Insect two-dimensional head orientation inversion method based on MLP-attention

Abstract

The present application belongs to the technical field of insect radar, and particularly relates to a two-dimensional head orientation inversion method based on MLP-Attention, which comprises the following steps: step one, constructing a data set based on the polarization echo data obtained by a multi-frequency full polarization coherent scattering radar measurement system under the microwave darkroom scene of experimental migratory insects and the polarization echo data collected by a high-resolution multi-frequency full polarization radar under the outside field flying scene; step two, designing an MLP-Attention architecture model constructed by taking a multilayer perceptron MLP as a base, adding a residual link and an attention module Attention, wherein the MLP-Attention architecture model realizes the mapping between the polarization scattering matrix of insects and the head orientation thereof, and completes the high-precision inversion of the head orientation, and the performance of the MLP-Attention architecture model under a low signal-to-noise ratio is particularly outstanding. The present application significantly improves the inversion performance under different noise environments, and provides strong support for improving the radar insect monitoring accuracy and optimizing the decision-making of agricultural pest control.

Description

Technical Field

[0001] This invention belongs to the field of insect radar technology, specifically relating to a two-dimensional head orientation inversion method for insects based on MLP-Attention. Background Technology

[0002] The head orientation of migratory insects is directly related to their flight direction, navigation behavior, and swarm migration patterns, making it a key parameter for agricultural pest early warning and ecological research. Insect radar technology provides technical support for long-distance, non-contact monitoring of migratory insects. Among them, high-resolution fully polarimetric radar can capture rich polarization scattering information, laying the data foundation for head orientation inversion.

[0003] Traditional vertical observation radar (VLR) determines insect orientation by rotating the polarization signal and determining the polarization direction at which the echo intensity is maximum. However, relying solely on echo amplitude information limits accuracy under low signal-to-noise ratio (SNR) conditions. High-resolution multi-frequency fully polarimetric insect radar (HMFPR) systems can directly acquire the polarization scattering matrix (SM). Based on the theory that the echo intensity is maximum when the polarization direction coincides with the principal eigenvector of the SM, a orientation estimation method based on the eigenvector of the scattering matrix has been developed. This method outperforms VLR, but still struggles to meet the high-precision monitoring requirements under low SNR conditions.

[0004] Deep neural networks possess powerful high-dimensional nonlinear mapping capabilities, enabling the discovery of complex correlations between polarization scattering and orientation angle. Multilayer perceptrons (MLPs) are simple and efficient, converging quickly at low feature dimensions, making them suitable for regression tasks. Therefore, methods integrating neural networks and polarization scattering information have become an important direction for improving the accuracy of insect head orientation inversion; to this end, this invention provides a two-dimensional insect head orientation inversion method based on MLP-Attention. Summary of the Invention

[0005] The purpose of this invention is to provide a two-dimensional head orientation inversion method for insects based on MLP-Attention, which significantly improves the inversion performance under different noise environments and provides strong support for improving the accuracy of radar insect monitoring and optimizing agricultural pest control decisions.

[0006] The specific technical solution adopted by this invention is as follows: A method for inverting the two-dimensional head orientation of insects based on MLP-Attention includes the following steps: Step 1: Based on the polarization echo data of the migratory insects in the microwave anechoic chamber scenario obtained by the multi-frequency fully polarized coherent scattering radar measurement system, and the polarization echo data collected by the high-resolution multi-frequency fully polarized radar in the field flying scenario, a dataset is constructed. Step 2: Based on the dataset, an MLP-Attention architecture model was designed, which is based on a multilayer perceptron (MLP) and incorporates residual links and an attention module. The MLP-Attention architecture model realizes the mapping between the insect polarization scattering matrix and its head orientation, and completes the high-precision inversion of head orientation. The MLP-Attention architecture model performs particularly well under low signal-to-noise ratio conditions.

