Insect type identification method based on multi-dimensional radar electromagnetic scattering characteristic sequence

Abstract

The invention relates to an insect type identification method based on a multi-dimensional radar electromagnetic scattering feature sequence, and aims to accurately identify insect types through electromagnetic scattering features of radar echo signals in combination with a deep learning technology. According to the invention, a multi-dimensional feature sequence is constructed based on a fully-polarized radar scattering matrix, and feature learning is carried out by using an insect identification neural network. And the insect identification neural network extracts spatial features of different scales through multi-scale convolution, updates a time dependency relationship in an effective capture feature sequence through a state input based on a time sequence, and finally identifies insect types based on the multi-scale spatial features and the time dependency thereof. According to the technical scheme, the accuracy and robustness of insect type identification are effectively improved, the average accurate identification rate of nine kinds of migratory flying pests in China can reach 98.1%, and the method has wide application prospects and is particularly suitable for the fields of insect migratory flying monitoring, agricultural pest control and ecological research.

Description

Technical Field

[0001] The present invention relates to the technical field of insect identification based on radar data, and in particular to an insect species identification method based on a multi-dimensional radar electromagnetic scattering feature sequence. Background Art

[0002] Traditional insect species identification methods usually rely on the biological characteristics of insects, such as weight, body length and other parameters, but these methods are limited by environmental factors and the diversity of insect samples, and have poor accuracy and robustness.

[0003] In recent years, with the development of radar technology, insect identification based on radar echo signals has become an emerging research direction. The radar echo signals of insects contain rich electromagnetic scattering characteristics, which are closely related to the body parameters and movement characteristics of insects. Therefore, by collecting and analyzing the radar echo signals of insects, a new technical means can be provided for the automatic identification of insect species. Applicant patents CN118656739A, CN118839257A and CN118656678A respectively study random forest, XGBoost and stack generalization insect identification methods based on radar hash features, and prove the effectiveness of multidimensional radar electromagnetic scattering features in identifying insects. However, the current radar insect identification methods mainly rely on multidimensional radar electromagnetic scattering feature extraction and traditional classification algorithms. It is often impossible to fully utilize the time series dependency and multi-scale feature information contained in the radar signal, and the recognition accuracy is limited. Therefore, a method for insect species identification based on multidimensional radar electromagnetic scattering feature sequence is proposed, which can make full use of the spatial, temporal and multi-scale characteristic information of multidimensional radar electromagnetic scattering features in insect echo signals, thereby significantly improving the accuracy and robustness of insect species identification. Summary of the invention

[0004] In order to solve the above technical problem or at least partially solve the above technical problem, the present invention provides an insect species identification method based on a multi-dimensional radar electromagnetic scattering feature sequence.

[0005] In a first aspect, the present invention provides an insect species identification method based on a multi-dimensional radar electromagnetic scattering feature sequence, comprising: The full polarization radar scattering matrix of insects is collected by full polarization insect radar, and the multi-dimensional electromagnetic scattering characteristics in the insect radar echo signal are obtained based on the full polarization radar scattering matrix of insects at time t. , multi-dimensional electromagnetic scattering characteristics Forming a multi-dimensional electromagnetic scattering feature sequence in chronological order; Among them, the full polarization radar scattering matrix of insects is expressed as: ; in, , , and denote the square roots of the radar cross sections for horizontal horizontal polarization, horizontal vertical polarization, vertical horizontal polarization, and vertical vertical polarization, respectively; , and are horizontal vertical polarization, vertical horizontal polarization and the relative phase of vertical vertical polarization relative to horizontal horizontal polarization respectively; the multi-dimensional electromagnetic scattering characteristics include: polarization pattern characteristics, radar scattering cross section characteristics, scattering matrix difference characteristics and polarization invariant characteristics; The obtained multi-dimensional electromagnetic scattering feature sequence is subjected to data preprocessing, which includes feature normalization processing to map each feature to the interval [0,1]; The preprocessed multidimensional electromagnetic scattering feature sequence is input into a pretrained insect recognition neural network. The insect recognition neural network classifies and recognizes insect species based on the multidimensional electromagnetic scattering feature sequence. The insect recognition neural network includes: a multi-scale feature extraction module. The multi-scale feature extraction module includes: three parallel one-dimensional convolutional layers. The convolution kernel sizes of the one-dimensional convolutional layers are k, 2k and 4k respectively. The output channels of each one-dimensional convolutional layer are the same. Each one-dimensional convolutional layer keeps the length of the output feature unchanged relative to the input through a zero padding operation. The spatial features of different scales extracted by each one-dimensional convolutional layer are combined and spliced ​​to form a comprehensive multi-scale feature representation. The multi-scale feature representation based on the current time t and the hidden state at the previous moment t-1 Iteratively update the hidden state as follows: ; in, is the hidden state at the current time t. The hidden state contains the information represented by the previous multi-scale features. , To obtain The multi-scale feature representation weights are obtained The hidden state weights and Bias; Multi-scale feature representation is mainly controlled by weight, and the hidden state at the previous moment t-1 Seek for assistance ; , To obtain The multi-scale feature representation weights are obtained The hidden state weights and Bias; through the weight control multi-scale feature representation as a supplement, the hidden state of the previous moment t-1 Ask for the Lord ; The multi-scale feature representation weights of the intermediate hidden state, the hidden state weights of the intermediate hidden state, and the bias of the intermediate hidden state are obtained respectively; is the Sigmoid activation function, For the tanh activation function, the initial state update provides an initial hidden state ; The classification module based on the fully connected network is based on Classify and identify insect species.

