A method for estimating chlorophyll content based on inversion of optical characteristic parameters

Abstract

The present invention discloses a chlorophyll content estimation method based on optical characteristic parameter inversion, which comprises the following steps: (1) constructing a leaf model based on a surface light source to simulate photon transmission and obtain a simulated spectral image; (2) inputting the simulated spectral image into a convolutional neural network for training and obtaining a pre-trained model; (3) based on the pre-trained model, performing maximum mean difference (MMD) migration on measured spectral data and obtaining an optical parameter inversion model of all training layers; (4) using the optical parameter inversion model obtained by MMD migration to estimate the chlorophyll content of the leaf. The optical parameter inversion method of green radish leaf based on MMD migration has better performance than the common model migration method.

Description

Technical Field

[0001] The invention relates to the field of high-quality cultivation, and in particular to a chlorophyll content estimation method based on optical characteristic parameter inversion. Background Art

[0002] Green radish is an excellent indoor foliage plant. In addition to its own ornamental value, it also plays an important role in purifying indoor chemical pollution. Chlorophyll content, as an important factor affecting plant photosynthesis, is a key indicator for evaluating the growth status of green plants. Therefore, rapid detection of chlorophyll content in green radish is of great significance for improving indoor air pollution. Chlorophyll content detection methods can be divided into two categories: destructive chemical detection and non-destructive detection. Non-destructive detection methods mainly use chlorophyll meters and hyperspectral imaging technology. Traditional chemical methods have defects such as time-consuming, labor-intensive, poor timeliness, and damage to plant body structure, and have poor applicability. The chlorophyll meter is easy to operate, but because it can only analyze local point spectra, it is impossible to fully obtain the chlorophyll information of the leaves or canopy as a whole, and there are still great limitations in practical applications.

[0003] Spectral analysis technology combines spectral and image information to achieve fast, non-destructive measurements and has been the most widely used in recent years. Spectrometers are widely used because of their portability, speed, and non-destructiveness. However, since portable spectrometers use point light source measurement, the spectral information obtained from samples is limited. The hyperspectral images captured by hyperspectral cameras contain a large amount of image information and spectral information at the same time, which makes up for the shortcomings of traditional spectrometers and digital cameras in obtaining image information. They have been widely used in the field of agricultural engineering.

[0004] Most of the hyperspectral methods used today to detect chlorophyll content directly link spectral images with chlorophyll content by constructing simple models, ignoring the absorption and scattering of light inside plant tissues, and failing to express complete information about the chlorophyll content inside the entire plant tissue. Therefore, when using spectral technology for analysis, it is necessary to study the transmission mechanism of light in plant tissues. Combining spectral information with optical characteristic parameters can make up for the shortcomings of single spectral information. Reconstruction of optical characteristic parameters of strongly scattering biological tissues [J]. Journal of Hubei University (Natural Science Edition), 2006, 28(1): 45-47. The genetic annealing algorithm is used to reconstruct the optical characteristic parameters and obtain μ a and μ sThe inversion errors are 8% and 0.1% respectively, but this method uses traditional algorithms. Considering the use of deep learning methods, more feature information can be mined. Monte Carlo simulation of biological tissue optical model [D]. Wuhan University of Technology, 2013. The optical characteristic parameters of the multi-layer model are inverted using BP neural network. The error of the double-layer plate model is 1.84%, and the error of the three-layer plate model is 1.73%, which is higher than the linear fitting and perturbation Monte Carlo methods. The deep learning method performs well in the inversion of optical characteristic parameters, but this method relies on a large number of samples for model training and requires more manpower and material resources to obtain data. Based on transfer learning and Mask Rice planthopper image classification method based on R-CNN[J]. Transactions of the Chinese Society of Agricultural Machinery, 2019, 50(07): 201-207.DOI:10.6041 / j.issn.1000-1298.2019.07.021. Plant leaf image recognition method based on convolutional neural network based on transfer learning[J]. Transactions of the Chinese Society of Agricultural Machinery, 2018, 49(S1): 354-359.DOI:10.6041 / j.issn.1000-1298.2018.S0.047 and Camellia oleifera disease image recognition based on convolutional neural network and transfer learning [J]. Transactions of the Chinese Society of Agricultural Engineering, 2018, 34(18): 194-201.DOI:10.11975 / j.issn.1002-6819.2018.18.024. Transfer learning is applied to the field of image recognition, and better results are obtained than direct training. However, the above methods have certain defects. When the measured data is small, the results will be inaccurate. Therefore, a simulation method is proposed to generate samples with a large similarity to the measured data distribution for the training of the pre-training model. Convolutional Neural Network for Corn Disease Image Recognition Based on Transfer Learning[J]. Transactions of the Chinese Society of Agricultural Machinery, 2020, 51(02): 230-236+253.DOI:10.6041 / j.issn.1000-1298.2020.02.025 Different training methods are used to migrate the VGG-16 model to corn disease images. Transfer learning reduces the data quantity requirements, but the gap between the source data set and the corn disease samples is large. Simple parameter fine-tuning is difficult to achieve better results. Adding distribution difference measurement methods can improve model performance. Summary of the invention

