Digital Beamforming Method for Pushbroom Radiometer Based on Multilayer Perceptron Network

Abstract

The present invention relates to a digital beam synthesis method for a pushbroom radiometer based on a multi-layer perceptron network. The method first establishes a brightness temperature image database as the input signal for the observation of the dense feed array of the pushbroom radiometer. The output signals of all feed observations are used as the input layer of the multi-layer perceptron network, and are fully connected through multiple neurons in the hidden layer. Each neuron in the output layer is obtained by the weighted sum of the output values of all neurons in the previous hidden layer. A loss function is established between the output result of the output layer and the set expected output result through forward propagation, and it is judged whether it is lower than the set threshold. If it is lower, the model training is completed, and the model parameters after digital beam synthesis are output. Otherwise, the first-order partial derivative of the loss function is taken, the network connection weights and biases of the hidden layer are updated, and backpropagation is performed to the hidden layer for a new round of model training until the model parameters when the output meets the set threshold are obtained. Thus, the digital beam synthesis of the pushbroom radiometer is completed.

Description

Technical Field

[0001] The present invention relates to a digital beam synthesis method for a pushbroom radiometer based on a multi-layer perceptron network, and belongs to the field of space microwave remote sensing technology. Background Art

[0002] Different from the real aperture radiometer and synthetic aperture radiometer systems, the beam synthesis pushbroom radiometer system performs beam synthesis in the digital domain to achieve a narrow beam and ultra-high main beam efficiency of the antenna electrical performance, avoid the contradiction between mechanical scanning and large-aperture antennas, and can be used to fill the gap in nearshore high-resolution and high-precision data detection. The pushbroom radiometer system synthesizes multiple digital beams by weighted summation of the outputs of a dense feed array. The beam width of the digital beam directly determines the spatial resolution of the system, and the main beam efficiency determines the detection distance from the coastline and the detection accuracy of the system. The spaceborne application of the digital beam synthesis pushbroom radiometer system is still blank, and there are also few beam requirements based on ultra-high main beam efficiency and ultra-high resolution. The traditional beam synthesis method can no longer meet the application requirements.

[0003] [[ID=ID=11]]In order to utilize limited feeds, a digital beam synthesis method is adopted to form pattern information with a narrow beam and low sidelobes to meet the application index requirements of high resolution, high detection accuracy, and near-offshore detection distance. The most common digital beam synthesis methods are: (1) using an adaptive filter combined with the LMS algorithm, with the minimum mean square error as the judgment criterion, to update the weighted system of beam synthesis; (2) using a genetic-sequence quadratic programming combined optimization method, first searching for the global optimal value, and then strengthening the local search. Disadvantages: The above methods all search for the optimal value of the weighting coefficient based on a single linear summation and the least squares criterion. In order to obtain higher application indexes, the traditional methods have low search efficiency and it is difficult to obtain the amplitude-phase weighting value that meets the optimization target through a single linear combination. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to overcome the above-mentioned deficiencies of the prior art, provide a digital beam synthesis method for a pushbroom radiometer based on a multi-layer perceptron network, improve the resolution, detection accuracy, and offshore detection distance of the digital beam synthesis pushbroom radiometer system, and provide a method for realizing a high-performance pushbroom radiometer system.

[0005] The technical solution of the present invention is as follows: A digital beam synthesis method for a pushbroom radiometer based on a multi-layer perceptron network, and the steps of the method are as follows:

[0006] S1. Establish a multi-layer perceptron network and initialize the network connection weights and biases. The multi-layer perceptron network includes an input layer, L - 1 hidden layers, and an output layer. The input layer is used to receive the output signals of k feeds of the pushbroom radiometer. The L - 1 hidden layers and the output layer perform complex weighted summation on the received signals of the k feeds, and H optimized digital beams are obtained at the output layer.