[0007] Preferably, in step one, to obtain the true value of the insect's head orientation, this invention calculates the eigenvalues ​​of the insect's polarization scattering matrix and assigns them to the HH and VV common polarization channels of the scattering matrix in descending order, while setting the cross-polarization channel data of the scattering matrix to 0; through the above data correction method, an insect polarization scattering matrix with a true orientation value of 0 is obtained; the corrected scattering matrix It can be represented as: (1) in For larger eigenvalues, These are relatively small eigenvalues.

[0008] The head orientation of migratory insects in nature is randomly distributed. Using only a scattering matrix with an orientation of 0 as training data cannot accurately reflect the actual head orientation of migratory insects. Therefore, this invention uses a rotation matrix method to randomly rotate the insect's head orientation by an angle, obtaining a rotated correction matrix; the rotation matrix... Represented as: (2) in The angle of random rotation; the corrected scattering matrix after rotation. Represented as: (3) in Scattering matrix Larger eigenvalues For smaller eigenvalues, The angle of random rotation This is the correction matrix after rotation.

[0009] By extracting the channel elements from the correction matrix and dividing them according to their real and imaginary parts, the original dataset of 8-channel features is obtained, thus completing the data construction.

[0010] Preferably, in step two, the MLP-Attention architecture model is divided into three core modules: feature mapping and enhancement module, attention weighting module, and regression prediction module. Each module works together to achieve efficient inversion from scattering features to head orientation angle.

[0011] The feature mapping and enhancement module consists of cascaded residual MLP sub-modules, each of which contains a complete link including a fully connected layer, batch normalization, activation function, random deactivation, and residual connection. The dimensions of the fully connected layer decrease in a stepwise manner, enabling the gradual extraction from low-order scattering features to high-order abstract features; the fully connected layer maps the current feature dimension to the target dimension through linear transformation; Batch normalization reduces internal covariate bias by standardizing the input distribution, thereby accelerating model convergence. The activation function introduces a nonlinear transformation, enabling the model to fit the complex nonlinear relationship between scattering characteristics and orientation angle. Randomly inactivating and discarding some neurons can suppress overfitting. Residual connections sum the inputs and outputs of submodules to alleviate the gradient vanishing problem in deep networks.

[0012] The attention weighting module first maps the enhanced features to attention scores through a fully connected layer; then, it converts the scores into a probability distribution using a Softmax function; finally, it multiplies the weights element-wise with the input features to achieve adaptive scaling of the features. The attention scores are generated by learning the correlation strength between scattering features and orientation angle. During the weight scaling process, the weights of important features are amplified while the weights of noise-dominated features are weakened. The attention weighting module does not require manual definition of feature priorities; instead, it automatically learns the correlation weights between scattering features and orientation angle through data-driven learning, enabling the model to focus on key information even in complex noisy environments.

[0013] The regression prediction module employs a single fully connected layer to map features to predicted values. To ensure the physical meaning of the predicted values, the output layer does not use an activation function and directly outputs the linear transformation result. The regression prediction module compresses high-dimensional abstract features into specific orientation angle values ​​and guides the preceding modules to optimize feature extraction and weight allocation strategies through loss calculation with the real angle.

[0014] The technical effects achieved by this invention are as follows: The proposed MLP-Attention-based two-dimensional head orientation inversion method for insects can achieve an inversion accuracy of less than 1.4 degrees and a standard deviation of less than 5 degrees under high signal-to-noise ratio conditions. Under low signal-to-noise ratio conditions, the average inversion error and standard deviation are reduced by more than 50% compared with traditional methods, significantly improving the inversion performance under different noise environments. This provides strong support for improving the accuracy of radar insect monitoring and optimizing agricultural pest control decisions. Attached Figure Description

[0015] Figure 1 This is a flowchart of an insect two-dimensional head orientation inversion method based on MLP-Attention according to the present invention; Figure 2 This is a graph showing the trend of the average error of the MLP-Attention architecture model compared with the feature vector method as a function of signal-to-noise ratio. Figure 3 This is a graph showing the trend of the error standard deviation of the MLP-Attention architecture model compared with the feature vector method as a function of signal-to-noise ratio. Detailed Implementation

[0016] To make the objectives and advantages of this invention clearer, the invention will be specifically described below with reference to embodiments. It should be understood that the following text is merely used to describe one or more specific embodiments of the invention and does not strictly limit the scope of protection specifically claimed by the invention.