[0005] In view of the problems existing in the prior art, the present invention proposes a chlorophyll content estimation method based on optical characteristic parameter inversion.

[0006] Technical solution:

[0007] The chlorophyll content estimation method based on optical characteristic parameter inversion includes the following steps:

[0008] (1) Construct a blade model based on a surface light source to simulate photon transmission and obtain a simulated spectral image. The simulation method makes the acquired source domain samples more similar to the measured sample distribution, and the surface light source can obtain more spectral and image information than the point light source method;

[0009] (2) Input the simulated spectral image into the convolutional neural network for training to obtain a pre-trained model. Deep learning methods can extract more feature information than traditional methods;

[0010] (3) Based on the pre-trained model, MMD migration is performed on the measured spectral data. Compared with model migration, the MMD migration method can better improve the model performance, and the linear method reduces the computational burden without affecting the model performance;

[0011] (4) The depth features obtained by MMD migration are used to estimate the chlorophyll content of leaves. Compared with traditional spectral methods, the detection effect is greatly improved.

[0012] In step (1), a leaf model based on a surface light source is constructed to simulate photon transmission. The specific steps are as follows:

[0013] (1.1) Photon emission: Photons are emitted vertically from random coordinate positions in the plane to the tissue surface, and the photons move at random steps and then transfer to (1.2);

[0014] (1.2) Calculate the step length and direction based on the absorption coefficient ua and the scattering coefficient us, and perform out-of-bounds judgment (1.3);

[0015] (1.3) Determine whether the photon has crossed the boundary. If so, proceed to step (1.4b); otherwise, proceed to step (1.4a);

[0016] (1.4a) Determine the next position of the photon based on the current position, step size and direction of the photon, and proceed to (1.5);

[0017] (1.4b) Reflection and transmission at the boundary: When a photon reaches the upper and lower surfaces of the tissue, if it is reflected back into the tissue by the boundary, proceed to (1.5); otherwise, if the photon escapes from the tissue, the tracking of the photon is stopped;

[0018] (1.5) Determine whether the photon weight is too small, if yes, go to (1.6), otherwise go to (1.2);

[0019] (1.6) Play Russian roulette: Generate a random number. If it is greater than the threshold, the photon is considered dead. Otherwise, the photon weight is updated and transferred to (1.2).

[0020] (1.7) Take a large number of photons and different leaf thicknesses d, and different optical characteristic parameter combinations (d, ua, us); repeat steps (1.1)-(1.6) to obtain multiple simulated leaf spectrum images. (1.5) Determine whether the photon weight is too small, if so, proceed to (1.6), otherwise proceed to (1.2);

[0021] In step (2), the pre-trained model is obtained. The specific steps are as follows:

[0022] 10,000 images are extracted from the simulated leaf spectral images as training sets and 2,000 images are extracted as test sets. The training sets are input into the convolutional neural network for training. The network parameters are adjusted according to the test set results to obtain a stable pre-training model. The pre-training network structure includes:

[0023] The first layer is a convolutional layer with a convolution kernel size of 2×2, 10 channels, and a stride of 1.

[0024] The second layer is a convolutional layer with a convolution kernel size of 3×3, 20 channels, and a stride of 1.

[0025] The third layer is the average pooling layer, with a pooling window size of 2×2 and a step size of 1;

[0026] The fourth layer is a convolutional layer with a convolution kernel size of 2×2, 20 channels, and a stride of 1.