[0007] S2. Collect color pictures of different observation scenes, convert the color pictures into grayscale pictures, regard the grayscale pictures as the brightness temperature images obtained by the pushbroom radiometer, and establish a brightness temperature image database; each pixel coordinate in the grayscale picture is equivalent to the feed observation angle. The grayscale value of each pixel is equivalent to the brightness temperature value.

[0008] S3. Extract the brightness temperature image from the brightness temperature image database, deduce the output signals of the k feeds according to the brightness temperature image, and use the output signals of the k feeds as the input of the multi-layer perceptron network. After being processed by the multi-layer perceptron network, the output value of the output layer is obtained.

[0009] S4. According to the antenna pattern of the H beams synthesized by the desired digital beams Construct the expected output signals (t1, t2, …, t h , …, t H ) of all neurons in the output layer. Combine the output value of the output layer in S4 to establish a loss function.

[0010] S5. Determine whether the loss function is lower than the set threshold. If so, the training of the multi-layer perceptron network is completed. Otherwise, find the first-order partial derivative of the loss function, update the network connection weights and biases, and backpropagate to all hidden layers. Repeat steps S3 to S5, continuously iterate, and stop the backpropagation until the loss function is lower than the set threshold, and complete the training of the multi-layer perceptron network.

[0011] S6. Use the trained multi-layer perceptron network to perform digital beam synthesis on the pushbroom radiometer.

[0012] Preferably, in the multi-layer perceptron network:

[0013] The output signals of the k feeds received by the input layer are respectively output to the neurons of the first hidden layer. The neurons of the input layer and the hidden layer correspond one by one to the feeds. Each neuron of the hidden layer and the output layer is connected to all neurons of the previous layer, that is, fully connected, and the neurons in the same layer are not connected to each other. Denote the hidden layer and the output layer as L neural network layers in total, and define as the input value of the jth neuron in the lth neural network layer of the multi-layer perceptron network. and are the output value and bias value of the j-th neuron in the l-th neural network layer respectively, is the network connection weight between the n-th neuron in the (l - 1)-th layer and the j-th neuron in the l-th layer, then there is:

[0014]

[0015] In the formula, f(·) is the activation function, and N is the number of neurons in the (l - 1)-th layer, which is the same as the number of feeds.

[0016] Preferably, the activation function is a rectified linear unit function, sigmoid function, tanh function or radial basis function.

[0017] Preferably, in step S3, the output signals of k feeds are expressed as The output signal of feed i is:

[0018]

[0019] In the formula, η feed,Mi is the main beam efficiency corresponding to the secondary antenna pattern of feed i corresponding to, is the average radiation brightness temperature of the observed scene corresponding to the main lobe range of the feed main lobe, represents the average radiation brightness temperature of the observed scene corresponding to the sidelobe beam range other than the main lobe.

[0020] Preferably, the desired output signal of the h-th neuron in the output layer is:

[0021]

[0022] In the formula, η beam,Mh is the main beam efficiency corresponding to the antenna pattern of the h-th beam after digital beamforming corresponding to, is the average radiation brightness temperature of the observed scene corresponding to the main lobe range of the beam antenna pattern, represents the average radiation brightness temperature of the observed scene corresponding to the sidelobe beam range other than the main lobe of the beam antenna pattern.

[0023] Preferably, the main lobe is defined as the 2.5 times 3dB beam width range of the secondary antenna pattern of the feed, and the 3dB beam width of the feed is Then the main lobe range of the feed is defined as

[0024] Preferably, the loss function E is:

[0025]

[0026] Among them, is the output of the h-th neuron in the output layer L, and t h is the expected output of the h-th neuron in the output layer.

[0027] Preferably, the formula for updating the network connection weights is:

[0028] Among them, is the network connection weight between the n neurons in the (l - 1)-th layer and the j-th neuron in the l-th layer after backpropagation update;

[0029] Preferably, the update formula for the bias is:

[0030]

[0031] is the bias of the j-th neuron in the l-th layer after backpropagation update.