[0006] Furthermore, the polarization pattern characteristics include: polarization direction Figure 1 Subharmonics and Second Harmonic Factors , , and , polarization pattern average radar cross section , polarization pattern radar cross section first harmonic component , polarization pattern radar cross section second harmonic component , polarization pattern shape parameters and , the relative phase of the horizontal vertical polarization with respect to the horizontal horizontal polarization , the relative phase of the vertical polarization with respect to the horizontal polarization , the calculation process of the polarization pattern characteristics is as follows: The insect polarization pattern measured by the fully polarized insect radar is calculated based on the insect's fully polarized radar scattering matrix: , in, is the normalized effective antenna length of the radar transmitting antenna, , Indicates the polarization direction of the antenna; It is obtained by the following formula: , , , , , , , Two parameters that characterize the shape of the polarization pattern and : , , is the phase difference between the two eigenvalues ​​of the insect's full polarization radar scattering matrix, is the coefficient factor related to the direction of the polarization pattern parallel to the insect body axis, is the coefficient factor related to the direction of the polarization pattern perpendicular to the insect body axis.

[0007] Furthermore, the radar cross-section characteristics include: the maximum value of the scattering cross-section of the insect polarization pattern ,and Vertical radar cross section , and Ratio and average , the calculation process of the radar cross section characteristics is as follows: Obtain the full polarization radar scattering matrix of the insect and calculate the average radar scattering cross section of the polarization pattern, the first and second harmonic components and factors of the radar scattering cross section: , , , , , , , Calculation based on the obtained parameters: Maximum Radar Cross Section of Insect Polarimetric Patterns , and Vertical radar cross section , , .

[0008] Furthermore, the scattering matrix differential feature includes the differential amplitude feature of the insect's full polarization radar scattering matrix Differential Phase Characteristics , the calculation process of the scattering matrix differential feature includes: The fully polarized radar scattering matrix of insects calculates the two eigenvalues ​​of insects and Without loss of generality, assume ,but: , , in, and Respectively and The amplitude, and Respectively and The phase of Differential Amplitude Characteristics of Full Polarimetric Radar Scattering Matrix of Insects Differential Phase Characteristics The calculation formula is as follows: , .

[0009] Furthermore, the polarization pattern characteristic polarization invariant characteristic comprises: two eigenvalues ​​of the insect Graves power matrix and , Determinant ,trace and the Frobenius coefficient ; The calculation process of polarization invariant characteristics includes: Calculate the insect's Graves power matrix based on its full polarization radar scattering matrix: ; in, , , , ; Calculate the two eigenvalues ​​of the Graves power matrix of insects and for: , , Determinant of the Graves power matrix of insects is the square root of the product of two eigenvalues; Traces of the Graves power matrix of insects For two eigenvalues and The sum of: Frobenius coefficient of the Graves power matrix of insects For two eigenvalues and Calculates the square root of the sum of squares.

[0010] Furthermore, the feature normalization uses a linear normalization method to map the value of each electromagnetic scattering feature to the interval [0,1] to ensure that different features have the same range and scale.

[0011] Furthermore, the convolutional layer is followed by a maximum pooling layer with a kernel size of 2, which compresses the input feature sequence and highlights the most significant features, ensuring that key dynamic features in the feature sequence are captured.

[0012] Furthermore, the training process of the neural network for insect recognition includes: constructing a data set based on multi-dimensional electromagnetic scattering features and insect species labels, and dividing the data set into a training set and a validation set in a ratio of 7:3; During the training of the insect recognition neural network, the initial learning rate was set to 0.001, and a segmented learning rate decay strategy was used to reduce the learning rate to 10% of the original value after every 30 training rounds. The cross entropy loss between the predicted insect species and the real insect species was used as the loss function in the training process, and the goal was to minimize the loss function. The parameters of the insect recognition neural network were adjusted using the Adam optimization algorithm. After the training is completed, the insect recognition neural network is verified through a validation set that was not exposed during the insect recognition neural network training process.

[0013] In a second aspect, the present invention provides an insect species identification device based on a multi-dimensional radar electromagnetic scattering feature sequence, comprising: at least one processing unit, the processing unit is connected to a storage unit and a collection unit via a bus unit, the storage unit stores a computer program, and when the computer program is executed by the processing unit, the insect species identification method based on the multi-dimensional radar electromagnetic scattering feature sequence is implemented.

[0014] In a third aspect, the present invention provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the insect species identification method based on a multi-dimensional radar electromagnetic scattering feature sequence.

