Satellite remote sensing target data recognition method based on artificial intelligence

Abstract

The present invention discloses a satellite remote sensing target data recognition method based on artificial intelligence, which belongs to the field of data processing technology. By performing satellite remote sensing target recognition on the collected spectral vectors, a distributed federated learning model training architecture is adopted to realize distributed model training and data mining of a multi-node joint model; the collected spectral vector data is manually labeled; a generative adversarial network based on multi-scale spectral decomposition is adopted to generate samples, thereby realizing data expansion; and a satellite remote sensing target data recognition model is adopted to perform data recognition. The present invention expands the data set by adopting a generative adversarial network based on multi-scale spectral decomposition to generate samples. An autoencoder algorithm is used to perform feature dimensionality reduction, and feature independence is strengthened through a nonlinear transformation layer, thereby reducing redundant information between features. A kernel extreme learning machine algorithm based on conjugate gradient compensation is adopted, and the kernel function is mapped to a high-dimensional space, thereby improving the accuracy and efficiency of classification.

Description

Technical Field

[0001] The invention belongs to the technical field of data processing, and in particular relates to a satellite remote sensing target data recognition method based on artificial intelligence. Background Art

[0002] With the development of remote sensing technology, hyperspectral satellite sensors can capture spectral data of different wavelengths reflected by the surface and generate a large number of hyperspectral remote sensing vectors. These data are of great value in many fields such as environmental monitoring, resource exploration, and urban planning. However, the traditional centralized data processing model has the problem of data islands, which limits the effective aggregation and utilization of data resources. In addition, due to the high dimensionality and complexity of hyperspectral data, traditional data processing methods face challenges in feature extraction, dimensionality reduction, and classification, especially when the number of samples is insufficient, the generalization ability and classification accuracy of the model are difficult to guarantee.

[0003] In the prior art, a Chinese invention patent with publication number CN111144423A proposes a method for extracting multi-scale spectral features from hyperspectral remote sensing data based on a one-dimensional group convolutional neural network, which belongs to the technical field of remote sensing data feature extraction. In order to solve the problem of the single scale of spectral feature extraction of hyperspectral remote sensing data based on convolutional neural networks, and at the same time solve the problem of dimensionality disaster caused by high-dimensional signals of hyperspectral remote sensing data, the high-dimensional hyperspectral vector signal is directly divided into several groups in the spectral band dimension based on the principle that the product of the value after each scale value is reduced by one is equal to the corresponding number of groups. Each scale uses multiple different filters to perform a separate one-dimensional conventional convolution operation on each segment of the high-dimensional hyperspectral vector signal, and the characteristic dimension of the spectral vector is quickly reduced through several groups of operations.

[0004] The Chinese invention patent with publication number CN114493318A proposes a separation and unmixing method of hyperspectral reflectance curves based on weighted NMF, which is used for separation and unmixing of sintered ore hyperspectra and qualitative detection of Fe~(2+) and Fe~(3+), including the following steps: S1, establishing a hyperspectral benchmark data set M of sintered ore samples with different iron grades; S2, performing correlation analysis on the hyperspectral benchmark data set M and the Fe~(2+) standard hyperspectral curve; S3, performing averaging and whitening processing on the hyperspectral benchmark data set M to obtain a data set MT; S4, determining initial weights based on correlation coefficients, performing weighted NMF iterative decomposition on the data set MT, decomposing and identifying Fe~(2+) spectral vector SF1 and Fe~(3+) spectral vector SF2, and using the hyperspectral reflectance curve as the basis and the mathematical method of weighted NMF iterative decomposition to quickly separate and unmix the Fe~(2+) and Fe~(3+) component spectra in the sintered ore, which is a non-destructive qualitative detection method for the Fe content in the sintered ore.

[0005] The Chinese invention patent with publication number CN114416768A proposes a spectral matching method, device, equipment and storage medium of an enhanced genetic algorithm, the method comprising: constructing a matching mathematical model and a matching spectral database; wherein the spectral database is composed of n m-dimensional discretized spectral vectors ln(λ), and the matching mathematical model is expressed as kn is the matching coefficient, and L(λ) is the matching spectral vector; obtaining a target spectral vector; establishing a fitness function based on the Euclidean distance and angle between the target spectral vector and the matching spectral vector; inputting the matching spectral database, the target spectral vector and the fitness function into the genetic algorithm according to the matching mathematical model, optimizing and solving the matching coefficient through the genetic algorithm to output the optimal matching coefficient matrix and matching results, having better local search capability, improving the optimization capability, improving the accuracy of the matching results, and reducing the running time of the algorithm.

[0006] The existing technology has the following problems: in the traditional centralized training mode, the data resources of each node are relatively independent, which makes it impossible to effectively aggregate the data resources, limiting the effect of model training; when processing hyperspectral remote sensing data, the existing technology often affects the generalization ability and classification accuracy of the model due to insufficient sample number; traditional methods may have problems of inefficiency and feature redundancy in the process of feature extraction and dimensionality reduction, affecting the final performance of the model; when processing remote sensing data classification, the existing technology may be limited in classification accuracy and efficiency due to the insufficient processing ability of the classifier for nonlinear features. Summary of the invention

[0007] The technical problem to be solved by the present invention is to overcome the shortcomings of the above-mentioned prior art and provide a satellite remote sensing target data recognition method based on artificial intelligence.

[0008] The technical solution adopted to solve the above technical problems is: a satellite remote sensing target data recognition method based on artificial intelligence, comprising the following steps:

[0009] S1. By performing satellite remote sensing target recognition on the collected spectral vectors, the distributed federated learning model training architecture is adopted to realize the distributed model training and data mining of the multi-node joint model; the distributed federated learning model training architecture is: in each round of iteration, each node performs local model training respectively, and uploads the trained model parameters to the central server. The central server completes parameter aggregation and update, and sends the updated parameters to each node to start a new round of iteration. Repeating this iterative operation indicates that the global model of the central server and the model training of each node are completed; S2. Manually annotate the collected spectral vector data, The labeled categories include: urban building target areas, water body target areas, vegetation target areas, and bare land target areas; S3. A generative adversarial network based on multi-scale spectral decomposition is used to generate samples to achieve data expansion. The generative adversarial network algorithm based on multi-scale spectral decomposition includes two main components: a generator and a discriminator. The generator is used to generate data of real spectral vectors, and the discriminator is used to distinguish between real input data and input data generated by the generator; S4. A satellite remote sensing target data recognition model is used for data recognition. The satellite remote sensing target data recognition model includes a spectral vector feature dimensionality reduction module and a remote sensing target data classification module.