[0032] Preferably, the secondary antenna pattern of the feed i is obtained through the following method: In the spherical near field, place the ring focus reflector and the dense feed array of the pushbroom radiometer system at the central target point specified by the spherical near field robotic arm with the geometric position center of the central unit of the feed array as the center; Rotate and move the robotic arm of the spherical near field by controlling the motor scanning, so that the spherical near field RF emission signal is at different elevation angles θ and azimuth angles in the feed array coordinate system, covering the full solid angle of the feed, and all RF signals of the feed in the full solid angle can be obtained; Divide the RF signals of all feeds by the amplitude and subtract the phase of the spherical near field RF emission signal, and the secondary antenna pattern of each feed at each elevation angle θ and azimuth angle can be obtained, where i = 1, …, k, and k is the number of feeds of the pushbroom radiometer.

[0033] The advantages of the present invention compared with the prior art are as follows:

[0034] (1) The pushbroom radiometer digital beamforming method based on a multi-layer perceptron network of the present invention uses multiple neurons and multiple levels of perception to form a multi-layer perceptron network, solves the problem of linear inseparability in digital beamforming, uses the multi-layer perceptron network to form forward propagation, improves the efficiency of model training, establishes a loss function using the results of the output layer, updates the network weights through first-order partial derivatives, backpropagates to the hidden layer, and trains the network connection weights until the loss function meets the requirements, thus completing the training of the multi-layer perceptron network and completing the synthesis optimization of the digital beam.

[0035] (2) In the present invention, the complex double integral expression of the input layer signal and the output layer signal of the digital beamforming problem is simplified into a linear combination of the main beam efficiency of the feed and the beam and the average radiation brightness temperature of the observation scene. The simplified parameters directly characterize the electrical performance of the secondary antenna pattern of the feed before beamforming and the beam antenna pattern after beamforming. The optimization problem is more intuitive, and the process of model training is simplified.

[0036] (3) The beamforming method based on the multi-layer perceptron network proposed in the present invention uses the complex activation functions of the multi-layer network and multiple neurons for non-linear combination, which is different from the traditional synthesis method that uses a single linear summation and the least squares criterion to search for the optimal value of the weighting coefficient, and is conducive to the realization of the high-resolution and high-precision requirement indicators of the pushbroom radiometer.

[0037] (4) In the present invention, by collecting color pictures of different scenes, converting the color pictures into grayscale pictures, and defining the grayscale value of the pictures as the brightness temperature, a brightness temperature image is formed, and a brightness temperature image database for training the multi-layer perceptron network is established. The scene pictures have characteristics such as complex details, strong contrast, and various styles, which are the same as those of the pushbroom radiometer observing the Earth scene in spaceborne applications. The training of the multi-layer perceptron network with the complex image database enhances the robustness and fault tolerance of digital beamforming and has extremely strong adaptability.

[0038] (5) The present invention is a solution method proposed for the optimization problem in digital beamforming for multiple feeds, and can be applied to radiometers and phased array type radar systems for multi-beam synthesis in spaceborne, airborne, and ground applications to improve the performance indicators of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] FIG. 1(a) is a schematic diagram of the antenna of the pushbroom radiometer system according to an embodiment of the present invention;

[0040] FIG. 1(b) is a schematic diagram of the observation beam footprint of the pushbroom radiometer system according to an embodiment of the present invention;

[0041] FIG. 1(c) is a schematic diagram of the angular coordinate of the antenna pattern according to an embodiment of the present invention;

[0042] FIG. 2(a) is a schematic diagram of the secondary antenna pattern of the feed obtained by testing according to an embodiment of the present invention;

[0043] FIG. 2(b) is a schematic diagram of the desired beam antenna pattern according to an embodiment of the present invention;

[0044] Figure 3 is a flowchart for establishing the brightness temperature image database according to an embodiment of the present invention;

[0045] Figure 4 is a topological structure diagram of the multi-layer perceptron network according to an embodiment of the present invention;

[0046] Figure 5 Block diagram of the digital beam synthesis neuron structure according to an embodiment of the present invention;

[0047] Figure 6 Method flowchart according to an embodiment of the present invention. Detailed implementation manners

[0048] The following further describes in detail the specific implementation manners of the present invention with reference to the accompanying drawings.