[0015] The above technical solution provided by the embodiment of the present invention has the following advantages compared with the prior art: The present invention constructs a multidimensional feature sequence by extracting polarization scattering features from insect radar echoes, and uses an insect recognition neural network for feature learning. The insect recognition neural network extracts spatial features of different scales of the multidimensional feature sequence through multi-scale convolution, and effectively captures the time dependency in the feature sequence through state updates based on time series input, and finally identifies insect species based on multi-scale spatial features and their time dependency. The technical solution of the present invention effectively improves the accuracy and robustness of insect species identification, and the average accurate recognition rate of the country's nine major categories of migratory pests can reach 98.1%. This application reduces the dependence on insect biological parameters, and at the same time has strong environmental adaptability and efficient computing performance, and can be widely used in practical scenarios such as insect monitoring and migratory pest control. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0017] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0018] Figure 1 A flowchart of an insect species identification method based on a multi-dimensional radar electromagnetic scattering feature sequence provided by an embodiment of the present invention; Figure 2 A schematic diagram of an insect recognition neural network provided by an embodiment of the present invention; Figure 3 The training and testing accuracy curves provided by the embodiment of the present invention; Figure 4 Confusion matrix for identifying migratory pests of 9 countries provided in the embodiment of the present invention; Figure 5 A schematic diagram of an insect species identification device based on a multi-dimensional radar electromagnetic scattering feature sequence provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0020] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.

[0021] Example 1 like Figure 1 As shown, the technology of the present invention realizes an insect species identification method based on a multi-dimensional radar electromagnetic scattering feature sequence, including: The full polarization radar scattering matrix of insects is collected by full polarization insect radar, and the multi-dimensional electromagnetic scattering characteristics in the insect radar echo signal are obtained based on the full polarization radar scattering matrix of insects at time t. , multi-dimensional electromagnetic scattering characteristics A multi-dimensional electromagnetic scattering feature sequence is formed in chronological order.

[0022] The full polarization radar scattering matrix of the insect after polarization and radar cross section calibration is expressed as: ; in, , , and denote the square roots of the radar cross sections for horizontal horizontal polarization, horizontal vertical polarization, vertical horizontal polarization, and vertical vertical polarization, respectively; , and They are the relative phases of horizontal vertical polarization, vertical horizontal polarization and vertical vertical polarization relative to horizontal horizontal polarization respectively.

[0023] The present invention extracts 22-dimensional electromagnetic scattering features 22-dimensional electromagnetic scattering features for insect species identification Including: 11-dimensional polarization pattern features, 4-dimensional radar scattering cross-section features, 2-dimensional scattering matrix difference features and 5-dimensional polarization invariant features.

[0024] The 11-dimensional polarization pattern features include: Figure 1 Subharmonics and Second Harmonic Factors , , and , polarization pattern average radar cross section , polarization pattern radar cross section first harmonic component , polarization pattern radar cross section second harmonic component , polarization pattern shape parameters and , the relative phase of the horizontal vertical polarization with respect to the horizontal horizontal polarization , the relative phase of the vertical polarization with respect to the horizontal polarization , the calculation process of the polarization pattern characteristics is as follows: The insect polarization pattern measured by the fully polarized insect radar is calculated based on the insect's fully polarized radar scattering matrix: , in, is the normalized effective antenna length of the radar transmitting antenna, , Indicates the polarization direction of the antenna; It is obtained by the following formula: , , , , , , , Two parameters that characterize the shape of the polarization pattern and : , , is the phase difference between the two eigenvalues ​​of the insect's full polarization radar scattering matrix, is the coefficient factor related to the direction of the polarization pattern parallel to the insect body axis, is the coefficient factor related to the direction of the polarization pattern perpendicular to the insect body axis.

[0025] The radar cross-section characteristics include: the maximum value of the scattering cross-section of the insect polarization pattern ,and Vertical radar cross section , and Ratio and average , the calculation process of the radar cross section characteristics is as follows: Obtain the full polarization radar scattering matrix of the insect to calculate the average radar scattering cross section of the polarization pattern, the first and second harmonic components and factors of the radar scattering cross section; Then calculate based on the obtained parameters: Maximum Radar Cross Section of Insect Polarimetric Patterns , and Vertical radar cross section , , .

[0026] The scattering matrix differential feature includes the differential amplitude feature of the insect's full polarization radar scattering matrix Differential Phase Characteristics , the calculation process of the scattering matrix differential feature includes: The fully polarized radar scattering matrix of insects calculates the two eigenvalues ​​of insects and Without loss of generality, assume ,but: , , in, and Respectively and The amplitude, and Respectively and The phase of Differential Amplitude Characteristics of Full Polarimetric Radar Scattering Matrix of Insects Differential Phase Characteristics The calculation formula is as follows: , .

[0027] The polarization pattern feature polarization invariant feature includes: two eigenvalues ​​of the insect's Graves power matrix and , Determinant ,trace and the Frobenius coefficient ; The calculation process of polarization invariant characteristics includes: Calculate the insect's Graves power matrix based on its full polarization radar scattering matrix: ; in, , , , ; Calculate the two eigenvalues ​​of the Graves power matrix of insects and for: , , Determinant of the Graves power matrix of insects is the square root of the product of the two eigenvalues ; Traces of the Graves power matrix of insects For two eigenvalues and of ; Frobenius coefficient of the Graves power matrix of insects For two eigenvalues and Square root of the sum of squares .

[0028] The obtained multidimensional electromagnetic scattering feature sequence is subjected to data preprocessing. The insect recognition neural network proposed in the present invention needs to perform strict data preprocessing on the multidimensional electromagnetic scattering feature sequence before inputting data to ensure that the training and reasoning of the model can obtain efficient and accurate results.