[0010] Furthermore, the training process of the generative adversarial network algorithm based on multi-scale spectral decomposition in step S3 is:

[0011] S301, using a weight initialization strategy based on wavelet transform to initialize the generator and discriminator of the generative adversarial network, the initialization method is expressed as:

[0012] ,

[0013] In the formula, is the initialization value of the weight of the generated adversarial network, is the wavelet transform basis strategy function, is the scale parameter of wavelet transform;

[0014] The wavelet transform basis strategy function is used to generate initial weights with multi-scale characteristics, and the calculation method is expressed as:

[0015] ,

[0016] In the formula, is the circumference of a circle, is the wave number of the wavelet function, which determines the oscillation frequency of the wavelet function, and x is the independent variable of the function. is the phase shift of the wavelet function, which is used to adjust the starting phase of the wavelet. is an exponential function with a natural constant as base, is the standard deviation of the Gaussian window, controlling the locality of the wavelet, Set to 0.04;

[0017] S302, perform rough tuning of the generative adversarial network. The generator and the discriminator are trained using a standard adversarial training framework. The loss function of the rough tuning stage is expressed as:

[0018] ,

[0019] In the formula, To combat the loss function, For the generator, is the discriminator, For the real data set, is a random noise distribution, For expectations, From the real data distribution The data sampled in To obey a specific distribution, is the distribution of real data, is the discriminator function, is the noise distribution, is the generator function, From the noise distribution The vector sampled in ;

[0020] S303, fine-tune the generative adversarial network, use the piecewise linear mapping strategy of the spectral vector to adjust the generator in detail, the generator The output is realized through a piecewise linear function, expressed as:

[0021] ,

[0022] In the formula, is a piecewise linear mapping function responsible for input noise To transform, is the parameter set of the mapping function, is the number of segments, Set to 5, is the linear coefficient within the i-th segment, is the ReLU activation function, is the nonlinear function, is the slope of the i-th segment, is the offset within the i-th segment, is the quadratic term coefficient used to increase the model's ability to handle the nonlinear characteristics of spectral data. Set to 0.5;

[0023] is the Sigmoid activation function, which is a nonlinear function used to further process the mapped data to approximate the real data distribution. is the adjustment coefficient that controls the angle adjustment range, is a sinusoidal function of the spectral angle, which is used to ensure that the adjustment amplitude is proportional to the angular deviation, is the spectral angle mapping function, is the target spectral vector, i.e., the reference vector selected from the real data set;

[0024] A dynamic feature correction strategy based on spectral angle mapping is used to improve the spectral authenticity of the data generated in the data augmentation algorithm. A spectral angle mapping function is defined to calculate and adjust the angle of the generated spectral data. The calculation method is expressed as:

[0025] ,

[0026] In the formula, For input noise Piecewise linear mapping function for transformation ;

[0027] S304, after the generator can reproduce the single-scale spectral data well, a multi-scale optimization strategy is adopted to optimize the generation performance of the generator at each scale by processing spectral data of different scales in parallel. The data generation method of the multi-scale optimization strategy stage is expressed as:

[0028] ,

[0029] In the formula, is a scale-specific processing function, is the weight matrix of the generator, nd is the number of scales of the multi-scale optimization strategy, is the nonlinear activation function of the jth scale, is the weight matrix of the jth scale;

[0030] is the generator function of the multi-scale optimization stage, It is the concatenation of processing results at different scales. is the weight matrix of the first scale, is the weight matrix of the ndth scale;

[0031] In order to adapt to the data processing requirements of different scales, the calculation method of the nonlinear activation function of the jth scale is expressed as:

[0032] ,

[0033] In the formula, x is the independent variable of the function, is the first adjustable parameter of the activation function, is the second adjustable parameter of the activation function;

[0034] S305. In the fine-tuning stage and the multi-scale optimization stage, the loss function calculation method of the fine-tuning stage and the multi-scale optimization stage is expressed as:

[0035] ,

[0036] In the formula, is the loss function of the fine-tuning stage and the multi-scale optimization stage, is the first loss weight of the loss function in the fine-tuning stage and the multi-scale optimization stage, Set to 0.3, is the second loss weight of the loss function in the fine-tuning stage and the multi-scale optimization stage, Set to 0.7, Dynamic weights are training parameters that allow the model to optimize loss weights based on the specific characteristics of the data and enhance sensitivity to key spectral features. The number of spectral data entered for the current batch, is the kth real spectrum data, is the noise input corresponding to the kth real spectrum data, for Norm, calculated in the same way as Euclidean distance, is used to measure the spectral distance between generated data and real data;

[0037] S306, repeat the above steps until a preset stop iteration condition is met. The preset stop iteration condition is reaching a preset maximum number of iterations, which means that the model training is completed.

[0038] Furthermore, the spectral vector feature dimension reduction module in step S4 uses an autoencoder algorithm to reduce the dimension of the spectral vector feature;

[0039] The training process of the autoencoder algorithm is as follows:

[0040] S4101. Initialize the weights and biases of the autoencoder model. The initialization method is expressed as:

[0041] ,

[0042] In the formula, is the initial value of the autoencoder weight, is the mean shift of the autoencoder weights, is the standard deviation of the autoencoder weight initialization, is a normal distribution with mean 0 and covariance matrix as the identity matrix, is a normal distribution, is the identity matrix; is the initial value of the autoencoder bias, is the mean shift of the autoencoder bias;

[0043] S4102. In the encoding stage, the input data is firstly subjected to preliminary feature extraction through multiple convolutional layers or fully connected layers, and then subjected to feature independence processing through a nonlinear transformation layer. The nonlinear transformation layer uses a special activation function and a constrained optimization algorithm to enhance the independence of features. The processing method of the nonlinear transformation layer is expressed as follows:

[0044] ,

[0045] In the formula, is the transformed feature vector, is the hyperbolic tangent function, which is used to increase the nonlinearity of the output characteristics. and are the weights and biases of the nonlinear transformation layer, is the input feature vector;

[0046] The weight update method of the nonlinear transformation layer is expressed as:

[0047] ,

[0048] In the formula, is the loss function of the autoencoder, is the learning rate of the autoencoder;

[0049] S4103. In the decoding stage, feature data is processed layer by layer. Each layer focuses on a specific set of feature dimensions, and uses the hierarchical feature deconstruction strategy to gradually restore the output close to the original data. The decoding method of each layer is expressed as:

[0050] ,

[0051] In the formula, is the reconstructed output vector, is the adaptive Sigmoid activation function, and are the weights and biases of the decoding layer, respectively. is the vector element multiplication symbol, is the feature reweighting function, is the reconstruction error vector of the feature corresponding to the transformed feature vector;

[0052] In the process of feature dimensionality reduction, the feature reweighting function allows the model to adjust the emphasis on each feature during the training process, thereby improving the target relevance and efficiency of feature dimensionality reduction. The calculation method of the feature reweighting function is expressed as:

[0053] ,

[0054] In the formula, is the adjustment coefficient of the feature reweighting function, which controls the sensitivity of the error impact, and nf is the number of features after dimensionality reduction; is the i-th element of the reconstruction error vector corresponding to the feature after the conversion, is the i-th element of the transformed eigenvector;

[0055] The parameters of the adaptive Sigmoid activation function are dynamically adjusted according to the reconstruction error, and the calculation method is expressed as:

[0056] ,

[0057] In the formula, x is the independent variable of the function, is the adaptive parameter of the activation function;

[0058] The calculation method of the adaptive parameters of the activation function is expressed as:

[0059] ,

[0060] In the formula, is the adjustment parameter of the initial activation function, Set to 0.3, is the adjustment factor of the activation function, for norm;

[0061] S4104. Calculate the loss function based on the reconstruction error and feature independence index to ensure the validity and independence of the features during model training. The calculation method of the loss function is expressed as:

[0062] ,

[0063] In the formula, is the eigenvector The mutual information of is used to evaluate the independence of features. is the regularization parameter, Set to 0.3;

[0064] S4105. Repeat the above steps until a preset stop iteration condition is met. The preset stop iteration condition is reaching a preset maximum number of iterations, which means that the model training is completed.