[0049] Different from the traditional mechanical scanning radiometer system, as shown in Fig. 1(a), the pushbroom radiometer is composed of a ring focus reflector combined with a dense feed array, forming multiple feed secondary antenna patterns with different pointing directions at the same time, as shown in Fig. 2(a). Since the key application of the pushbroom radiometer system is to achieve high-resolution and high-precision nearshore detection, the antenna pattern obtained in Fig. 2(a) cannot meet the requirements of resolution and detection accuracy. In order to further obtain an antenna pattern with a narrow beam and low sidelobes, as shown in Fig. 2(a), it is necessary to perform complex weighted summation on the outputs of multiple feeds and obtain the required beam antenna pattern through digital beam synthesis. As shown in Fig. 1(b), the ellipse in the figure is the observation footprint (i.e., the beam footprint) formed by the 3dB beam angle range of each beam antenna pattern after digital beam synthesis. The pushbroom radiometer system does not need to perform mechanical scanning and can obtain multiple footprints in one observation. The footprints are connected and distributed to form a swath, thereby overcoming the engineering implementation problem of large-aperture antenna mechanical scanning for high resolution and large swath, and becoming a research hotspot in the field of microwave remote sensing.

[0050] In order to utilize the limited feeds of the pushbroom radiometer and form an antenna pattern with a narrow beam and low sidelobes by means of digital beam synthesis to meet the application index requirements of high resolution, high detection accuracy, and near-offshore detection distance. The most common digital beam synthesis method uses an adaptive filter combined with the LMS algorithm, and updates the weighted system of beam synthesis with the minimum mean square error as the judgment criterion; or uses a genetic-sequence quadratic programming combined optimization method, first searches for the global optimal value, and then strengthens the local search. The above methods all search for the optimal value of the weighting coefficient based on a single linear summation and the least squares criterion. In order to obtain higher application indexes, the traditional methods have low search efficiency and it is difficult to obtain the weighting value that meets the optimization goal through a single linear combination.

[0051] The push-broom radiometer observes the Earth by means of satellite-borne means, and can measure the microwave radiation brightness temperature of Earth target scenes such as the atmosphere, ocean, polar regions, and nearshore areas. The present method invents a digital beam synthesis method for a push-broom radiometer based on a multi-layer perceptron network, which can improve the resolution and detection accuracy of the system. The high-precision detection results will provide important information for weather forecasting, ocean environmental monitoring, disaster forecasting, etc., and play a very important role in improving the accuracy of short-term extreme weather and long-term climate prediction, and have extremely important significance for China's national defense construction and national life.

[0052] The present invention provides a digital beam synthesis method for a push-broom radiometer based on a multi-layer perceptron network, and the specific steps are as follows:

[0053] S1. Measure all the secondary antenna patterns of the feeders under the condition of the ring focus reflector surface combined with the dense feeder array of the push-broom radiometer (i = 1, …, k), where k is the number of feeders of the push-broom radiometer;

[0054] The preferred solution is specifically as follows:

[0055] Each secondary antenna pattern is specifically: in the spherical near field, place the ring focus reflector surface and the dense feeder array of the push-broom radiometer system at the central target point specified by the spherical near field robotic arm with the geometric position center of the central unit of the feeder array as the center (provided by the measurement site); by controlling the rotation and movement of the robotic arm of the spherical near field through the motor scanning method, make the spherical near field radio frequency emission signal (provided by the measurement site) be at different elevation angles θ and azimuth angles in the feeder array coordinate system covering the full solid angle of the feeders, and all the radio frequency signals of the feeders can be obtained; divide the amplitude of the radio frequency signals of all the feeders by the amplitude of the spherical near field radio frequency emission signal (provided by the measurement site) and subtract the phase, and obtain the secondary antenna pattern of each feeder at each elevation angle θ and azimuth angle (i = 1, …, k), where k is the number of feeders of the push-broom radiometer. (i = 1, …, k), where k is the number of feeders of the push-broom radiometer.