[0027] The fully connected layer FC1 reduces the dimension of the output of the 4th layer network to 1×50;

[0028] The last layer is the network output layer.

[0029] In step (3), MMD migration is performed. The specific steps are as follows:

[0030] 3000 leaf spectral images were obtained through actual measurement, and the training set and test set were divided into a ratio of 4:1. The measured data training set was input into the pre-trained model obtained in step (2), and the distribution matching metric was added after the fully connected layer of the network. The cross entropy loss and MMD metric loss together constituted the network loss function. The back propagation algorithm was used to adjust the parameters of the neural network to complete the migration of the pre-trained model on the measured data. The transfer learning method shortened the convergence time of the model and solved the problem that a large amount of measured data must be obtained to improve the model performance. The introduction of the MMD idea made the migrated network have better results.

[0031] In step (4), the leaf chlorophyll content is estimated, and the specific steps are as follows:

[0032] The measured leaf spectral image is input into the convolutional neural network migrated in step (3) to obtain its fully connected layer features, that is, the deep features combining the spectrum and light parameters. The partial least squares regression method is used to take this feature as input to construct a chlorophyll content regression model to estimate the chlorophyll content of the leaves.

[0033] Beneficial effects of the present invention

[0034] (1) Compared with direct training with measured data, the transfer learning method shortens the model convergence time, solves the problem of poor network model effect caused by insufficient measured data, and improves the accuracy of network recognition. The optical parameter inversion method of green radish leaves based on MMD transfer has better performance than the ordinary model transfer method.

[0035] (2) The data obtained by the Monte Carlo simulation method based on the surface light source has similar characteristic distribution to the measured data, which reduces the number of samples actually required and reduces the human and material costs caused by obtaining a large number of measured samples. Moreover, compared with point light sources, surface light sources can obtain more image and spectral information, further improving the robustness of the model.

[0036] (3) The deep optical parameter features obtained by transfer learning method were used to establish a regression model for the chlorophyll content of Pothos sinensis. The results were better than those directly correlating the chlorophyll content with hyperspectral images and optical parameters, which proved the effectiveness and necessity of the optical parameter inversion method.

[0037] (4) In the maximum mean difference (MMD) migration calculation, this application takes into account the distribution differences between the simulation data and the measured data and adopts a linear embedding method to reduce the computational burden while ensuring that the samples between the fields are fully adapted. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 Schematic diagram of obtaining a hyperspectral image of a green radish leaf using the HSI-VNIR-0001 hyperspectral imaging system in the embodiment

[0039] Figure 2 The simulated photon distribution diagram in the embodiment

[0040] Figure 3 This is the network structure diagram of the pre-trained network model in the embodiment

[0041] Figure 4 The training loss curve of the convolutional neural network in the embodiment is

[0042] FIG5(a) is a high-spectral original image in the range of 400 to 900 nm in the embodiment

[0043] FIG5( b ) is a hyperspectral image after SG smoothing preprocessing in the embodiment.

[0044] FIG. 6( a ) is a prediction result of a prediction model constructed using optical parameters in an embodiment.

[0045] FIG. 6( b ) is a prediction result of constructing a prediction model using hyperspectral data screened by SPA in the embodiment

[0046] FIG6(c) is a prediction result of constructing a prediction model using the deep features obtained by migration in the embodiment. DETAILED DESCRIPTION

[0047] The present invention is further described below in conjunction with embodiments to verify the beneficial effects of the solution, but the protection scope of the present invention is not limited thereto:

[0048] 1 Experimental data acquisition

[0049] 1.1 Acquisition of measured data

[0050] 1.1.1 Test samples

[0051] The experimental object is the long-vine green radish. 20 pots of green radish with similar growth were selected as samples. Ten leaves were evenly collected from all sides of each pot of green radish plant, totaling 200 green radish leaves. In order to avoid water loss in the leaves or decomposition of chlorophyll under high temperature and light over a long period of time, experimental data collection was carried out immediately after the leaves were picked.