[0017] like Figure 1 As shown, a two-dimensional head orientation inversion method for insects based on MLP-Attention includes the following steps: Step 1: Construct a dataset based on the polarization echo data of the migratory insects obtained by the multi-frequency fully polarized coherent scattering radar measurement system in the microwave anechoic chamber scenario and the polarization echo data collected by the high-resolution multi-frequency fully polarized radar in the field flying scenario. Preferably, in step one, theoretically, due to the symmetry of insects about their body axis, when the insect's head orientation is perfectly parallel to the horizontal polarization direction of the radar, the cross-polarization channel data of its polarization scattering matrix must be 0. However, in experiments, it cannot be guaranteed that the insect's head orientation is perfectly parallel to the horizontal polarization direction of the radar, and there will inevitably be a very small error between the cross-polarization channel data of its scattering matrix and 0. In this case, it is impossible to obtain the accurate true head orientation of the measured insect. To obtain the true value of the measured insect head orientation, this invention calculates the eigenvalues ​​of the insect's polarization scattering matrix and then assigns them to the HH and VV common polarization channels of the scattering matrix in descending order, while setting the cross-polarization channel data of its scattering matrix to 0. Through the above data correction method, an insect polarization scattering matrix with a true orientation value of 0 is obtained; the corrected scattering matrix... It can be represented as: (1) in For larger eigenvalues, These are relatively small eigenvalues.

[0018] The head orientation of migratory insects in nature is randomly distributed. Using only a scattering matrix with an orientation of 0 as training data cannot accurately reflect the actual head orientation of migratory insects. Therefore, this invention uses a rotation matrix method to randomly rotate the insect's head orientation by an angle, obtaining a rotated correction matrix; the rotation matrix... Represented as: (2) in The angle of random rotation; the corrected scattering matrix after rotation. Represented as: (3) in Scattering matrix Larger eigenvalues For smaller eigenvalues, The angle of random rotation This is the correction matrix after rotation.

[0019] By extracting the channel elements from the correction matrix and dividing them according to their real and imaginary parts, the original dataset of 8-channel features is obtained, thus completing the data construction.

[0020] Step 2: Based on the dataset, an MLP-Attention architecture model was designed, which is based on a multilayer perceptron (MLP) and incorporates residual connections and an attention module. The MLP-Attention architecture model realizes the mapping between the insect polarization scattering matrix and its head orientation, and completes the high-precision inversion of head orientation. The MLP-Attention architecture model performs particularly well under low signal-to-noise ratio conditions.

[0021] The MLP-Attention architecture proposed in this invention aims to construct a precise mapping between the scattering matrix and head orientation of migratory insects. By fusing residual learning and attention mechanisms, it effectively extracts key information from polarization scattering features and suppresses noise interference. The MLP-Attention architecture consists of three core modules: a feature mapping and enhancement module, an attention weighting module, and a regression prediction module. These modules work together to achieve efficient inversion from scattering features to head orientation angle. The overall architecture of the MLP-Attention model is as follows: Figure 1 As shown, Feature mapping and enhancement module: It consists of cascaded residual MLP sub-modules. Each sub-module contains a complete link of fully connected layer, batch normalization, activation function, random deactivation and residual connection. The dimensions of the fully connected layer decrease in a stepwise manner, enabling the gradual extraction from low-order scattering features to high-order abstract features; the fully connected layer maps the current feature dimension to the target dimension through linear transformation; Batch normalization reduces internal covariate bias by standardizing the input distribution, thereby accelerating model convergence. The activation function introduces a nonlinear transformation, enabling the model to fit the complex nonlinear relationship between scattering characteristics and orientation angle. Randomly inactivating and discarding some neurons can suppress overfitting. Residual connections sum the inputs and outputs of submodules to alleviate the gradient vanishing problem in deep networks.