[0056] For the requirements of the secondary antenna pattern test: as shown in Figure 1(c), first take the geometric center of the measured feeder in the dense feeder array as the origin o of the coordinate system, the axis perpendicular to the plane of the dense feeder array as the z-axis, and the direction to the right along the array as the x-axis, and establish an xyz coordinate system according to the right-hand rule. Define the angle between the straight line connecting the test position on the antenna pattern and the origin and the positive z-axis as the elevation angle θ, and the intersection angle of the projection of this straight line on the xoy plane and the positive x-axis as the azimuth angle Measure the secondary and antenna patterns of each feeder in the dense feeder array as where θ ∈ [0~180], that is, corresponding to the full solid angle.

[0057] S2. Collect color pictures of different observation scenarios, convert the color pictures into grayscale pictures, regard the grayscale pictures as the brightness temperature images obtained by the pushbroom radiometer, and establish a brightness temperature image database; each pixel coordinate in the grayscale picture is equivalent to the feed observation angle The grayscale value of each pixel is equivalent to the brightness temperature value;

[0058] The preferred solution is as follows:

[0059] As Figure 3 shown, collect color pictures of different observation scenarios, which can be pictures with complex details and strong contrast, and then convert the color pictures into grayscale pictures. The grayscale value of the pictures is equivalent to the brightness temperature value, and the coordinate of each grayscale value is the pixel coordinate, thus forming a brightness temperature image; a large number of brightness temperature images form a database, adjust the image size, and the pixel coordinate is equivalent to the feed observation angle The grayscale value of each observation angle is the brightness temperature As the original scene brightness temperature image database during the training of the multi-layer perceptron network.

[0060] S3. Extract the brightness temperature image from the brightness temperature image database, and derive the output signals of k feeds according to the brightness temperature image;

[0061] The feed receives the radiation brightness temperature After the weighted integration of the feed secondary antenna pattern the output signal of the i-th feed is obtained (i = 1,..., k); specifically as follows:

[0062] The radiation brightness temperature of the brightness temperature image is observed by the feed array composed of k feeds The output signals of the k feeds in the array are expressed as The relationship expression between and the brightness temperature is:

[0063]

[0064] In the formula, represents the secondary antenna pattern data of the i-th feed.

[0065] In actual situations, the secondary antenna pattern information of the feed is not close to the ideal Dealt function and is distributed in the entire spatial solid angle region. In addition to receiving the radiation brightness temperature of the target area through the main beam, it also receives the radiation brightness temperature of the non-target area from other azimuths. The formula (1) is written as the expression of the main lobe and non-main lobe of the antenna pattern:

[0066]

[0067] The main lobe is defined as the range of 2.5 times the 3dB beam width of the secondary antenna pattern of the feed, and the 3dB beam width of the feed is Then the main lobe range of the feed is defined as Since the double integral form shown in Equation (2) is complex in expression and not conducive to the calculation of the mathematical model for network training, it is necessary to further simplify Equation (2) into the expression of the main beam efficiency, which can be written as:

[0068]

[0069] In the formula, is the secondary antenna pattern of feed i corresponding main beam efficiency, is within the main lobe range of the feed (i.e., the angular range, represents 2.5 times the angular range), the average value of the radiation brightness temperature of the observation scene corresponding thereto, represents the sidelobe beam range except for the main lobe the average value of the radiation brightness temperature of the observation scene corresponding thereto, as shown in Equation (4).