[0052] 1.1.2 Hyperspectral data collection of green radish leaves

[0053] The HSI-VNIR-0001 hyperspectral imaging system is used to obtain the hyperspectral image of the leaves of the green radish. Figure 1 As shown in FIG. 1 , the imaging system mainly includes a CCD camera 1, a spectrometer 2, an imaging lens 3, a light source 4, a mobile stage 5, a sample 6, a computer, and supporting spectral imaging and analysis software. The spectrum acquisition wavelength range is 375.54 to 1033.87 nm. After eliminating the front and rear noise bands, the hyperspectral effective wavelength range is set to 400 to 900 nm, with a total of 468 bands.

[0054] 1.1.3 Data collection of integrating sphere of green radish leaves

[0055] The double integrating sphere system of Shanghai Wuling Optoelectronics Technology Co., Ltd. is used to collect the reflectivity and transmittance of the leaves of the green radish. The double integrating sphere system is mainly composed of a transmission sphere, a reflection sphere, an optical fiber, a light source, a spectrometer, a computer and corresponding supporting software. After obtaining the reflectivity and transmittance of the leaves by the double integrating sphere system, the IAD algorithm is used to determine the optical characteristic parameters of the green radish leaves through iterative calculation, and finally the absorption coefficient μ of the green radish leaves is obtained. a (0.01~6mm -1 ) and the scattering coefficient μ s (0.9~150mm-1 ).

[0056] 1.1.4 Collection of chlorophyll values ​​of green radish leaves

[0057] The chlorophyll content data of the green radish samples were collected using a portable SPAD-502 chlorophyll meter. To avoid measurement errors in the instrument itself, the device was calibrated before each measurement. After the device was calibrated, the leaves of the green radish plants with normal growth and development were selected using the device to measure a 2×2 (number of pixels) rectangular area. The same area was measured three times and the average value was taken as the chlorophyll value of the area.

[0058] 1.2 Monte Carlo simulation of single-layer flat plate model based on surface light source

[0059] The Monte Carlo method is a simulated sampling or statistical test method. Its basic idea is to first simulate a random process or probability model similar to the actual situation through a series of random numbers, then perform sampling experiments to calculate relevant parameters, and finally obtain an approximate solution. Practice has shown that the MC method can most effectively describe the transmission process of photons in tissues, which is basically consistent with the actual measured parameters. Therefore, it has become a non-experimental standard for light transmission in biological tissues.

[0060] Reference to the process of simulating photon transmission using the Monte Carlo method. A light beam can be viewed as photons one by one. First, when a photon is emitted, the photon will move randomly by a step length, and then absorb and scatter with tiny particles in the tissue. Scattering will change the direction of movement of the photon, and absorption will reduce the weight of the photon. Then the photon continues to move and absorbs and scatters. When the photon escapes from the surface of the tissue, the photon escaping from the upper surface forms diffuse reflected light, and the photon escaping from the lower surface forms transmitted light. When the weight of the photon is small enough, it will be absorbed by the tissue. Finally, repeat all the photons in this light beam to complete the Monte Carlo simulation of the entire light beam.

[0061] Inversion of Apple Optical Characteristic Parameters Based on Model Migration [J]. Transactions of the Chinese Society of Agricultural Machinery, 2020. A double-layer flat plate model of apple under a point light source is constructed to simplify the calculation of complex models. However, the application of point light sources has great limitations. This paper uses a surface light source to construct a Monte Carlo flat plate model to simulate the most realistic photon transmission process under a parallel light source. Because the leaf structure is very similar to the single-layer flat plate structure, a single-layer flat plate structure based on a surface light source is finally used to construct a photon transmission model of green radish leaves. The simulation experiment parameters are set according to the optical characteristic parameters measured by the integrating sphere experiment, and the leaf thickness d is set to 25 values ​​covering the entire thickness of the measured samples, and μ a and μ s Take 7 and 8 uniformly distributed values ​​of the measured absorption coefficient and scattering coefficient respectively. This paper uses 1*10 6The whole process is simulated by using photons, and the statistical distribution of photons on the surface of leaf tissue is obtained as follows: Figure 2 shown. Figure 2 From left to right in the figure are the simulated photon distribution diagrams under point light source and surface light source respectively. It can be seen that compared with the point light source image, the surface light source photon distribution diagram is brighter and the photon coverage area is wider, which to a certain extent explains that more image and spectral information can be obtained by using surface light source data.