[0022] Attention weighting module: First, the enhanced features are mapped to attention scores through a fully connected layer; then, the scores are converted into a probability distribution through a Softmax function; finally, the weights are multiplied element-wise with the input features to achieve adaptive scaling of the features; the attention scores are generated by learning the correlation strength between scattering features and orientation angle. During the weight scaling process, the weights of important features are amplified while the weights of noise-dominated features are weakened; the attention weighting module does not require manual definition of feature priorities, but automatically learns the correlation weights between scattering features and orientation angle through data-driven learning, enabling the model to focus on key information even in complex noisy environments.

[0023] The regression prediction module uses a single fully connected layer to map features to predicted values. To ensure the physical meaning of the predicted values, the output layer does not use an activation function and directly outputs the linear transformation result. The regression prediction module compresses high-dimensional abstract features into specific orientation angle values ​​and guides the preceding modules to optimize feature extraction and weight allocation strategies by calculating the loss with the real angle.

[0024] The working principle of this invention is as follows: To verify the effectiveness of the insect two-dimensional head orientation inversion method based on the MLP-Attention architecture model of this invention, a database construction experiment was conducted using insect planned scattering echo data obtained from microwave anechoic chamber experiments and outdoor hoisting experiments. The specific data construction is as follows: Microwave anechoic chamber The microwave anechoic chamber experimental setup is a multi-band, fully polarized, coherent scattering radar measurement system, mainly composed of a four-port vector network analyzer, three pairs of dual-polarized horn antennas (covering the X-band (8-12GHz), Ku-band (12-18GHz), and Ka-band (34-36GHz)), and a wooden horn-shaped shielded box filled with absorbing material. Insect samples are suspended on a metal ring at the top of an experimental support approximately 2m high. This ring, approximately 1.3m in diameter, is rotatable to control the insect's body axis orientation. The insects are suspended by 0.1mm diameter polyethylene (PE) threads, with both ends fixed to the metal ring to ensure that the insect's abdomen faces downwards, its body axis is parallel to the H-polarization direction of the antenna, and it is accurately positioned at the center of the antenna beam, thus ensuring the repeatability and measurement accuracy of the experiment. The anechoic chamber data used in this invention consists of anechoic polarization scattering echo data of 251 insects from 2018 to 2019, covering 20 species. The measured insect body size parameters have a wide range, with body length ranging from 10.3 mm to 47 mm, body width from 2.2 mm to 14 mm, and weight ranging from 20.2 mg to 1002.8 mg.

[0025] Outdoor hoisting The field suspension experiment setup is a high-resolution, multi-band, fully polarimetric radar system. The radar system can simultaneously perform measurements in the X, Ku, and Ka bands, with the X and Ku bands each containing two sub-bands, providing a total of five operating frequencies: 9.5 GHz, 11.5 GHz, 15.5 GHz, 17.5 GHz, and 35 GHz. The radar employs independent horizontal (H) and vertical (V) polarimetric (V) transceiver channels, acquiring four polarimetric echoes (HH, HV, VH, VV) at each frequency, thus achieving fully polarimetric measurements and constructing the polarimetric scattering matrix of the insect target. A small amount of glue was used to fix the insect's back to an ultra-fine PE line. This PE line, with a diameter of only 0.02 mm, had negligible scattering effects under experimental conditions. Subsequently, the two ends of a 100 m long PE line were fixed to the mounting brackets of two UAVs, ensuring they were taut and straight to reduce additional swaying interference. The two UAVs simultaneously ascended vertically to a height of 200 m at the same speed, suspending the insect at the predetermined measurement position. To minimize the impact of polarization mismatch, the direction of the PE line is strictly controlled to remain parallel to the polarization direction of the radar antenna. Once the insect has ascended to the designated position, the radar system enters target search mode until the beam center is precisely aligned with the insect, then switches to target tracking mode for stable measurement. The field data used in this invention comprises polarization scattering echo data from 442 insects, covering 11 species, collected between 2018 and 2020. The measured insect weight ranged from 15.65 to 1040 mg, body length from 9.9 to 45 mm, and body width from 2.5 to 8.5 mm.