[0070]

[0071] S4. The output signals of k feeds are used as the input of the multi-layer perceptron network. After being processed by the multi-layer perceptron network, the output value

[0072] of the output layer is obtained. The preferred solution is as follows:

[0073] The output signals of k feeds received by the input layer are respectively output to the neurons of the first hidden layer. After all neurons in the first hidden layer perform weighted sum processing on the input signals, the output value of the first hidden layer is obtained. And so on, the output values transmitted to all neurons in the l-1 hidden layer. After completing the operations of all hidden layers, weighted sum processing is performed again to obtain the output value

[0074] Such as Figure 4As shown, the received signals of k feeds are used as the input layer. After being optimized through the 1st layer, 2nd layer, …, (l - 1)th layer, lth layer, …, (L - 1)th layer of hidden layers, H optimized digital beams are obtained at the Lth layer output layer. The forward propagation model of digital beamforming adopts a multi-layer perceptron network, which is a network model composed of an input layer, hidden layer (one or more layers) and an output layer, and can solve the problem of linear inseparability that cannot be solved by a single-layer perceptron network. In a specific embodiment of the present invention, the relationship between the input layer and the output layer is established, and a three-layer shallow model is used to establish a multi-layer perceptron network to improve the efficiency of model training.

[0075] The neurons in the input layer receive input signals. Each neuron in the hidden layer and the output layer is connected to all neurons in the adjacent layer, that is, fully connected, and the neurons in the same layer are not connected. As Figure 4 shown, the line segments with arrows represent the connection between neurons and the direction of signal transmission, and each connection has a network connection weight. As Figure 5 shown, it represents the signal flow of the outputs of all neurons in the (l - 1)th hidden layer to the jth neuron in the lth output layer. is the input value of the jth neuron in the lth layer of the multi-layer perceptron network. and are the output value and bias value of the jth neuron in the lth layer respectively. is the network connection weight between the nth neuron in the (l - 1)th layer and the jth neuron in the lth layer, then there is:

[0076]

[0077] In the formula, N is the number of neurons in the (l - 1)th layer, and f(·) is an activation function, which can be a linear rectification function, sigmoid function, tanh function, radial basis function, etc.

[0078] S5. According to the expected antenna pattern of the H beams after digital beamforming Construct the expected output signals (t1, t2, …, t h , …, t H ) of all neurons in the output layer, combine with the output values of the output layer in S4, establish a loss function, and by judging whether the loss function is lower than the set threshold. If so, the training of the multi-layer perceptron network is completed. Otherwise, find the first-order partial derivative of the loss function, update the network connection weights and biases, and backpropagate to all hidden layers. Repeat steps S3 to S5, and continuously iterate until the loss function is lower than the set threshold and stop backpropagation. At this point, the obtained model parameters are the parameters of the digital beamforming of the pushbroom radiometer.

[0079] The preferred solution is as follows:

[0080] As shown in Fig. 2(b), set the antenna pattern of H beams after digital beamforming When the bright temperature image is input to S3 then the expected output value t of the h-th neuron in the output layer of S4 h is:

[0081]

[0082] In the formula, is the antenna pattern of the h-th beam after digital beamforming corresponding main beam efficiency, is the 3dB beam width of the beam antenna pattern after digital beamforming, is within the main lobe range of the beam antenna pattern (i.e., angle range, represents 2.5 times angle range) corresponding average observed scene radiation bright temperature, represents the sidelobe beam range of the beam antenna pattern except for the main lobe corresponding average observed scene radiation bright temperature.

[0083]

[0084] Before digital beamforming, the electrical performance of the secondary antenna pattern of the feed is characterized by: 3dB beam width main beam efficiency η feed,Mi ; after digital beamforming, the electrical performance of the beam antenna pattern is characterized by: 3dB beam width characterizes the resolution index of the pushbroom radiometer system, and the main beam efficiency η beam,Mh then characterizes the offshore detection distance and detection accuracy of the pushbroom radiometer system.

[0085] The preferred solution is as follows:

[0086] As Figure 6 shown, S4 completes the forward propagation model training of the multi-layer perceptron network, and uses the output value of the L-th output layer and the expected output value (t1, t2,..., t h ,..., t H ) to establish the loss function E based on the expectation of the difference between them:

[0087]

[0088] Among them, is the output of the h-th neuron in the output layer L, and t his the expected output of the h-th neuron in the output layer. In a specific embodiment of the present invention, L takes the value of 4.