[0029] Description of the data format and multidimensional electromagnetic scattering feature sequence: The data consists of polarization scattering features of radar echo signals. Each piece of data contains 22 electromagnetic scattering features, which are obtained through signal processing and feature extraction algorithms within a radar pulse accumulation period (Coherent Processing Interval, CPI). The sequence length is 100, which represents the feature extraction results of 100 consecutive pulse accumulation periods. In other words, each multidimensional electromagnetic scattering feature sequence sample can be represented as a feature matrix of size 22×100, where each column of the matrix represents a multidimensional electromagnetic scattering feature, and each multidimensional electromagnetic scattering feature contains 22 electromagnetic scattering features.

[0030] In order to eliminate the differences in the magnitude of each feature and improve the convergence speed and training effect of the model, the 22 electromagnetic scattering features are normalized respectively. The normalization adopts the linear normalization method to map the feature value to the [0,1] interval. The normalization formula is as follows: , in, is the normalized eigenvalue, is the original eigenvalue, and The normalization operation is performed on a feature-by-feature basis to ensure that each feature has the same range and scale.

[0031] The preprocessed multidimensional electromagnetic scattering feature sequence is input into a pretrained insect recognition neural network, and the insect recognition neural network classifies and recognizes insect species based on the multidimensional electromagnetic scattering feature sequence.

[0032] like Figure 2 As shown, the insect recognition neural network includes: a multi-scale feature extraction module, the multi-scale feature extraction module includes: three parallel one-dimensional convolutional layers, the convolution kernel sizes of the one-dimensional convolutional layers are k, 2k and 4k respectively, and the output channels of each one-dimensional convolutional layer are the same; wherein each one-dimensional convolutional layer keeps the length of the output feature unchanged relative to the input through a zero padding operation, and the spatial features of different scales extracted by each one-dimensional convolutional layer are combined and concatenated to form a comprehensive multi-scale feature representation; based on the multi-scale feature representation at the current time t and the hidden state at the previous moment t-1 Update the hidden state as follows: ; in, is the hidden state at the current time t. The hidden state contains the information represented by the previous multi-scale features. , To obtain The multi-scale feature representation weights are obtained The hidden state weights and Bias; Multi-scale feature representation is mainly controlled by weight, and the hidden state at the previous moment t-1 Seek for assistance ; , To obtain The multi-scale feature representation weights are obtained The hidden state weights and Bias; through the weight control multi-scale feature representation as a supplement, the hidden state of the previous moment t-1 Ask for the Lord ; The multi-scale feature representation weights of the intermediate hidden state, the hidden state weights of the intermediate hidden state, and the bias of the intermediate hidden state are obtained respectively; is the Sigmoid activation function, For the tanh activation function, the initial state update provides an initial hidden state ; The classification module based on the fully connected network is based on Classify and identify insect species.

[0033] The input is a multidimensional electromagnetic scattering feature sequence, and the dimension of the multidimensional electromagnetic scattering feature sequence is , 22 corresponds to the 22-dimensional polarization scattering characteristics, and the sequence length 100 represents the continuous sampling time of the radar.

[0034] As a multi-scale feature extraction module, three parallel one-dimensional convolutional layers are mainly responsible for extracting multi-scale information in different time windows from the radar multi-dimensional scattering features, and improving the comprehensive modeling capabilities of local and global features. The convolution kernel sizes of the three parallel one-dimensional convolutional layers are 2k, 4k, and 8k, respectively. An example of k value is 1, which is used to extract small-scale, medium-scale, and large-scale features respectively: The convolution kernel size is 2: It mainly captures small-scale local features, such as the rapid changes in the insect body size in the radar signal. The convolution kernel size is 4: Modeling the dynamic characteristics of the medium scale and the dynamic characteristics of the insect's flight posture. The convolution kernel size is 8: Extracting large-scale global trends and revealing the overall change law of migration behavior.

[0035] The spatial features of different scales extracted by each one-dimensional convolutional layer are combined and stitched to form a comprehensive multi-scale feature representation. In the present application, the output channels of each one-dimensional convolutional layer are the same. In an example, the number of output channels of each convolutional layer is fixed to 64. The multi-scale features output by different convolution kernels are spliced ​​in the channel dimension to obtain a unified multi-scale feature representation of 192 channels. The spliced ​​features are dimensionally compressed by a maximum pooling operation. Specifically, the convolutional layer is followed by a maximum pooling layer with a kernel size of 2. The maximum pooling layer compresses the input feature sequence and highlights the most significant features, ensuring that the key dynamic features in the feature sequence are captured while reducing the computational complexity.

[0036] Timing modeling via hidden state updates: Multi-scale feature representation based on the current time t and the hidden state at the previous moment t-1 Update the hidden state as follows: ; in, is the hidden state at the current time t. The hidden state contains the information represented by the previous multi-scale features. , To obtain The multi-scale feature representation weights are obtained The hidden state weights and Bias; Multi-scale feature representation is mainly controlled by weight, and the hidden state at the previous moment t-1 Seek for assistance ; , To obtain The multi-scale feature representation weights are obtained The hidden state weights and Bias; through the weight control multi-scale feature representation as a supplement, the hidden state of the previous moment t-1 Ask for the Lord ; The multi-scale feature representation weights of the intermediate hidden state, the hidden state weights of the intermediate hidden state, and the bias of the intermediate hidden state are obtained respectively; is the Sigmoid activation function, For the tanh activation function, the initial state update provides an initial hidden state ; The significance of time series modeling: Time series modeling can capture the significant dynamic changes in the insect migration process, such as the differences in flight frequency and scattering intensity of different insect species, thereby improving the recognition ability of the model. Helps capture short-term dependencies in a sequence; It helps to capture long-term dependencies in the sequence. It can better capture the dependencies on sequences with long time step distances in various situations.