[0026] Noise-free insect polarization scattering matrix data was used as the dataset, combined with microwave anechoic chamber and field-flying data, covering frequencies of 9.5, 11.5, 15.5, and 17.5 GHz. The microwave anechoic chamber dataset contained 251 insects, and the field-flying dataset contained 442 insects, totaling 693 migratory insects of 25 species. However, the limited data volume was insufficient for achieving a good model fit. Therefore, this invention employs a rotational augmentation method to expand the dataset. Each preprocessed scattering matrix (SM) is rotated 30 random angles between -90 and +90 degrees. The rotated scattering matrix is ​​considered a new orientation scattering matrix measured for that insect at that frequency. The rotational augmentation does not change the eigenvalues ​​or other scattering characteristics of the scattering matrix and can still be considered as measured data. After rotational augmentation, the dataset size reached 83,160 sets, which were then divided into training, validation, and test sets in a 7:1:2 ratio.

[0027] The experiment uses the mean square error (MAE) and mean squared error (RMSE) as evaluation metrics for the model inversion. MAE can be expressed as: (4) in The amount of data in the test set. These are the model's predicted values. It is a true value. RMSE can be represented as: (5) in The amount of data in the test set. These are the model's predicted values. The values ​​are considered true. The accuracy of the inversion results was assessed using evaluation metrics. The average error of the inversion results was less than 1.4°, and the standard deviation was less than 5°. Specific performance indicators are shown in Table 1. It can be seen that the model has high inversion accuracy and strong robustness.

[0028] Table 1 Model Inversion Performance Indicators

[0029] To verify the insect 2D head orientation inversion method based on the MLP-Attention architecture model of this invention, improvements in estimation accuracy, robustness, and noise resistance of existing orientation inversion methods based on free polarization extraction of scattering matrices are achieved. Test data with different signal-to-noise ratios (SNRs) are used to test and verify the model. The original test data with scattering parameters and angle labels are read, noise is generated by iterating through different SNRs, the noise intensity is calculated, and complex noise is generated and superimposed on the original matrix. Finally, the real and imaginary parts are extracted from the noisy matrix to reconstruct the feature vector, obtaining noisy SMs with SNRs ranging from 1 to 50 dB. Noise testing uses the trained model to make batch predictions on the noisy test set, and the average and standard deviation of the inversion error under different SNRs are summarized to quantify the model's performance stability under noise interference.

[0030] The trend of the mean error of the MLP-Attention architecture model versus the feature vector method as a function of signal-to-noise ratio is as follows: Figure 2 As shown, the red line represents the average error of the proposed model in retrieving orientation from noisy test data at various angles as a function of signal-to-noise ratio (SNR), while the blue line represents the error result obtained using the eigenvector method. It can be seen that both the red and blue lines gradually decrease and converge as the SNR increases. The red line is lower than the blue line when the SNR is less than approximately 38 dB, and slightly higher than the blue line when the SNR is greater than or equal to 38 dB, but the difference is not significant. This indicates that the average error of the proposed model gradually decreases as the SNR increases, significantly outperforming the eigenvector method when the SNR is less than 38 dB, while its average error is slightly higher than the eigenvector method when the SNR is greater than or equal to 38 dB. The trend of the standard deviation of the error of the MLP-Attention architecture model compared to the eigenvector method with SNR is shown below. Figure 3As shown, the red line represents the standard deviation of the inversion error of the proposed model on noisy test data as a function of the signal-to-noise ratio (SNR), while the blue line represents the error result obtained using the eigenvector method. It can be seen that both the red and blue lines gradually decrease and converge as the SNR increases, with the red line consistently below the blue line, and its change is more stable. This indicates that the standard deviation of the proposed model's inversion error gradually decreases with increasing SNR, and is significantly lower than that of the eigenvector method at all SNR levels. Furthermore, its change with SNR is more gradual, demonstrating its strong robustness and significantly superior noise resistance compared to the eigenvector method. Therefore, this verifies the performance improvement of the present invention compared to existing methods.