[0062] 2 Experimental methods

[0063] This paper proposes a method for inverting the optical characteristic parameters of green radish leaves based on transfer learning, and constructs a prediction model for the chlorophyll content of green radish according to the inversion results of optical parameters. First, a Monte Carlo single-layer plate model under a surface light source is constructed based on the measured data of green radish leaves, and the motion trajectory of photons in the leaves is simulated to obtain a large amount of noise-free ideal spectral data. The simulated spectral data is then trained using a convolutional neural network to obtain a pre-trained model. Based on the trained model, the maximum mean difference between the simulated data and the measured data is added to the loss function of the neural network, and the network is fine-tuned using the measured data to obtain a better inversion accuracy than directly training the neural network and model migration. Finally, the optical characteristic parameters are inverted using hyperspectral and migrated models, and a prediction model for the chlorophyll content of green radish is established. The model is compared with the optical parameters, hyperspectral data and chlorophyll content directly to verify the necessity of inverting optical characteristic parameters.

[0064] 2.1 Convolutional Neural Network Model

[0065] Convolutional neural networks have stronger feature extraction and expression capabilities than traditional methods due to their multi-layer convolution, weight sharing, rotation invariance and other characteristics. They perform outstandingly in image recognition and classification tasks. They are generally composed of input layers, convolution layers, pooling layers, fully connected layers and output layers. Convolutional neural networks can directly input the original image into the network for feature extraction. Convolution layers and pooling layers are important links in extracting image features. The output layer combines multi-dimensional vectors into one-dimensional vector outputs to achieve image classification and recognition.

[0066] This paper selects convolutional neural network for pre-training. The network structure is as follows Figure 3 As shown in the figure, the network model consists of three convolutional layers, one pooling layer and one fully connected layer. The network input is a 30×30 simulated leaf brightness distribution picture. The first layer is a convolutional layer with a convolution kernel size of 2×2, 10 channels and a step size of 1. The second layer is a convolutional layer with a convolution kernel size of 3×3, 20 channels and a step size of 1. The third layer is a pooling layer with a pooling window size of 2×2 and a step size of 1. The fourth layer is a convolutional layer with a convolution kernel size of 2×2, 20 channels and a step size of 1. The fully connected layer FC1 reduces the output of the fourth layer network to 1×50, and finally the optical parameter classification is realized through the Softmax layer.

[0067] 2.2 Transfer Learning

[0068] Transfer learning refers to a learning process that uses the similarities between data, tasks, or models to apply the model learned in the source domain to the target domain. The convolutional layer in the pre-trained model can already extract features from the source domain samples, and adding a fully connected layer can achieve better classification results. The general model transfer method only needs to change the final fully connected layer based on the pre-trained model and apply it to the target dataset, and then fine-tune the network to achieve transfer learning. However, this method requires that the source domain and target domain images have the same distribution, ignoring the differences between the two types of data. If the distribution differences between the two are added to the model training, the migration ability of the network can be improved.

[0069] This application uses the maximum mean discrepancy (MMD) as a metric in transfer learning. The MMD square distance L of two random variables mmd (F s ,F t ):

[0070]

[0071] In the formula, F s and F t Represent the features extracted from the source domain and the target domain respectively; M, N represent the number of samples in the source domain and the target domain respectively; φ represents the kernel function that maps the original data to the regenerated Hilbert space; represents the characteristics of the i-th sample in the source domain, Represents the characteristics of the jth sample in the target domain.

[0072] This paper introduces the distribution matching method of maximum mean difference based on the pre-training model, adds the distribution matching metric after the fully connected layer of the network, reduces the difference in feature distribution between the measured green radish leaf spectral image and the simulated brightness image, and improves the migration ability of the network. The cross entropy loss function L of the network c for:

[0073] L c =λ1L1+λ2L2

[0074]

[0075] In the formula, L1 and L2 represent two tasks μ a , μ s The cross entropy loss of λ1 and λ2 represent the two tasks μ a , μ s The weight of

[0076] L is the general formula of L1 and L2; m is the total number of samples, n is the number of task classification categories, and x is i represents the learned feature vector; and b j Respectively represent the weight vector and bias term of the j-th category fully connected layer output; and Respectively represent the y i The weights and biases of each sample.