[0037] The classification module based on the fully connected network is based on Classify and identify insect species. The classification module first uses a fully connected network to convert the 128-dimensional hidden state output of the time series modeling into Mapped to a 64-dimensional feature space, the key features required for classification are further extracted, and the nonlinear representation capability of the features is enhanced through the ReLU activation function. Subsequently, the second-layer fully connected network further maps the 64-dimensional features to the number of insect categories, and the final output is the probability distribution of each insect category. The category with the highest probability can be selected as the recognition result through post-processing.

[0038] In summary, the insect recognition neural network architecture of the present invention extracts spatial features through multi-scale convolution, models temporal dynamics through hidden state update timing, and combines the high-dimensional mapping capability of the classification module to provide an accurate and efficient solution for insect species recognition.

[0039] In order to achieve the above-mentioned purpose, the insect recognition neural network needs to be trained. The training process of the insect recognition neural network includes: constructing a data set based on multi-dimensional electromagnetic scattering characteristics and insect species labels, dividing the data set into a training set and a validation set in a ratio of 7:3; training the insect recognition neural network with the training set. During the training process, the present invention reasonably sets the training parameters to ensure that the insect recognition neural network model can converge efficiently and obtain optimal performance. Including: the initial learning rate is set to 0.001, and a segmented learning rate decay strategy is adopted to reduce the learning rate to 10% of the original every 30 training rounds. The update method is as follows: ,in l =0.1. The batch size is set to 32 to achieve a balance between video memory usage and training efficiency. The total number of training rounds is set to 100 to ensure that the insect recognition neural network can fully learn the data feature distribution. The optimizer selected the Adam algorithm, which has good performance in dynamically adjusting the learning rate and sparse gradient optimization. In addition, in order to measure the deviation between the model prediction results and the true label, the training process uses the cross entropy loss of the predicted insect species and the true insect species as the loss function, minimizes the loss function as the goal, adjusts the parameters of the insect recognition neural network, and the loss function expression is: , in, is the total number of samples, is the total number of categories, is the true label, is the predicted probability.

[0040] During the training process, in order to further optimize the model performance, the present invention has made many improvements to the network architecture. First, the network pruning technology is used to analyze the importance of each layer of network weights and prune redundant weights to reduce the complexity of the model. Secondly, the model quantization technology is used to compress the 32-bit floating-point weights of the insect recognition neural network into 8-bit integers, which not only effectively reduces the storage requirements but also improves the computing efficiency. In terms of hyperparameter tuning, the present invention uses the grid search method to experimentally optimize the key hyperparameters.

[0041] In order to avoid overfitting of the model in the later stage of training, the present invention introduces an early stopping strategy (Early Stopping). Specifically, when the accuracy of the validation set no longer improves within 10 consecutive rounds, the training process automatically terminates. Combined with the segmented learning rate decay and early stopping strategy, the model can avoid overfitting while ensuring sufficient learning, thereby improving generalization performance. The training and optimization process described in the present invention not only ensures the accuracy and stability of the model, but also significantly reduces the deployment complexity and improves efficiency in practical applications.

[0042] In order to verify the effectiveness of the insect species identification method based on multi-dimensional radar electromagnetic scattering feature sequence of the present invention, a database construction experiment was carried out on the nine major migratory pests in the country using a high-resolution fully polarized insect radar. The species and body size parameters of the nine major migratory pests are shown in Table 1.

[0043] Table 1 Pest species and body size parameters

[0044] Under experimental conditions, the fully polarized insect radar accurately collects echoes of each pest in flight state to obtain a multi-dimensional polarization scattering feature sequence. The time length of each group of multi-dimensional polarization scattering feature sequences is 100 time steps, and each time step corresponds to the scattering characteristic extraction result of the radar under a single coherent processing interval. Through the method of the present invention, the multi-dimensional polarization scattering feature sequence is input into the insect recognition neural network for training and testing after normalization.

[0045] During the experiment, the datasets of the above nine types of pests were randomly divided into training set and validation set in a ratio of 7:3, and the specific number of pest samples was 7171 and 3074 respectively. During the training process, the Adam optimizer was used to optimize the parameters of the insect recognition neural network. The initial value of the learning rate was 0.001, and it decreased proportionally every 30 training rounds. An exemplary learning rate reduction factor was 0.1. A total of 100 rounds of training were conducted in the experiment, and the performance of the insect recognition neural network was evaluated by the cross entropy loss function.