[0065] Furthermore, the remote sensing target data classification module in step S4 adopts a kernel extreme learning machine algorithm based on conjugate gradient compensation as a classifier model. The kernel extreme learning machine algorithm based on conjugate gradient compensation includes an input layer, a kernel transformation layer and an output layer. The input layer receives data from the feature dimensionality reduction model, the kernel transformation layer uses a kernel function to map the data to a high-dimensional space, and the output layer is responsible for generating a final classification decision.

[0066] The training process of the kernel extreme learning machine algorithm based on conjugate gradient compensation is:

[0067] S4201, using a polynomial kernel function as the kernel function of the kernel extreme learning machine, expressed as:

[0068] ,

[0069] In the formula, is the kernel function of the extreme learning machine, is the first input vector of the kernel extreme learning machine, is the second input vector of the kernel extreme learning machine, is the scaling parameter of the kernel function, To calculate the inner product operation, is the free term of the kernel extreme learning machine, de is the degree of the polynomial;

[0070] The way to calculate the inner product in the kernel function is expressed as:

[0071] ,

[0072] in, and Respectively, vector and The i-th element of un is is the number of input features of the kernel extreme learning machine;

[0073] Initialize the weights of the kernel extreme learning machine. The initialization method is expressed as:

[0074] ,

[0075] In the formula, is the weight matrix of the kernel extreme learning machine, To generate standard normally distributed random numbers, is the number of input features of the kernel extreme learning machine, is the number of output categories of the kernel extreme learning machine;

[0076] S4202, map the input data to a high-dimensional space through a kernel function, and calculate the mapped features as follows:

[0077] ,

[0078] In the formula, is the feature mapping function of the kernel extreme learning machine, is the kernel function of the extreme learning machine, is the kernel function correlation diagonal matrix, whose diagonal elements are composition, is the ui-th element of the kernel function correlation diagonal matrix, representing the weight of each feature;

[0079] Adopting the adaptive feature weight adjustment strategy to dynamically adjust the weight of the input features, the influence of each feature is enhanced or weakened according to its actual performance in the classification task, and the initial value of each element of the kernel function correlation diagonal matrix is ​​set to 1, that is, it means that all features have an equal influence on the classification result in the initial state;

[0080] In each iteration, the weight of the feature is automatically adjusted according to its contribution to the loss function, and the update method of the elements in the diagonal matrix of the kernel function correlation is expressed as:

[0081] ,

[0082] In the formula, is the ui-th element of the diagonal matrix of the kernel function correlation at the t+1-th iteration, is the ui-th element of the diagonal matrix of the kernel function correlation at the t-th iteration, is the diagonal matrix adjustment factor, Set to 0.01, is the loss function of the kernel extreme learning machine The partial derivative of

[0083] S4203, using the conjugate gradient method optimization strategy to optimize the weight vector in the kernel space, expressed as:

[0084] ,

[0085] In the formula, is the conjugate gradient of the t+1th iteration, is the gradient of the loss function of the kernel extreme learning machine with respect to the weights at the tth iteration, is the loss function of the kernel extreme learning machine, is the weight matrix of the kernel extreme learning machine at the tth iteration, is the learning rate adjustment coefficient of the kernel extreme learning machine at the tth iteration, is the conjugate gradient of the tth iteration;

[0086] is the weight matrix of the kernel extreme learning machine at the t+1th iteration, is the direction adjustment coefficient of the kernel extreme learning machine at the tth iteration;

[0087] The calculation method of the learning rate adjustment coefficient of the kernel extreme learning machine at the tth iteration is expressed as:

[0088] ,

[0089] In the formula, for norm, The gradient of the loss function of the kernel extreme learning machine at the tth iteration with respect to the weight is an indicator of the current gradient size and is used to adjust the search direction for the next step. is the gradient of the loss function of the kernel extreme learning machine with respect to the weights at the t-1th iteration, is the weight matrix of the kernel extreme learning machine at the t-1th iteration;

[0090] S4204, calculating the loss function, the loss function of the kernel extreme learning machine used is the logistic regression loss with a regularization term, and the calculation method is expressed as:

[0091] ,

[0092] In the formula, is the loss function of the kernel extreme learning machine, mfd is the total number of samples input in the current batch, is the true label of the jth input sample, is the Sigmoid activation function, is the transpose of the weight matrix of the kernel extreme learning machine, is the jth input sample, is the regularization parameter, Set to 0.2;

[0093] S4205, weights are updated through back propagation algorithm and conjugate gradient compensation mechanism, and the updating method is expressed as:

[0094] ,

[0095] In the formula, is the weight matrix of the kernel extreme learning machine at the t+1th iteration, is the learning rate of the kernel extreme learning machine, Set to 0.01;

[0096] S4206. Use the trained model to make classification decisions and output the specific category of the remote sensing target data according to the characteristics of the input data. The classification function is expressed as:

[0097] ,

[0098] In the formula, is the true label of the sample, is the function used to convert the model output into classification labels.

[0099] Furthermore, the distributed federated learning model training architecture in step S1 adopts the federated average algorithm as the model aggregation algorithm. The federated average algorithm realizes the collaborative training of local models through multiple global iterations. For each global iteration, the number of nodes is set to The total number of samples is , and the number of samples of the kth node is , then the objective function of federated learning is defined as:

[0100] ,

[0101] In the formula, is the objective function of federated learning, are model parameters, is the model parameter Loss prediction for the jth sample data;

[0102] is the training loss, the training loss The cross entropy loss function is used for calculation. is the sample feature of the jth sample data, The label of the jth sample data.

[0103] Furthermore, the multiple nodes in step S1 include k nodes, and for the kth node, the objective function of the node is defined as:

[0104] ,

[0105] In the formula, is the number of samples of the kth node, is the objective function of the kth node, is the data distribution of the kth node;

[0106] At the tth iteration, let the parameter gradient of the kth node be , then the way the model updates parameters at the tth iteration is expressed as:

[0107] ,

[0108] In the formula, is the model parameter of the t+1th iteration, is the model parameter of the tth iteration, is the learning rate for the current parameter update, is the number of samples of the kth node;

[0109] The parameter update method of the global model of the central server is expressed as:

[0110] ,

[0111] In the formula, are the parameters of the global model of the central server at the t+1th iteration.

