claim 7, wherein: The square base is made of aluminum alloy material and the surface is blackened; the conveying unit (50) uses a PVC black belt; the illumination unit (30) includes two halogen lamps (301) and two blue lamps (302); the two halogen lamps (301) are symmetrically placed on both sides of the tomato plant to be identified; and the two blue lamps (302) are placed in an oblique symmetry about the tomato plant to be identified; the collection device further includes an infrared distance sensor (60) and a display unit (80); the infrared distance sensor (60) is arranged on the spectrometer (70) to detect the height of the spectrometer (70); and the display unit (80) displays a scanning view in real time to show the spectral curve of the tomato leaf.

Images