[0046] Performance evaluation based on the validation set: The insect recognition neural network showed good convergence and stability in 100 training rounds. The accuracy curves of the training set and the validation set are shown in Figure 4 As shown. Figure 3 As shown in the figure, the training and test accuracy curves show that the training and test accuracy stabilizes at 98.1% after about 50 rounds of training. This result shows that the insect recognition neural network not only achieves a good fit on the training data, but also shows excellent generalization ability for unseen test data. Figure 4 In the validation set, the identification confusion matrix of 9 migratory pests was plotted, from which the classification effect of each type of pest can be intuitively observed. The values ​​on the diagonal of the confusion matrix represent the number of correct classifications for each pest type, indicating that the classification accuracy of the insect identification neural network is extremely high in most categories. The off-diagonal elements represent classification errors, and their values ​​are low, indicating that there is less confusion between different pest types.

[0047] Example 2 See also Figure 5As shown, an embodiment of the present invention provides an insect species identification device based on a multi-dimensional radar electromagnetic scattering feature sequence, comprising: at least one processing unit, the processing unit is connected to a storage unit via a bus unit, the storage unit is a computer-readable storage medium, and can be used to store software programs, computer executable programs, and modules, such as the software programs, computer executable programs, and modules corresponding to an insect species identification method based on a multi-dimensional radar electromagnetic scattering feature sequence in an embodiment of the present invention. The processing unit implements the above-mentioned insect species identification method based on a multi-dimensional radar electromagnetic scattering feature sequence by running the software programs, computer executable programs, and modules stored in the storage unit, comprising: The full polarization radar scattering matrix of insects is collected by full polarization insect radar, and the multi-dimensional electromagnetic scattering characteristics in the insect radar echo signal are obtained based on the full polarization radar scattering matrix of insects at time t. , multi-dimensional electromagnetic scattering characteristics Forming a multi-dimensional electromagnetic scattering feature sequence in chronological order; Among them, the full polarization radar scattering matrix of insects is expressed as: ; in, , , and denote the square roots of the radar cross sections for horizontal horizontal polarization, horizontal vertical polarization, vertical horizontal polarization, and vertical vertical polarization, respectively; , and are horizontal vertical polarization, vertical horizontal polarization and the relative phase of vertical vertical polarization relative to horizontal horizontal polarization respectively; the multi-dimensional electromagnetic scattering characteristics include: polarization pattern characteristics, radar scattering cross section characteristics, scattering matrix difference characteristics and polarization invariant characteristics; The obtained multi-dimensional electromagnetic scattering feature sequence is subjected to data preprocessing, which includes feature normalization processing to map each feature to the interval [0,1]; The preprocessed multidimensional electromagnetic scattering feature sequence is input into a pretrained insect recognition neural network. The insect recognition neural network classifies and recognizes insect species based on the multidimensional electromagnetic scattering feature sequence. The insect recognition neural network includes: a multi-scale feature extraction module. The multi-scale feature extraction module includes: three parallel one-dimensional convolutional layers. The convolution kernel sizes of the one-dimensional convolutional layers are k, 2k and 4k respectively. k is a positive integer. The output channels of each one-dimensional convolutional layer are the same. Each one-dimensional convolutional layer keeps the length of the output feature unchanged relative to the input through a zero padding operation. The spatial features of different scales extracted by each one-dimensional convolutional layer are combined and concatenated to form a comprehensive multi-scale feature representation. The multi-scale feature representation based on the current time t and the hidden state at the previous moment t-1 Iteratively update the hidden state as follows: ; in, is the hidden state at the current time t. The hidden state contains the information represented by the previous multi-scale features. , To obtain The multi-scale feature representation weights are obtained The hidden state weights and Bias; Multi-scale feature representation is mainly controlled by weight, and the hidden state at the previous moment t-1 Seek for assistance ; , To obtain The multi-scale feature representation weights are obtained The hidden state weights and Bias; through the weight control multi-scale feature representation as a supplement, the hidden state of the previous moment t-1 Ask for the Lord ; The multi-scale feature representation weights of the intermediate hidden state, the hidden state weights of the intermediate hidden state, and the bias of the intermediate hidden state are obtained respectively; is the Sigmoid activation function, For the tanh activation function, the initial state update provides an initial hidden state ; The classification module based on the fully connected network is based on Classify and identify insect species.

[0048] Of course, the computer program stored in the storage unit of the insect species identification device based on a multi-dimensional radar electromagnetic scattering feature sequence provided in an embodiment of the present invention is not limited to the method operations described above, but can also execute related operations in the insect species identification method based on a multi-dimensional radar electromagnetic scattering feature sequence provided in any embodiment of the present invention.