[0112] The beneficial effects of the present invention are as follows: (1) In view of the problem of data resource islands in the traditional centralized training mode, the present invention adopts a distributed federated learning model training architecture to realize distributed training and data mining of multi-node joint models, solves the problem of data islands, and protects the data privacy of each node. (2) In view of the poor generalization ability and accuracy of the model caused by insufficient training samples, the present invention adopts a generative adversarial network based on multi-scale spectral decomposition to generate samples, expand the data set, and improve the generalization ability and accuracy of the model. (3) In view of the high-dimensional and complex problem in the feature extraction of hyperspectral remote sensing data, the present invention uses an autoencoder algorithm to reduce the dimension of features, strengthens the independence of features through a nonlinear transformation layer, and reduces the redundant information between features. (4) In view of the need for nonlinear feature processing in remote sensing target data classification, the present invention adopts a kernel extreme learning machine algorithm based on conjugate gradient compensation, maps to a high-dimensional space through a kernel function, and improves the accuracy and efficiency of classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0113] Figure 1 It is a principle block diagram of an embodiment of the model training architecture of distributed federated learning of the present invention.

[0114] Figure 2 It is a principle block diagram of data recognition using a satellite remote sensing target data recognition model.

[0115] Figure 3 It is a spectral data curve graph of satellite remote sensing target data. DETAILED DESCRIPTION

[0116] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0117] The satellite remote sensing target data recognition method based on artificial intelligence of this embodiment includes the following steps:

[0118] S1. Satellite remote sensing target recognition is performed on the collected spectral vectors. The target recognition task of satellite remote sensing requires a large amount of effective training data for feature extraction, analysis and learning. This embodiment adopts a distributed federated learning model training architecture to achieve distributed model training and data mining of multi-node joint models, breaking the data resource island problem of the traditional centralized training model and aggregating more effective data. The distributed federated learning architecture can prevent the leakage of local data of each node. The data is used and trained locally, effectively ensuring data security.

[0119] In the federated learning framework, each node does not need to share the data of its local training model. Instead, it trains the local model and sends the updated model to the centralized learning unit for aggregation. The model training architecture of distributed federated learning is as follows: Figure 1 shown.

[0120] The model training architecture of distributed federated learning is as follows: in each round of iteration, each node performs local model training and uploads the trained model parameters to the central server. The central server completes parameter aggregation and update, and sends the updated parameters to each node to start a new round of iteration. Repeating this iterative operation means that the global model of the central server and the model training of each node are completed.

[0121] The model training architecture of distributed federated learning uses the federated average algorithm as the model aggregation algorithm. The federated average algorithm realizes the collaborative training of local models through multiple global iterations. For each global iteration, the number of nodes is set to The total number of samples is , and the number of samples of the kth node is , then the objective function of federated learning is defined as:

[0122] ,

[0123] In the formula, is the objective function of federated learning, are model parameters, is the model parameter Loss prediction for the jth sample data;

[0124] is the training loss, the training loss The cross entropy loss function is used for calculation. is the sample feature of the jth sample data, The label of the jth sample data.

[0125] The multi-node includes k nodes. For the kth node, the objective function of the node is defined as:

[0126] ,

[0127] In the formula, is the number of samples of the kth node, is the objective function of the kth node, is the data distribution of the kth node.

[0128] At the tth iteration, let the parameter gradient of the kth node be , then the way the model updates parameters at the tth iteration is expressed as:

[0129] ,

[0130] In the formula, is the model parameter of the t+1th iteration, is the model parameter of the tth iteration, is the learning rate for the current parameter update, is the number of samples of the kth node.

[0131] The parameter update method of the global model of the central server is expressed as:

[0132] ,

[0133] In the formula, are the parameters of the global model of the central server at the t+1th iteration.

[0134] S2. Manually label the collected spectral vector data, and the labeled categories include: urban building target area, water body target area, vegetation target area, and bare land target area.

[0135] The data collection source of each node is a hyperspectral satellite sensor, which can capture spectral data of different wavelengths reflected from the surface. The collected data is stored in the form of hyperspectral remote sensing vector data in a structured JSON format. Figure 3 As shown, the dimension of a piece of data is 200, that is, the wavelength of the spectrum vector is 200.

[0136] S3. The acquisition, labeling and preprocessing of training data are time-consuming and labor-intensive, and insufficient training samples can easily lead to poor generalization ability of the model, while affecting the accuracy of the model. This embodiment uses a generative adversarial network based on multi-scale spectral decomposition to generate samples, thereby achieving data expansion. The generative adversarial network algorithm based on multi-scale spectral decomposition includes two main components: a generator and a discriminator. The generator is used to generate data of a real spectral vector, and the discriminator is used to distinguish between real input data and input data generated by the generator.

[0137] This embodiment adopts a multi-scale spectral decomposition mechanism, which can capture and simulate subtle differences in surface features at different spectral scales. At the same time, a piecewise linear mapping strategy of spectral vectors is adopted to more accurately simulate the nonlinear characteristics of spectral data, so that the generated data has a higher similarity with the real data in terms of spectral characteristics.

[0138] The training process of the generative adversarial network algorithm based on multi-scale spectral decomposition is:

[0139] S301. A weight initialization strategy based on wavelet transform is used to initialize the generator and discriminator of the generative adversarial network to facilitate the model to capture the multi-scale characteristics of spectral data at an early stage. The initialization method is expressed as:

[0140] ,

[0141] In the formula, is the initialization value of the weight of the generated adversarial network, is the wavelet transform basis strategy function, is the scale parameter of wavelet transform.

[0142] The wavelet transform basis strategy function is used to generate initial weights with multi-scale characteristics, and the calculation method is expressed as:

[0143] ,

[0144] In the formula, is the circumference of a circle, is the wave number of the wavelet function, which determines the oscillation frequency of the wavelet function, and x is the independent variable of the function. is the phase shift of the wavelet function, which is used to adjust the starting phase of the wavelet. is an exponential function with a natural constant as base, is the standard deviation of the Gaussian window, controlling the locality of the wavelet, Set to 0.04.

[0145] S302, perform rough tuning of the generative adversarial network. The generator and the discriminator are trained through a standard adversarial training framework. The goal is to enable the generator to macroscopically approximate the distribution of real data. The main optimization is the overall structure and parameters of the generator to ensure that the generated spectral vector is similar to the real data as a whole. The loss function of the rough tuning stage is expressed as:

[0146] ,

[0147] In the formula, To combat the loss function, For the generator, is the discriminator, For the real data set, is a random noise distribution, For expectations, From the real data distribution The data sampled in To obey a specific distribution, is the distribution of real data, is the discriminator function, is the noise distribution, is the generator function, From the noise distribution The vector sampled from .