[0089] When the loss function is lower than the set threshold, output the parameters of the model training to complete the digital beam synthesis of the pushbroom radiometer for the multi-layer perceptron network; otherwise, take the first-order partial derivative of the loss function, update the network connection weights and biases, and then perform as Figure 6 shown in the backpropagation model training, and the update formula is:

[0090]

[0091] where is the network connection weight between the n neurons in the (l - 1)-th layer and the j neurons in the l-th layer after backpropagation update, is the network connection weight between the n neurons in the (l - 1)-th layer and the j neurons in the l-th layer before update; is the bias of the j neurons in the l-th layer after backpropagation update, is the bias of the j neurons in the l-th layer before update.

[0092] The updated network connection weights are backpropagated to all hidden layers of the perceptron network, and continuously iterated until the backpropagation stops when the output is lower than the expected value of the set threshold. At this point, the obtained model parameters are the parameters for the digital beam synthesis of the pushbroom radiometer.

[0093] S6. Perform digital beam synthesis on the pushbroom radiometer using the trained multi-layer perceptron network.

[0094] In summary, the present invention innovatively uses a multi-layer perceptron network model, combines the constraints of backpropagation and the update of weighted coefficients, solves the problem of linear inseparability of the optimization objective, and is conducive to the realization of higher demand indicators through the combination of multi-layer complex activation functions; at the same time, since the feed output signal undergoes multi-layer complex non-linear combinations, the robustness and fault tolerance of digital beam synthesis are enhanced, that is, when applied in spaceborne applications, the change in the electrical performance of the feed has little impact on the digital beam performance.

[0095] The content not described in detail in the specification of the present invention belongs to the well-known technology of those skilled in the art.

Claims

1. A digital beamforming method for a pushbroom radiometer based on a multi-layer perceptron network, characterized in that The steps are as follows: S1. Establish a multi-layer perceptron network and initialize the network connection weights and biases. The multi-layer perceptron network includes an input layer, L - 1 hidden layers, and an output layer. The input layer is used to receive the output signals of k feeds of the pushbroom radiometer. The L - 1 hidden layers and the output layer perform complex weighted summation on the received signals of the k feeds, and H optimized digital beams are obtained at the output layer; S2. Collect color pictures of different observation scenes, convert the color pictures into grayscale pictures, regard the grayscale pictures as the brightness temperature images obtained by the pushbroom radiometer, and establish a brightness temperature image database; Each pixel coordinate in the grayscale image is equivalent to the feed observation angle The grayscale value of each pixel is equivalent to the brightness temperature value; S3. Extract the brightness temperature image from the brightness temperature image database, derive the output signals of k feeds based on the brightness temperature image, and use the output signals of the k feeds as the input of the multi-layer perceptron network. After being processed by the multi-layer perceptron network, the output value of the output layer is obtained S4. According to the antenna patterns of the H beams after the digital beams to be obtained are synthesized Construct the expected output signals (t1, t2, …, t h , …, t H ), …, t ) of all neurons in the output layer, and establish a loss function by combining the output values of the output layer in S4; S5. Determine whether the loss function is lower than the set threshold. If so, the training of the multi-layer perceptron network is completed. Otherwise, find the first-order partial derivative of the loss function, update the network connection weights and biases, and backpropagate to all hidden layers. Repeat steps S3 to S5, continuously iterate, and stop the backpropagation until the loss function is lower than the set threshold, and complete the training of the multi-layer perceptron network; S6. Use the trained multi-layer perceptron network to perform digital beam synthesis on the pushbroom radiometer.