[0049] Example 3 An embodiment of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program. When the computer program is executed, the method for identifying insect species based on a multi-dimensional radar electromagnetic scattering feature sequence is implemented, including: The full polarization radar scattering matrix of insects is collected by full polarization insect radar, and the multi-dimensional electromagnetic scattering characteristics in the insect radar echo signal are obtained based on the full polarization radar scattering matrix of insects at time t. , multi-dimensional electromagnetic scattering characteristics Forming a multi-dimensional electromagnetic scattering feature sequence in chronological order; Among them, the full polarization radar scattering matrix of insects is expressed as: ; in, , , and denote the square roots of the radar cross sections for horizontal horizontal polarization, horizontal vertical polarization, vertical horizontal polarization, and vertical vertical polarization, respectively; , and are horizontal vertical polarization, vertical horizontal polarization and the relative phase of vertical vertical polarization relative to horizontal horizontal polarization respectively; the multi-dimensional electromagnetic scattering characteristics include: polarization pattern characteristics, radar scattering cross section characteristics, scattering matrix difference characteristics and polarization invariant characteristics; The obtained multi-dimensional electromagnetic scattering feature sequence is subjected to data preprocessing, which includes feature normalization processing to map each feature to the interval [0,1]; The preprocessed multidimensional electromagnetic scattering feature sequence is input into a pretrained insect recognition neural network. The insect recognition neural network classifies and recognizes insect species based on the multidimensional electromagnetic scattering feature sequence. The insect recognition neural network includes: a multi-scale feature extraction module. The multi-scale feature extraction module includes: three parallel one-dimensional convolutional layers. The convolution kernel sizes of the one-dimensional convolutional layers are k, 2k and 4k respectively. k is a positive integer. The output channels of each one-dimensional convolutional layer are the same. Each one-dimensional convolutional layer keeps the length of the output feature unchanged relative to the input through a zero padding operation. The spatial features of different scales extracted by each one-dimensional convolutional layer are combined and concatenated to form a comprehensive multi-scale feature representation. The multi-scale feature representation based on the current time t and the hidden state at the previous moment t-1 Iteratively update the hidden state as follows: ; in, is the hidden state at the current time t. The hidden state contains the information represented by the previous multi-scale features. , To obtain The multi-scale feature representation weights are obtained The hidden state weights and Bias; Multi-scale feature representation is mainly controlled by weight, and the hidden state at the previous moment t-1 Seek for assistance ; , To obtain The multi-scale feature representation weights are obtained The hidden state weights and Bias; through the weight control multi-scale feature representation as a supplement, the hidden state of the previous moment t-1 Ask for the Lord ; The multi-scale feature representation weights of the intermediate hidden state, the hidden state weights of the intermediate hidden state, and the bias of the intermediate hidden state are obtained respectively; is the Sigmoid activation function, For the tanh activation function, the initial state update provides an initial hidden state ; The classification module based on the fully connected network is based on Classify and identify insect species.

[0050] A computer-readable storage medium provided in an embodiment of the present invention stores a computer program which is not limited to the method operations described above, but can also execute related operations in a method for insect species identification based on a multi-dimensional radar electromagnetic scattering feature sequence provided in any embodiment of the present invention.

[0051] In the embodiments provided by the present invention, it should be understood that the disclosed structures and methods can be implemented in other ways. For example, the structural embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, structures or units, which can be electrical, mechanical or other forms.

[0052] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0053] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0054] The foregoing is merely a specific embodiment of the present invention, which enables those skilled in the art to understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A method for insect species identification based on multi-dimensional radar electromagnetic scattering feature sequence, characterized in that: include: The full polarization radar scattering matrix of insects is collected by full polarization insect radar, and the multi-dimensional electromagnetic scattering characteristics in the insect radar echo signal are obtained based on the full polarization radar scattering matrix of insects at time t. , multi-dimensional electromagnetic scattering characteristics Forming a multi-dimensional electromagnetic scattering feature sequence in chronological order; Among them, the full polarization radar scattering matrix of insects is expressed as: ； in, , , and denote the square roots of the radar cross sections for horizontal horizontal polarization, horizontal vertical polarization, vertical horizontal polarization, and vertical vertical polarization, respectively; , and are horizontal vertical polarization, vertical horizontal polarization and the relative phase of vertical vertical polarization relative to horizontal horizontal polarization respectively; the multi-dimensional electromagnetic scattering characteristics include: polarization pattern characteristics, radar scattering cross section characteristics, scattering matrix difference characteristics and polarization invariant characteristics; The obtained multi-dimensional electromagnetic scattering feature sequence is subjected to data preprocessing, which includes feature normalization processing to map each feature to the interval [0,1]; The preprocessed multidimensional electromagnetic scattering feature sequence is input into a pretrained insect recognition neural network. The insect recognition neural network classifies and recognizes insect species based on the multidimensional electromagnetic scattering feature sequence. The insect recognition neural network includes: a multi-scale feature extraction module. The multi-scale feature extraction module includes: three parallel one-dimensional convolutional layers. The convolution kernel sizes of the one-dimensional convolutional layers are k, 2k and 4k respectively. k is a positive integer. The output channels of each one-dimensional convolutional layer are the same. Each one-dimensional convolutional layer keeps the length of the output feature unchanged relative to the input through a zero padding operation. The spatial features of different scales extracted by each one-dimensional convolutional layer are combined and concatenated to form a comprehensive multi-scale feature representation. The multi-scale feature representation based on the current time t and the hidden state at the previous moment t-1 Iteratively update the hidden state as follows: ； in, is the hidden state at the current time t. The hidden state contains the information represented by the previous multi-scale features. , To obtain The multi-scale feature representation weights are obtained The hidden state weights and Bias; Multi-scale feature representation is mainly controlled by weight, and the hidden state at the previous moment t-1 Seek for assistance ; , To obtain The multi-scale feature representation weights are obtained The hidden state weights and Bias; through the weight control multi-scale feature representation as a supplement, the hidden state of the previous moment t-1 Ask for the Lord ; The multi-scale feature representation weights of the intermediate hidden state, the hidden state weights of the intermediate hidden state, and the bias of the intermediate hidden state are obtained respectively; is the Sigmoid activation function, For the tanh activation function, the initial state update provides an initial hidden state ; The classification module based on the fully connected network is based on Classify and identify insect species.