[0148] S303, fine-tune the generative adversarial network, adopt the piecewise linear mapping strategy of the spectral vector, adjust the generator in detail, and focus on improving the local detail accuracy of the generated data. By adjusting the mapping function parameters of each segment to approximate the spectral details in the real data, the generator The output is realized through a piecewise linear function, expressed as:

[0149] ,

[0150] In the formula, is a piecewise linear mapping function responsible for input noise To transform, is the parameter set of the mapping function, is the number of segments, Set to 5, is the linear coefficient within the i-th segment, is the ReLU activation function, is the nonlinear function, is the slope of the i-th segment, is the offset within the i-th segment, is the quadratic term coefficient used to increase the model's ability to handle the nonlinear characteristics of spectral data. Set to 0.5;

[0151] is the Sigmoid activation function, which is a nonlinear function used to further process the mapped data to approximate the real data distribution. is the adjustment coefficient that controls the angle adjustment range, is a sinusoidal function of the spectral angle, which is used to ensure that the adjustment amplitude is proportional to the angular deviation, is the spectral angle mapping function, is the target spectral vector, i.e., the reference vector selected from the real data set.

[0152] A dynamic feature correction strategy based on spectral angle mapping is used to improve the spectral authenticity of the data generated in the data augmentation algorithm. The feature correction strategy dynamically adjusts the spectral angle of the generated data through the spectral angle mapping function to ensure that it is consistent with the real data in the spectral space. In remote sensing data processing, the spectral angle describes the angular difference of the spectra of different substances. The spectral angle of the generated data is dynamically adjusted to ensure that it is consistent with the real data in the spectral space. The spectral angle mapping function is defined to calculate and adjust the angle of the generated spectral data. The calculation method is expressed as:

[0153] ,

[0154] In the formula, For input noise Piecewise linear mapping function for transformation .

[0155] S304, after the generator can reproduce the single-scale spectral data well, a multi-scale optimization strategy is adopted to optimize the generation performance of the generator at each scale by processing spectral data of different scales in parallel, so as to improve its simulation ability of complex spectral features. The data generation method of the multi-scale optimization strategy stage is expressed as:

[0156] ,

[0157] In the formula, is a scale-specific processing function, is the weight matrix of the generator, nd is the number of scales of the multi-scale optimization strategy, is the nonlinear activation function of the jth scale, is the weight matrix of the jth scale;

[0158] is the generator function of the multi-scale optimization stage, It is the concatenation of processing results at different scales. is the weight matrix of the first scale, is the weight matrix of the ndth scale.

[0159] In order to adapt to the data processing requirements of different scales, the calculation method of the nonlinear activation function of the jth scale is expressed as:

[0160] ,

[0161] In the formula, x is the independent variable of the function, is the first adjustable parameter of the activation function, is the second adjustable parameter of the activation function.

[0162] S305. In the fine-tuning stage and the multi-scale optimization stage, the loss function is used to guide the generator to more accurately simulate the data features within a specific spectral distance. Based on the spectral distance loss, the training process pays more attention to the authenticity and accuracy of the spectral data. The loss function calculation method of the fine-tuning stage and the multi-scale optimization stage is expressed as:

[0163] ,

[0164] In the formula, is the loss function of the fine-tuning stage and the multi-scale optimization stage, is the first loss weight of the loss function in the fine-tuning stage and the multi-scale optimization stage, Set to 0.3, is the second loss weight of the loss function in the fine-tuning stage and the multi-scale optimization stage, Set to 0.7, Dynamic weights are training parameters that allow the model to optimize loss weights based on the specific characteristics of the data and enhance sensitivity to key spectral features. The number of spectral data entered for the current batch, is the kth real spectrum data, is the noise input corresponding to the kth real spectrum data, for Norm, the same as Euclidean distance calculation method, is used to measure the spectral distance between generated data and real data.

[0165] S306, repeat the above steps until a preset stop iteration condition is met. The preset stop iteration condition is reaching a preset maximum number of iterations. The preset maximum number of iterations is set to 1000 times, which means that the model training is completed.

[0166] S4. A satellite remote sensing target data recognition model is used for data recognition. The satellite remote sensing target data recognition model includes a spectral vector feature dimension reduction module and a remote sensing target data classification module, such as Figure 2 shown.

[0167] The spectral vector feature dimensionality reduction module of this embodiment uses an autoencoder algorithm to reduce the spectral vector features. Traditional autoencoders learn the compressed representation of input data through the framework of an encoder and a decoder, and then reconstruct the data through the decoding process. This embodiment adds a special nonlinear transformation layer to the encoder. The nonlinear transformation layer can strengthen the independence between features and reduce the redundant information between features, so that the model can capture more complex data structures.

[0168] The training process of the autoencoder algorithm is:

[0169] S4101. Initialize the weights and biases of the autoencoder model. The initialization method is expressed as:

[0170] ,

[0171] In the formula, is the initial value of the autoencoder weight, is the mean shift of the autoencoder weights, is the standard deviation of the autoencoder weight initialization, is a normal distribution with mean 0 and covariance matrix as the identity matrix, is a normal distribution, is the identity matrix; is the initial value of the autoencoder bias, is the mean shift of the autoencoder bias.

[0172] S4102. In the encoding stage, the input data is firstly subjected to preliminary feature extraction through multiple convolutional layers or fully connected layers, and then subjected to feature independence processing through a nonlinear transformation layer. The nonlinear transformation layer uses a special activation function and a constrained optimization algorithm to enhance the independence of features. The processing method of the nonlinear transformation layer is expressed as follows:

[0173] ,

[0174] In the formula, is the transformed feature vector, is the hyperbolic tangent function, which is used to increase the nonlinearity of the output characteristics. and are the weights and biases of the nonlinear transformation layer, is the input feature vector.

[0175] The weight update method of the nonlinear transformation layer is expressed as:

[0176] ,

[0177] In the formula, is the learning rate of the autoencoder, is the loss function of the autoencoder.

[0178] S4103. In the decoding stage, feature data is processed layer by layer. Each layer focuses on a specific set of feature dimensions, and uses the hierarchical feature deconstruction strategy to gradually restore the output close to the original data. The decoding method of each layer is expressed as:

[0179] ,

[0180] In the formula, is the reconstructed output vector, is the adaptive Sigmoid activation function, and are the weights and biases of the decoding layer, respectively. is the vector element multiplication symbol, is the feature reweighting function, is the reconstruction error vector of the features corresponding to the transformed feature vector.

[0181] Since some features may be more important to the final classification or regression task during the feature dimensionality reduction process, the feature reweighting function allows the model to adjust the importance of each feature during the training process to improve the target relevance and efficiency of feature dimensionality reduction. The calculation method of the feature reweighting function is expressed as:

[0182] ,

[0183] In the formula, is the adjustment coefficient of the feature reweighting function, which controls the sensitivity of the error impact, and nf is the number of features after dimensionality reduction; is the i-th element of the reconstruction error vector corresponding to the feature after the conversion, is the i-th element of the transformed eigenvector.