[0077] Combining the above cross entropy loss and distribution matching metric loss, the loss function of the domain adaptation method based on MMD in this paper is:

[0078]

[0079] Where, L c Represents the cross entropy loss function of the network; F s and F t Represent the features extracted from the source domain and the target domain respectively, L mmd (F s ,F t ) represents the MMD square distance between two random variables, and λ represents the weight of the mmd loss.

[0080] 2.3 Detection of chlorophyll content in green radish

[0081] Partial least squares regression (PLSR) is based on the idea of ​​component extraction and can solve the problem of multiple correlations between variables. Based on the optical characteristic parameter inversion method based on transfer learning proposed in this paper, the partial least squares prediction model is constructed using the deep features of optical parameters to obtain the regression results of chlorophyll content and compare them with the conventional hyperspectral analysis method. The coefficient of determination (R) between the measured value and the predicted value is used to calculate the chlorophyll content. 2 ) and root mean square error (RMSE) were used to evaluate the performance of the prediction model. 2 The calculation method is:

[0082]

[0083] In the formula, n represents the number of samples; y i , f i represents the measured value and predicted value of the i-th sample; y m Indicates the average of the actual values.

[0084] The closer the coefficient of determination is to 1, the better the predictive ability of the model.

[0085] The smaller the RMSE, the better the model prediction ability is. Its calculation method is:

[0086]

[0087] In the formula, n represents the number of samples; y i , f i Represents the measured value and predicted value of the i-th sample.

[0088] 3 Experimental results analysis

[0089] 3.1 Experimental Environment

[0090] The experiment was completed in Python 3.7.4 and Tensorflow-GPU 1.13.1 software environments. The CPU used in the hardware environment was AMD Ryzen 5 1600 Six-core Processor with a main frequency of 3.2GHz; the GPU used was NVIDIA GeForce RTX2080 Ti graphics card.

[0091] 3.2 Experimental design

[0092] Through the Monte Carlo simulation model, 12,000 simulated leaf brightness distribution maps were obtained, and the sample data was simply divided into 10,000 training sets and 2,000 test sets. The model parameters were adjusted by analyzing and evaluating the test results, and the batch size was set to 100, the learning rate was 0.0001, and the pre-trained model was obtained by training 2,000 times. The 3,000 hyperspectral images collected in the experiment were divided into training sets and test sets at a ratio of 4:1. Based on the pre-trained network, the maximum mean difference measurement method was introduced for transfer learning to generate a new optical parameter inversion model. In the stage of predicting chlorophyll content, the spectral data of the measured samples were input into the trained model, and the output results of the fully connected layer were obtained as the input of the partial least squares to obtain the chlorophyll content regression results. At the same time, in order to verify the effectiveness of the proposed method, the optical parameters and hyperspectral direct association chlorophyll content modeling were compared with the proposed method.

[0093] 3.3 Results and Analysis

[0094] 3.3.1 Inversion of optical characteristic parameters based on simulation data

[0095] The size of the simulated images obtained by the Monte Carlo method is 100dpi*100dpi, with the center point at (50,50). In order to reduce the amount of calculation and improve the efficiency without affecting the training effect, the simulated images of (21:80,21:80), (21:50,51:80), (21:50,21:50), (51:80,51:80), and (51:80,21:50) are taken as the input of the network respectively. The inversion results are shown in Table 1.

[0096] Table 1 Comparison of inversion results with different selection ranges

[0097]

[0098] As can be seen from Table 1, (21:80, 21:80) and (51:80, 51:80) obtain better inversion results. The difference between the inversion effects of the two is not large, but a larger image selection range will lead to an exponential increase in model parameters and training and testing costs. Considering the practicality of the model, the (51:80, 51:80) area image is finally selected to train the model.

[0099] Before training the network model, the data is normalized to facilitate feature extraction by the convolutional neural network. The linear method is used to normalize the data, and the (51:80, 51:80) of the simulation image is selected for model training. The methods of point division by 10, point division by 100, point division by 1000, and point division by 10000 are used respectively. The optical parameter inversion results of the simulation data are shown in Table 2. It can be seen that the normalization effect of point division by 10000 is better.