2. The insect species identification method based on multi-dimensional radar electromagnetic scattering feature sequence according to claim 1 is characterized in that: The polarization pattern characteristics include: polarization pattern first harmonic and second harmonic factors , , and , polarization pattern average radar cross section , polarization pattern radar cross section first harmonic component , polarization pattern radar cross section second harmonic component , polarization pattern shape parameters and , the relative phase of the horizontal vertical polarization with respect to the horizontal horizontal polarization , the relative phase of the vertical polarization with respect to the horizontal polarization , the calculation process of the polarization pattern characteristics is as follows: The insect polarization pattern measured by the fully polarized insect radar is calculated based on the insect's fully polarized radar scattering matrix: ， in, is the normalized effective antenna length of the radar transmitting antenna, , Indicates the polarization direction of the antenna; It is obtained by the following formula: ， ， ， ， ， ， ， Two parameters that characterize the shape of the polarization pattern and : ， ， is the phase difference between the two eigenvalues ​​of the insect's full polarization radar scattering matrix, is the coefficient factor related to the direction of the polarization pattern parallel to the insect body axis, is the coefficient factor related to the direction of the polarization pattern perpendicular to the insect body axis.

3. The insect species identification method based on multi-dimensional radar electromagnetic scattering feature sequence according to claim 1 is characterized in that: The radar cross-section characteristics include: the maximum value of the scattering cross-section of the insect polarization pattern ,and Vertical radar cross section , and Ratio and average , the calculation process of the radar cross section characteristics is as follows: Obtain the full polarization radar scattering matrix of the insect and calculate the average radar scattering cross section of the polarization pattern, the first and second harmonic components and factors of the radar scattering cross section: ， ， ， ， ， ， ， Calculation based on the obtained parameters: Maximum Radar Cross Section of Insect Polarimetric Patterns , and Vertical radar cross section , ， 。 4. The insect species identification method based on multi-dimensional radar electromagnetic scattering feature sequence according to claim 1 is characterized in that: The scattering matrix differential feature includes the differential amplitude feature of the insect's full polarization radar scattering matrix Differential Phase Characteristics , the calculation process of the scattering matrix differential feature includes: The fully polarized radar scattering matrix of insects calculates the two eigenvalues ​​of insects and Without loss of generality, assume ,but: ， ， in, and Respectively and The amplitude, and Respectively and The phase of Differential Amplitude Characteristics of Full Polarimetric Radar Scattering Matrix of Insects Differential Phase Characteristics The calculation formula is as follows: ， 。 5. The insect species identification method based on multi-dimensional radar electromagnetic scattering feature sequence according to claim 1 is characterized in that: The polarization pattern feature polarization invariant feature includes: two eigenvalues ​​of the insect's Graves power matrix and , Determinant ,trace and the Frobenius coefficient ; The calculation process of polarization invariant characteristics includes: Calculate the insect's Graves power matrix based on its full polarization radar scattering matrix: ； in, , , , ； Calculate the two eigenvalues ​​of the Graves power matrix of insects and for: ， ， Determinant of the Graves power matrix of insects is the square root of the product of two eigenvalues; Traces of the Graves power matrix of insects For two eigenvalues and The sum of: Frobenius coefficient of the Graves power matrix of insects For two eigenvalues and Calculates the square root of the sum of squares.

6. The insect species identification method based on multi-dimensional radar electromagnetic scattering feature sequence according to claim 1 is characterized in that: The feature normalization uses a linear normalization method to map the value of each electromagnetic scattering feature to the interval [0, 1] to ensure that different features have the same range and scale.

7. The insect species identification method based on multi-dimensional radar electromagnetic scattering feature sequence according to claim 1 is characterized in that: The convolutional layer is followed by a maximum pooling layer with a kernel size of 2, which compresses the input feature sequence and highlights the most significant features to ensure that the key dynamic features in the feature sequence are captured.

8. The insect species identification method based on multi-dimensional radar electromagnetic scattering feature sequence according to claim 1 is characterized in that: The training process of the insect recognition neural network includes: constructing a data set based on multi-dimensional electromagnetic scattering features and insect species labels, and dividing the data set into a training set and a validation set in a ratio of 7:3; During the training of the insect recognition neural network, the initial learning rate was set to 0.001, and a segmented learning rate decay strategy was used to reduce the learning rate to 10% of the original value after every 30 training rounds. The Adam optimization algorithm was used as the optimizer, and the cross entropy loss of the predicted insect species and the real insect species was used as the loss function in the training process. The goal was to minimize the loss function and adjust the parameters of the insect recognition neural network. After the training is completed, the insect recognition neural network is verified through a validation set that was not exposed during the insect recognition neural network training process.

9. An insect species identification device based on multi-dimensional radar electromagnetic scattering feature sequence, characterized in that: include: At least one processing unit, the processing unit is connected to the storage unit and the acquisition unit through a bus unit, the storage unit stores a computer program, and when the computer program is executed by the processing unit, the insect species identification method based on the multi-dimensional radar electromagnetic scattering feature sequence as described in any of claims 1-8 is implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the insect species identification method based on a multi-dimensional radar electromagnetic scattering feature sequence as described in any of claims 1 to 8 is implemented.

Images