2. The pushbroom radiometer digital beam synthesis method based on a multi-layer perceptron network according to claim 1, wherein In the multi-layer perceptron network: The output signals of k feeds received by the input layer are respectively output to the neurons of the first hidden layer. The neurons of the input layer and the hidden layer correspond one by one to the feeds. Each neuron of the hidden layer and the output layer is connected to all neurons of the previous layer, that is, fully connected. Neurons in the same layer are not connected to each other. The hidden layer and the output layer are collectively recorded as L neural network layers. Define as the input value of the j-th neuron in the l-th neural network layer of the multi-layer perceptron network, and are respectively the output value and the bias value of the j-th neuron in the l-th neural network layer. is the network connection weight between the n-th neuron in the (l - 1)-th layer and the j-th neuron in the l-th layer. Then there is: In the formula, f(·) is the activation function, and N is the number of neurons in the (l - 1)-th layer, which is the same as the number of feeds.

3. The pushbroom radiometer digital beam synthesis method based on a multi-layer perceptron network according to claim 2, characterized in that The activation function is a rectified linear unit function, sigmoid function, tanh function, or radial basis function.

4. The pushbroom radiometer digital beam synthesis method based on a multi-layer perceptron network according to claim 1, characterized in that In the step S3, the output signals of the k feeds are expressed as the output signal of feed i is: Where, η feed,Mi is the secondary antenna pattern of feed i corresponding main beam efficiency, is the average radiation brightness temperature of the observed scene corresponding to the main lobe range of the feed, represents the average radiation brightness temperature of the observed scene corresponding to the sidelobe beam range except the main lobe.

5. The pushbroom radiometer digital beam synthesis method based on a multi-layer perceptron network according to claim 1, characterized in that The expected output signal of the h-th neuron in the output layer is: where η beam,Mh is the antenna pattern of the h-th beam after digital beamforming corresponding main beam efficiency, is the average radiation brightness temperature of the observed scene corresponding to the main lobe range of the beam antenna pattern, represents the average radiation brightness temperature of the observed scene corresponding to the sidelobe beam range other than the main lobe of the beam antenna pattern.

6. The pushbroom radiometer digital beam synthesis method based on a multi-layer perceptron network according to claim 4 or 5, characterized in that The main lobe is defined as the range of 2.5 times the 3dB beam width of the feed secondary antenna pattern, and the 3dB beam width of the feed is Then the main lobe range of the feed is defined as 7. The pushbroom radiometer digital beam synthesis method based on a multi-layer perceptron network according to claim 1, characterized in that The loss function E is: where, is the output of the h-th neuron in the output layer L, and t h is the expected output of the h-th neuron in the output layer.

8. The digital beam synthesis method of a pushbroom radiometer based on a multi-layer perceptron network according to claim 7, characterized in that The formula for updating the network connection weights is: Among them, The network connection weight between the n neurons in the (l - 1)-th layer after backpropagation update and the j neurons in the l-th layer is the network connection weight between the n neurons in the (l - 1)-th layer before update and the j neurons in the l-th layer.

9. The digital beam synthesis method of a pushbroom radiometer based on a multi-layer perceptron network according to claim 7, characterized in that The formula for updating the bias is: is the bias of the j-th neuron in the l-th layer after backpropagation update, is the bias of the j-th neuron in the l-th layer before the update.

10. The pushbroom radiometer digital beam synthesis method based on a multi-layer perceptron network according to claim 1, characterized in that The secondary antenna pattern of the feed i It is obtained through the following method: in the spherical near-field, place the ring-focus reflector and the dense feed array of the pushbroom radiometer system at the central target point specified by the spherical near-field robotic arm with the geometric position center of the central unit of the feed array as the center; by controlling the motor scanning, rotate and move the robotic arm of the spherical near-field to make the spherical near-field RF transmission signal at different elevation angles θ and azimuth angles cover the full solid angle of the feed, and the RF signals of all feeds in the full solid angle can be obtained; divide the RF signals of all feeds by the amplitude of the spherical near-field RF transmission signal and subtract the phases to obtain the secondary antenna pattern of each feed at each elevation angle θ and azimuth angle k is the number of feeds of the pushbroom radiometer. ​

Images