[0100] Table 2 Effect of normalization on classification results of pre-trained networks

[0101]

[0102]

[0103] The data is normalized by dividing by 10000, and the normalized result is used as the input of the convolutional neural network to obtain the pre-trained model. The training loss curve is as follows: Figure 4 As shown in the figure, it can be seen that the model gradually stabilizes after 1100 iterations. Since the simulation data is generated under ideal conditions, the numerical distribution is relatively consistent, the model converges quickly during training, and the loss value does not fluctuate significantly as the number of iterations increases. In the final simulation data optical parameter inversion results, the inversion accuracy of the absorption coefficient and scattering coefficient both reached more than 87%.

[0104] 3.3.2 Inversion of optical property parameters based on measured data

[0105] Transfer learning enables the model to have better generalization ability and can effectively solve the overfitting problem. Model transfer is one of the most commonly used transfer methods. Non-transfer learning and model transfer are performed separately, and different parameter learning methods are used for training. The results are shown in Table 3.

[0106] Table 3 Optical parameter inversion accuracy of different training methods

[0107]

[0108] As can be seen from Table 3, the model with the lowest test accuracy is the one that only trains the fully connected layer. The reason is that although the convolutional layer of the pre-trained network has the ability to extract image features, the distribution of simulated data and measured data is not exactly the same. Only training to change the fully connected layer cannot achieve the ideal effect, while training to change all parameters can improve model performance. Therefore, the method of training all layers is finally adopted for optical parameter inversion.

[0109] The MMD measurement method is introduced on the basis of the model transfer method. The inversion results based on the MMD transfer method are: absorption coefficient μa 87.55%, scattering coefficient μs 86.67%. Compared with non-transfer learning and simple model transfer, the test accuracy is significantly improved.

[0110] From the above results, we can see that the non-transfer learning method has the lowest test accuracy. This is because the measured training data is small, and the loss value fluctuates with the increase in the number of training times, resulting in unsatisfactory final training results. Both transfer learning methods have higher accuracy than non-transfer learning and greatly shorten the model convergence time. Because there are still certain differences between the simulated data and the measured spectral data, the ideal effect cannot be achieved only through model migration, while the MMD-based migration method can significantly improve the test accuracy.

[0111] 3.3.3 Estimation of chlorophyll content

[0112] SG smoothing preprocessing has the best effect on the hyperspectral data of green radish leaves in the 400-900nm band, and can effectively eliminate the noise information in the original hyperspectral image. Therefore, the SG smoothing method is used to preprocess the hyperspectral data. Figure 5(a) is the original hyperspectral image in the 400-900nm range, and the data noise is large; Figure 5(b) is the hyperspectral image after SG smoothing preprocessing. It can be seen that this method effectively removes data noise and ensures the integrity of the data, providing a good basis for the extraction of characteristic bands. Then the continuous projection algorithm SPA is used to select the characteristic bands. Finally, the number of characteristic wavelength variables selected by the SPA algorithm is 8, which are 59, 77, 97, 128, 167, 305, 348, and 420 respectively.

[0113] In order to verify the necessity of the optical parameter inversion method, the chlorophyll content prediction model was constructed using optical parameters, SPA-screened hyperspectral data, and the depth characteristics obtained by migration. The results are shown in Figure 6. The results of regressing the chlorophyll content using optical parameters are: The regression correlation coefficient R 2 is 0.8842, and the root mean square error is 1.7686; the result of regression of chlorophyll content using hyperspectral image is: regression correlation coefficient R 2 is 0.869, and the root mean square error is 1.1549. The result of modeling chlorophyll content using the deep light parameter features obtained by the migration method proposed in this paper is: R 2 The value of the deep feature is 0.931, and the root mean square error is 0.8374. It can be seen that the deep feature that combines spectral information and optical parameter information has the best effect. Therefore, the inversion of optical parameters by hyperspectral can provide better detection results for the chlorophyll content of green radish.

[0114] The specific embodiments described herein are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.