[0184] The parameters of the adaptive Sigmoid activation function are dynamically adjusted according to the reconstruction error, and the calculation method is expressed as:

[0185] ,

[0186] In the formula, x is the independent variable of the function, is the adaptive parameter of the activation function.

[0187] The calculation method of the adaptive parameters of the activation function is expressed as:

[0188] ,

[0189] In the formula, is the adjustment parameter of the initial activation function, Set to 0.3, is the adjustment factor of the activation function, for Norm.

[0190] S4104. Calculate the loss function based on the reconstruction error and feature independence index to ensure the validity and independence of the features during model training. The calculation method of the loss function is expressed as:

[0191] ,

[0192] In the formula, is the eigenvector The mutual information of is used to evaluate the independence of features. is the regularization parameter, Set to 0.3;

[0193] S4105. Repeat the above steps until a preset stop iteration condition is met. The preset stop iteration condition is reaching a preset maximum number of iterations, which means that the model training is completed.

[0194] The remote sensing target data classification module adopts the kernel extreme learning machine algorithm based on conjugate gradient compensation as the classifier model. The kernel extreme learning machine algorithm based on conjugate gradient compensation includes an input layer, a kernel transformation layer and an output layer. The input layer receives data from the feature dimensionality reduction model, the kernel transformation layer uses a kernel function to map the data to a high-dimensional space, and the output layer is responsible for generating the final classification decision. In the training process of the kernel extreme learning machine, this embodiment adopts the conjugate gradient method to optimize the weight update, improve the learning speed and accuracy, and at the same time, adopts the polynomial fitting method to better handle the complex relationship between nonlinear features and enhance the model's recognition ability for remote sensing target data classification.

[0195] The training process of the kernel extreme learning machine algorithm based on conjugate gradient compensation is:

[0196] S4201, using a polynomial kernel function as the kernel function of the kernel extreme learning machine, expressed as:

[0197] ,

[0198] In the formula, is the kernel function of the extreme learning machine, is the first input vector of the kernel extreme learning machine, is the second input vector of the kernel extreme learning machine, is the scaling parameter of the kernel function, To calculate the inner product operation, is the free term of the kernel extreme learning machine, and de is the degree of the polynomial.

[0199] The way to calculate the inner product in the kernel function is expressed as:

[0200] ,

[0201] in, and Respectively, vector and The i-th element of un is is the number of input features of the kernel extreme learning machine.

[0202] Initialize the weights of the kernel extreme learning machine. The initialization method is expressed as:

[0203] ,

[0204] In the formula, is the weight matrix of the kernel extreme learning machine, To generate standard normally distributed random numbers, is the number of input features of the kernel extreme learning machine, is the number of output categories of the kernel extreme learning machine.

[0205] S4202, map the input data to a high-dimensional space through a kernel function, and calculate the mapped features as follows:

[0206] ,

[0207] In the formula, is the feature mapping function of the kernel extreme learning machine, is the kernel function of the extreme learning machine, is the kernel function correlation diagonal matrix, whose diagonal elements are composition, It is the ui-th element of the kernel function correlation diagonal matrix, representing the weight of each feature.

[0208] An adaptive feature weight adjustment strategy is used to dynamically adjust the weights of input features, and the influence of each feature is enhanced or weakened according to its actual performance in the classification task. The initial value of each element of the kernel function correlation diagonal matrix is ​​set to 1, which means that all features have an equal impact on the classification result in the initial state.

[0209] In each iteration, the weight of the feature is automatically adjusted according to its contribution to the loss function, and the update method of the elements in the diagonal matrix of the kernel function correlation is expressed as:

[0210] ,

[0211] In the formula, is the ui-th element of the diagonal matrix of the kernel function correlation at the t+1-th iteration, is the ui-th element of the diagonal matrix of the kernel function correlation at the t-th iteration, is the diagonal matrix adjustment factor, Set to 0.01, is the loss function of the kernel extreme learning machine The partial derivative of .

[0212] S4203, using the conjugate gradient method optimization strategy to optimize the weight vector in the kernel space to reduce computing resource consumption and improve training efficiency, expressed as:

[0213] ,

[0214] In the formula, is the conjugate gradient of the t+1th iteration, is the gradient of the loss function of the kernel extreme learning machine with respect to the weights at the tth iteration, is the loss function of the kernel extreme learning machine, is the weight matrix of the kernel extreme learning machine at the tth iteration, is the learning rate adjustment coefficient of the kernel extreme learning machine at the tth iteration, is the conjugate gradient of the tth iteration;

[0215] is the weight matrix of the kernel extreme learning machine at the t+1th iteration, is the direction adjustment coefficient of the kernel extreme learning machine at the tth iteration.

[0216] The calculation method of the learning rate adjustment coefficient of the kernel extreme learning machine at the tth iteration is expressed as:

[0217] ,

[0218] In the formula, for norm, The gradient of the loss function of the kernel extreme learning machine at the tth iteration with respect to the weight is an indicator of the current gradient size and is used to adjust the search direction for the next step. is the gradient of the loss function of the kernel extreme learning machine with respect to the weights at the t-1th iteration, is the weight matrix of the kernel extreme learning machine at the t-1th iteration.

[0219] S4204, calculating the loss function, the loss function of the kernel extreme learning machine used is the logistic regression loss with a regularization term, and the calculation method is expressed as:

[0220] ,

[0221] In the formula, is the loss function of the kernel extreme learning machine, mfd is the total number of samples input in the current batch, is the true label of the jth input sample, is the Sigmoid activation function, is the transpose of the weight matrix of the kernel extreme learning machine, is the jth input sample, is the regularization parameter, Set to 0.2.

[0222] S4205, weights are updated through the back propagation algorithm and the conjugate gradient compensation mechanism. Each step of weight update is based on the conjugate gradient optimization result of the previous step, ensuring that each step of weight adjustment is carried out in the direction of reducing the total loss. The update method is expressed as:

[0223] ,

[0224] In the formula, is the weight matrix of the kernel extreme learning machine at the t+1th iteration, is the learning rate of the kernel extreme learning machine, Set to 0.01.

[0225] S4206. Use the trained model to make classification decisions and output the specific category of the remote sensing target data according to the characteristics of the input data. The classification function is expressed as:

[0226] ,

[0227] In the formula, is the true label of the sample, is the function used to convert the model output into classification labels.

[0228] The above description is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention.