[0031] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention. Structures, devices, and operating methods not specifically described or explained in this invention are implemented according to conventional methods in the art unless otherwise specified or limited.

Claims

1. A method for inverting the two-dimensional head orientation of insects based on MLP-Attention, characterized in that: Includes the following steps: Step 1: Based on the polarization echo data of the migratory insects in the microwave anechoic chamber scenario obtained by the multi-frequency fully polarized coherent scattering radar measurement system, and the polarization echo data collected by the high-resolution multi-frequency fully polarized radar in the field flying scenario, a dataset is constructed. Step 2: Based on the dataset, an MLP-Attention architecture model was designed, which is based on a multilayer perceptron (MLP) and includes residual links and an attention module. The MLP-Attention architecture model realizes the mapping between the insect polarization scattering matrix and its head orientation, thus completing the head orientation inversion.

2. The method for inverting two-dimensional head orientation of insects based on MLP-Attention according to claim 1, characterized in that: In step one, to obtain the true value of the insect's head orientation, eigenvalues ​​are calculated from the insect's polarization scattering matrix. These eigenvalues ​​are then assigned to the HH and VV common polarization channels of the scattering matrix in descending order, while the cross-polarization channel data is set to 0. Through this data correction method, an insect polarization scattering matrix with a true orientation value of 0 is obtained. The corrected scattering matrix... It can be represented as: (1) in For larger eigenvalues, These are relatively small eigenvalues.

3. The method for inverting two-dimensional head orientation of insects based on MLP-Attention according to claim 2, characterized in that: By using a rotation matrix method to randomly rotate the insect's head by an angle, a correction matrix after rotation was obtained; rotation matrix Represented as: (2) in The angle of random rotation; the corrected scattering matrix after rotation. Represented as: (3) in Scattering matrix Larger eigenvalues For smaller eigenvalues, The angle of random rotation Rotation matrix transpose, This is the correction matrix after rotation; By extracting the channel elements from the correction matrix and dividing them according to their real and imaginary parts, the original dataset of 8-channel features is obtained, thus completing the data construction.

4. The method for inverting two-dimensional head orientation of insects based on MLP-Attention according to claim 3, characterized in that: In step two, the MLP-Attention architecture model is divided into three core modules: feature mapping and enhancement module, attention weighting module, and regression prediction module. These modules work together to achieve efficient inversion from scattering features to head orientation angle.

5. The method for inverting two-dimensional head orientation of insects based on MLP-Attention according to claim 4, characterized in that: The feature mapping and enhancement module consists of cascaded residual MLP sub-modules, each of which contains a complete link of fully connected layers, batch normalization, activation functions, random deactivation, and residual connections. The dimensions of the fully connected layer decrease in a stepwise manner, enabling the gradual extraction from low-order scattering features to high-order abstract features; the fully connected layer maps the current feature dimension to the target dimension through linear transformation; Batch normalization reduces internal covariate bias by standardizing the input distribution, thereby accelerating model convergence. The activation function introduces a nonlinear transformation, enabling the model to fit the complex nonlinear relationship between scattering characteristics and orientation angle. Randomly inactivating and discarding some neurons can suppress overfitting. Residual connections sum the inputs and outputs of submodules to alleviate the gradient vanishing problem in deep networks.

6. The method for inverting two-dimensional head orientation of insects based on MLP-Attention according to claim 5, characterized in that: The attention weighting module first maps the enhanced features to attention scores through a fully connected layer; then it converts the scores into a probability distribution through a Softmax function; finally, it multiplies the weights element-wise with the input features to achieve adaptive scaling of the features. Attention scores are generated by learning the correlation strength between scattering features and orientation angle. During weight scaling, the weights of important features are amplified while the weights of noise-dominated features are weakened.

7. The method for inverting two-dimensional head orientation of insects based on MLP-Attention according to claim 6, characterized in that: The regression prediction module uses a single fully connected layer to map features to predicted values; the output layer does not use an activation function and directly outputs the linear transformation result.

Images