Claims

1. A chlorophyll content estimation method based on optical characteristic parameter inversion, characterized in that It includes the following steps: (1) Construct a blade model based on a surface light source to simulate photon transmission and obtain a simulated blade spectrum image; (2) Input the simulated blade spectral image into the convolutional neural network for training to obtain a pre-trained model; the specific steps are as follows: extract 10,000 images from the simulated blade spectral image as a training set and 2,000 images as a test set, input the training set into the convolutional neural network for training, adjust the network parameters according to the test set effect and obtain a stable pre-trained model; the network structure of the pre-trained model includes: The first layer is a convolutional layer with a convolution kernel size of 2×2, 10 channels, and a stride of 1. The second layer is a convolutional layer with a convolution kernel size of 3×3, 20 channels, and a stride of 1. The third layer is the average pooling layer, with a pooling window size of 2×2 and a step size of 1; The fourth layer is a convolutional layer with a convolution kernel size of 2×2, 20 channels, and a stride of 1. The fully connected layer FC1 reduces the dimension of the output of the 4th layer network to 1×50; Finally, there is the network output layer; (3) Based on the pre-trained model, the maximum mean difference (MMD) migration is performed on the measured spectral data to obtain the optical parameter inversion model of all training layers; the specific steps of step (3) are as follows: A plurality of leaf spectral images are obtained by actual measurement, and the training set and the test set are divided into two sets in proportion. The measured data training set is input into the pre-trained model obtained in step (2), and a distribution matching metric is added after the fully connected layer of the network. The cross entropy loss and the MMD metric loss together constitute the loss function of the network. The back propagation algorithm is used to adjust the parameters of the network to complete the migration of the pre-trained model on the measured data. The loss function of the network composed of cross entropy loss and MMD metric loss is: Where, L c represents the cross entropy loss function of the network; F s and F t Represent the features extracted from the source domain and the target domain respectively; L mmd (F s ,F t ) represents the MMD square distance between two random variables, and λ represents the weight of the MMD metric loss; (4) The optical parameter inversion model obtained by MMD migration is used to estimate the chlorophyll content of leaves.

2. The method according to claim 1, characterized in that The specific steps of step (1) are as follows: (1.1) Photon emission: Photons are emitted vertically from random positions in the plane coordinates to the tissue surface, and the photons move in random steps and are transferred to (1.2); (1.2) Based on the absorption coefficient u a and the scattering coefficient u s Calculate the step length and direction, and perform out-of-bounds judgment (1.3); (1.3) Determine whether the photon has crossed the boundary. If so, proceed to step (1.4b); otherwise, proceed to step (1.4a); (1.4a) Determine the next position of the photon based on the current position, step size and direction of the photon, and proceed to (1.5); (1.4b) Reflection and transmission at the boundary: When a photon reaches the upper and lower surfaces of the tissue, if it is reflected back into the tissue by the boundary, proceed to (1.5); otherwise, if the photon escapes from the tissue, the tracking of the photon is stopped; (1.5) Determine whether the photon weight is too small, if yes, go to (1.6), otherwise go to (1.2); (1.6) Play Russian roulette: Generate a random number. If it is greater than the threshold, the photon is considered dead. Otherwise, the photon weight is updated and transferred to (1.2). (1.7) Take a large number of photons and different leaf thickness d, different optical characteristic parameter combinations (d, u a ,u s ); Repeat steps (1.1)-(1.6) to obtain multiple simulated leaf spectral images.

3. The method according to claim 1, characterized in that Cross entropy loss function L c : L c =λ1L1+λ2L2 In the formula, L1 and L2 represent two tasks μ a , μ s The cross entropy loss of λ1 and λ2 represent the two tasks μ a , μ s The weight of L is the general formula of L1 and L2; m is the total number of samples, n is the number of task classification categories, and x is i represents the learned feature vector; and b j Respectively represent the weight vector and bias term of the j-th category fully connected layer output; and b yi Respectively represent the y i The weights and biases of each sample.

4. The method according to claim 1, characterized in that MMD squared distance L between two random variables mmd (F s ,F t ): In the formula, F s and F t Represent the features extracted from the source domain and the target domain respectively; M, N represent the number of samples in the source domain and the target domain respectively; φ represents the kernel function that maps the original data to the regenerated Hilbert space; represents the characteristics of the i-th sample in the source domain, Represents the characteristics of the jth sample in the target domain.

5. The method according to claim 1, characterized in that The specific steps of step (4) are as follows: Input the spectral data of the leaf to be tested into the optical parameter inversion model obtained in step (3) to obtain the output result of its fully connected layer, and use the output result as the input of the partial least squares prediction model to obtain the chlorophyll content regression result; the partial least squares prediction model is constructed by the deep features of the optical parameters, and the output is the chlorophyll content value.

Images