Claims

1. A satellite remote sensing target data recognition method based on artificial intelligence, characterized in that: The following steps are involved: S1. By performing satellite remote sensing target recognition on the collected spectral vectors, the distributed federated learning model training architecture is adopted to realize distributed model training and data mining of multi-node joint models; The model training architecture of distributed federated learning is as follows: in each round of iteration, each node performs local model training and uploads the trained model parameters to the central server. The central server completes parameter aggregation and update, and sends the updated parameters to each node to start a new round of iteration. Repeating this iteration operation means that the global model of the central server and the model training of each node are completed. S2. Manually label the collected spectral vector data, and the labeling categories include: urban building target area, water body target area, vegetation target area, and bare land target area; S3. Generate samples using a generative adversarial network based on multi-scale spectral decomposition to achieve data expansion. The generative adversarial network algorithm based on multi-scale spectral decomposition includes two main components: a generator and a discriminator. The generator is used to generate data of a real spectral vector, and the discriminator is used to distinguish between real input data and input data generated by the generator. The training process of the generative adversarial network algorithm based on multi-scale spectral decomposition in step S3 is: S301, using a weight initialization strategy based on wavelet transform to initialize the generator and discriminator of the generative adversarial network, the initialization method is expressed as: ， In the formula, is the initialization value of the weight of the generated adversarial network, is the wavelet transform basis strategy function, is the scale parameter of wavelet transform; The wavelet transform basis strategy function is used to generate initial weights with multi-scale characteristics, and the calculation method is expressed as: ， In the formula, is the circumference of a circle, is the wave number of the wavelet function, which determines the oscillation frequency of the wavelet function. X is the independent variable of the function. is the phase shift of the wavelet function, which is used to adjust the starting phase of the wavelet. is an exponential function with a natural constant as base, is the standard deviation of the Gaussian window, controlling the locality of the wavelet, Set to 0.04; S302, perform rough tuning of the generative adversarial network. The generator and the discriminator are trained using a standard adversarial training framework. The loss function of the rough tuning stage is expressed as: ， In the formula, To combat the loss function, For the generator, is the discriminator, For the real data set, is a random noise distribution, For expectations, From the real data distribution The data sampled in To obey a specific distribution, is the distribution of real data, is the discriminator function, is the noise distribution, is the generator function, From the noise distribution The vector sampled in ; S303, fine-tune the generative adversarial network, use the piecewise linear mapping strategy of the spectral vector to adjust the generator in detail, the generator The output is realized through a piecewise linear function, expressed as: ， In the formula, is a piecewise linear mapping function responsible for input noise To transform, is the parameter set of the mapping function, md is the number of segments, md is set to 5, is the linear coefficient within the i-th segment, is the ReLU activation function, is the nonlinear function, is the slope of the i-th segment, is the offset within the i-th segment, is the quadratic term coefficient used to increase the model's ability to handle the nonlinear characteristics of spectral data. Set to 0.5; is the Sigmoid activation function, which is a nonlinear function used to further process the mapped data to approximate the real data distribution. is the adjustment coefficient that controls the angle adjustment range, is a sinusoidal function of the spectral angle, which is used to ensure that the adjustment amplitude is proportional to the angular deviation, is the spectral angle mapping function, is the target spectral vector, i.e., the reference vector selected from the real data set; A dynamic feature correction strategy based on spectral angle mapping is used to improve the spectral authenticity of the data generated in the data augmentation algorithm. A spectral angle mapping function is defined to calculate and adjust the angle of the generated spectral data. The calculation method is expressed as: ， In the formula, For input noise Piecewise linear mapping function for transformation ; S304, after the generator can reproduce the single-scale spectral data well, a multi-scale optimization strategy is adopted to optimize the generation performance of the generator at each scale by processing spectral data of different scales in parallel. The data generation method of the multi-scale optimization strategy stage is expressed as: ， In the formula, is a scale-specific processing function, is the weight matrix of the generator, is the number of scales for the multi-scale optimization strategy, is the nonlinear activation function of the jth scale, is the weight matrix of the jth scale; is the generator function of the multi-scale optimization stage, It is the concatenation of processing results at different scales. is the weight matrix of the first scale, For the The weight matrix of each scale; In order to adapt to the data processing requirements of different scales, the calculation method of the nonlinear activation function of the jth scale is expressed as: ， In the formula, x is the independent variable of the function, is the first adjustable parameter of the activation function, is the second adjustable parameter of the activation function; S305. In the fine-tuning stage and the multi-scale optimization stage, the loss function calculation method of the fine-tuning stage and the multi-scale optimization stage is expressed as: ， In the formula, is the loss function of the fine-tuning stage and the multi-scale optimization stage, is the first loss weight of the loss function in the fine-tuning stage and the multi-scale optimization stage, Set to 0.3, is the second loss weight of the loss function in the fine-tuning stage and the multi-scale optimization stage, Set to 0.7, Dynamic weights are training parameters that allow the model to optimize loss weights based on the specific characteristics of the data and enhance sensitivity to key spectral features. The number of spectral data entered for the current batch, is the kth real spectrum data, is the noise input corresponding to the kth real spectrum data, for Norm, calculated in the same way as Euclidean distance, is used to measure the spectral distance between generated data and real data; S306, repeating the above steps until a preset stop iteration condition is met, where the preset stop iteration condition is reaching a preset maximum number of iterations, indicating that the model training is completed; S4. A satellite remote sensing target data recognition model is used for data recognition. The satellite remote sensing target data recognition model includes a spectral vector feature dimension reduction module and a remote sensing target data classification module.

2. The method for identifying satellite remote sensing target data based on artificial intelligence according to claim 1, characterized in that: The spectral vector feature dimension reduction module in step S4 uses an autoencoder algorithm to reduce the dimension of the spectral vector feature; The training process of the autoencoder algorithm is as follows: S4101. Initialize the weights and biases of the autoencoder model. The initialization method is expressed as: ， In the formula, is the initial value of the autoencoder weight, is the mean shift of the autoencoder weights, is the standard deviation of the autoencoder weight initialization, is a normal distribution with mean 0 and covariance matrix as the identity matrix, is a normal distribution, is the identity matrix; is the initial value of the autoencoder bias, is the mean shift of the autoencoder bias; S4102. In the encoding stage, the input data is firstly subjected to preliminary feature extraction through multiple convolutional layers or fully connected layers, and then subjected to feature independence processing through a nonlinear transformation layer. The nonlinear transformation layer uses a special activation function and a constrained optimization algorithm to enhance the independence of features. The processing method of the nonlinear transformation layer is expressed as follows: ， In the formula, is the transformed feature vector, is the hyperbolic tangent function, which is used to increase the nonlinearity of the output characteristics. and are the weights and biases of the nonlinear transformation layer, is the input feature vector; The weight update method of the nonlinear transformation layer is expressed as: ， In the formula, is the loss function of the autoencoder, is the learning rate of the autoencoder; S4103. In the decoding stage, feature data is processed layer by layer. Each layer focuses on a specific set of feature dimensions, and uses the hierarchical feature deconstruction strategy to gradually restore the output close to the original data. The decoding method of each layer is expressed as: ， In the formula, is the reconstructed output vector, is the adaptive Sigmoid activation function, and are the weights and biases of the decoding layer, respectively. is the vector element multiplication symbol, is the feature reweighting function, is the reconstruction error vector of the feature corresponding to the transformed feature vector; In the process of feature dimensionality reduction, the feature reweighting function allows the model to adjust the emphasis on each feature during the training process, thereby improving the target relevance and efficiency of feature dimensionality reduction. The calculation method of the feature reweighting function is expressed as: ， In the formula, is the adjustment coefficient of the feature reweighting function, which controls the sensitivity of the error impact, and nf is the number of features after dimensionality reduction; is the i-th element of the reconstruction error vector corresponding to the feature after the conversion, is the i-th element of the transformed eigenvector; The parameters of the adaptive Sigmoid activation function are dynamically adjusted according to the reconstruction error, and the calculation method is expressed as: ， In the formula, x is the independent variable of the function, is the adaptive parameter of the activation function; The calculation method of the adaptive parameters of the activation function is expressed as: ， In the formula, is the adjustment parameter of the initial activation function, Set to 0.3, is the adjustment factor of the activation function, for norm; S4104. Calculate the loss function based on the reconstruction error and feature independence index to ensure the validity and independence of the features during model training. The calculation method of the loss function is expressed as: ， In the formula, is the eigenvector The mutual information of is used to evaluate the independence of features. is the regularization parameter, Set to 0.3; S4105. Repeat the above steps until a preset stop iteration condition is met. The preset stop iteration condition is reaching a preset maximum number of iterations, which means that the model training is completed.

3. The method for identifying satellite remote sensing target data based on artificial intelligence according to claim 1, characterized in that: The remote sensing target data classification module in step S4 adopts a conjugate gradient compensation-based kernel extreme learning machine algorithm as a classifier model. The conjugate gradient compensation-based kernel extreme learning machine algorithm includes an input layer, a kernel transformation layer, and an output layer. The input layer receives data from the feature dimensionality reduction model, the kernel transformation layer uses a kernel function to map the data to a high-dimensional space, and the output layer is responsible for generating a final classification decision. The training process of the kernel extreme learning machine algorithm based on conjugate gradient compensation is: S4201, using a polynomial kernel function as the kernel function of the kernel extreme learning machine, expressed as: ， In the formula, is the kernel function of the extreme learning machine, is the first input vector of the kernel extreme learning machine, is the second input vector of the kernel extreme learning machine, is the scaling parameter of the kernel function, To calculate the inner product operation, is the free term of the kernel extreme learning machine, de is the degree of the polynomial; The way to calculate the inner product in the kernel function is expressed as: ， in, and Respectively, vector and The i-th element of , nu is is the number of input features of the kernel extreme learning machine; Initialize the weights of the kernel extreme learning machine. The initialization method is expressed as: ， In the formula, is the weight matrix of the kernel extreme learning machine, To generate standard normally distributed random numbers, is the number of input features of the kernel extreme learning machine, is the number of output categories of the kernel extreme learning machine; S4202, map the input data to a high-dimensional space through a kernel function, and calculate the mapped features as follows: ， In the formula, is the feature mapping function of the kernel extreme learning machine, is the kernel function of the extreme learning machine, is the kernel function correlation diagonal matrix, whose diagonal elements are composition, is the ui-th element of the kernel function correlation diagonal matrix, representing the weight of each feature; Adaptive feature weight adjustment strategy is used to dynamically adjust the weight of input features, and the influence of each feature is enhanced or weakened according to its actual performance in the classification task. The initial value of each element of the kernel function correlation diagonal matrix is ​​set to 1, that is, , indicating that all features in the initial state affect the classification results equally; In each iteration, the weight of the feature is automatically adjusted according to its contribution to the loss function, and the update method of the elements in the diagonal matrix of the kernel function correlation is expressed as: ， In the formula, is the ui-th element of the diagonal matrix of the kernel function correlation at the t+1th iteration, is the ui-th element of the diagonal matrix of the kernel function correlation at the t-th iteration, is the diagonal matrix adjustment factor, Set to 0.01, is the loss function of the kernel extreme learning machine The partial derivative of S4203, using the conjugate gradient method optimization strategy to optimize the weight vector in the kernel space, expressed as: ， In the formula, is the conjugate gradient of the t+1th iteration, is the gradient of the loss function of the kernel extreme learning machine with respect to the weights at the tth iteration, is the loss function of the kernel extreme learning machine, is the weight matrix of the kernel extreme learning machine at the tth iteration, is the learning rate adjustment coefficient of the kernel extreme learning machine at the tth iteration, is the conjugate gradient of the tth iteration; is the weight matrix of the kernel extreme learning machine at the t+1th iteration, is the direction adjustment coefficient of the kernel extreme learning machine at the tth iteration; The calculation method of the learning rate adjustment coefficient of the kernel extreme learning machine at the tth iteration is expressed as: , In the formula, for norm, The gradient of the loss function of the kernel extreme learning machine at the tth iteration with respect to the weight is an indicator of the current gradient size and is used to adjust the search direction for the next step. is the gradient of the loss function of the kernel extreme learning machine with respect to the weights at the t-1th iteration, is the weight matrix of the kernel extreme learning machine at the t-1th iteration; S4204, calculating the loss function, the loss function of the kernel extreme learning machine used is the logistic regression loss with a regularization term, and the calculation method is expressed as: ， In the formula, is the loss function of the kernel extreme learning machine, mfd is the total number of samples input in the current batch, is the true label of the jth input sample, is the Sigmoid activation function, is the transpose of the weight matrix of the kernel extreme learning machine, is the jth input sample, is the regularization parameter, Set to 0.2; S4205, weights are updated by back propagation algorithm and conjugate gradient compensation mechanism, and the updating method is expressed as: ， In the formula, is the weight matrix of the kernel extreme learning machine at the t+1th iteration, is the learning rate of the kernel extreme learning machine, Set to 0.01; S4206. Use the trained model to make classification decisions and output the specific category of the remote sensing target data according to the characteristics of the input data. The classification function is expressed as: ， In the formula, is the true label of the sample, is the function used to convert the model output into classification labels.

4. The method for identifying satellite remote sensing target data based on artificial intelligence according to claim 1, characterized in that: The distributed federated learning model training architecture in step S1 adopts the federated average algorithm as the model aggregation algorithm. The federated average algorithm realizes the collaborative training of local models through multiple global iterations. For each global iteration, the number of nodes is set to The total number of samples is , and the number of samples of the kth node is , then the objective function of federated learning is defined as: ， In the formula, is the objective function of federated learning, are model parameters, is the model parameter Loss prediction for the jth sample data; is the training loss, the training loss The cross entropy loss function is used for calculation. is the sample feature of the jth sample data, The label of the jth sample data.

5. The method for identifying satellite remote sensing target data based on artificial intelligence according to claim 1, characterized in that: The multiple nodes in step S1 include k nodes. For the kth node, the objective function of the node is defined as: ， In the formula, is the number of samples of the kth node, is the objective function of the kth node, is the data distribution of the kth node; At the tth iteration, let the parameter gradient of the kth node be , then the way the model updates parameters at the tth iteration is expressed as: ， In the formula, is the model parameter of the t+1th iteration, is the model parameter of the tth iteration, is the learning rate for the current parameter update, is the number of samples of the kth node; The parameter update method of the global model of the central server is expressed as: ， In the formula, are the parameters of the global model of the central server at the t+1th iteration.

Images