CN114842328B

CN114842328B - Hyperspectral change detection method based on collaborative analysis autonomous perception network structure

Info

Publication number: CN114842328B
Application number: CN202210283172.9A
Authority: CN
Inventors: 侍佼; 谭春晖; 雷雨; 周德云; 周颖
Original assignee: Northwestern Polytechnical University
Current assignee: Northwestern Polytechnical University
Priority date: 2022-03-22
Filing date: 2022-03-22
Publication date: 2024-03-22
Anticipated expiration: 2042-03-22
Also published as: CN114842328A

Abstract

The invention discloses a hyperspectral change detection method based on collaborative analysis autonomous perception network structure, which comprises the following steps: acquiring two groups of hyperspectral change detection data sets acquired by the same sensor, and determining a training sample and a sample to be tested of each group; a network structure search task is detected for each group of established changes; generating an initial network structure population for each task; carrying out intra-population evolution on each current network structure population by utilizing a genetic algorithm to obtain an internal evolved network structure population; aiming at the pair of internal evolutionary network structure populations, excellent individual information is shared when the cross-task knowledge communication condition is met, and updated network structure populations corresponding to each internal evolutionary network structure population are obtained; judging whether the preset maximum iteration times are reached or not; if not, returning to the population for evolution; if the iteration is stopped, an optimal network structure is obtained from each updated network structure population, and a corresponding sample to be tested is utilized to obtain a change detection result. The invention can improve the classification precision of each CD task.

Description

Hyperspectral change detection method based on collaborative analysis autonomous perception network structure

Technical Field

The invention belongs to the technical field of image processing, and particularly relates to a hyperspectral change detection method based on collaborative analysis of an autonomous perception network structure.

Background

Change Detection (CD) is an important application of remote sensing technology to identify significant changes between two-phase images, and can quantitatively analyze and determine the characteristics and Change process of surface changes. Currently, remote sensing image change detection has been applied in a variety of different fields, such as farmland detection, disaster assessment, and military strikes. The hyperspectral images (Hyperspectral Images, HSIs) can provide more abundant spectrum information than other remote sensing images, and have the potential of distinguishing fine spectrum differences, so that the change detection of the hyperspectral images has become a research hot spot. However, the high dimensional data of HSIs makes the detection of changes in hyperspectral images a challenging problem.

Conventional methods of change detection, such as iterative re-weighted multivariate (Iteratively Reweighted Mad, IR-MAD), time-principal component analysis (Temporal Principal Components Analysis, TPCA), and change vector analysis (Change Vector Analysis, CVA), provide some reliable insight for measuring change correlation, but these methods tend to compress high-dimensional data into one or more dimensions of data, resulting in loss of some important information, limiting the ability to mine change features.

The depth network has strong modeling capability, can well characterize the relationship between the image and the surface characteristics, and has strong capability of processing high-dimensional data, and in recent years, a plurality of students begin to research on using the depth network for detecting changes. For example, du et al propose a change detection method based on a deep network and slow feature analysis (Deep Network and Slow Feature Analysis, DSFA), indicating that the image transformation can effectively highlight the changing information. In order to better deal with the high-Dimensional problem of HSIs data, a General End-to-End Two-Dimensional convolutional neural network (GETNET) framework is proposed for the change detection of hyperspectral images. In addition, saha et al propose a CD method using an untrained model and further use depth variation vector analysis to compare extracted features.

However, these methods analyze only one specific data set at a time, perform one CD task, and start performing CD tasks from the zero-basis state of knowledge, and cannot efficiently mine useful information between different data sets. In addition, conventional CD networks are typically designed by human experts, requiring a significant amount of time. The use of a fixed network architecture in processing multiple data sets can result in varying effects in detecting the variation of different data sets.

Disclosure of Invention

In order to solve the problems in the prior art, the invention provides a hyperspectral change detection method based on collaborative analysis of an autonomous perception network structure. The technical problems to be solved by the invention are realized by the following technical scheme:

a hyperspectral change detection method based on collaborative analysis of an autonomous perceived network structure comprises the following steps:

acquiring two groups of hyperspectral change detection data sets acquired by the same sensor, and determining a training sample and a sample to be tested in each group of hyperspectral change detection data sets; each group of hyperspectral change detection data sets contains hyperspectral remote sensing image data aiming at the front moment and the rear moment of the same area;

respectively establishing corresponding change detection network structure search tasks for each group of hyperspectral change detection data sets; aiming at each change detection network structure searching task, generating a corresponding initial network structure population by utilizing a preset individual gene coding mode; wherein the initial network structure population contains a preset number of individuals for representing different neural network structures;

aiming at each current network structure population, carrying out intra-population evolution by utilizing a genetic algorithm to obtain a corresponding intra-evolved network structure population; wherein, when iterating for the first time, the current network structure population is the initial network structure population;

Aiming at the pair of internal evolutionary network structure populations, when the cross-task knowledge communication condition is met, sharing excellent individual information based on the cross-task knowledge communication to obtain an updated network structure population corresponding to each internal evolutionary network structure population;

judging whether the current iteration number reaches a preset maximum iteration number or not;

if not, returning to the step of carrying out intra-population evolution by utilizing a genetic algorithm aiming at each current network structure population;

if yes, stopping iteration, acquiring an optimal network structure from each updated network structure population, and obtaining a change detection result by using a corresponding sample to be detected.

The invention has the beneficial effects that:

the method can execute respective CD tasks for two hyperspectral change detection data sets simultaneously, and based on the intra-population evolution of each task, the inter-population collaborative analysis sharing of excellent individual information is carried out by means of inter-task knowledge exchange, so that useful information among different data sets can be effectively mined, an autonomous perception optimal network structure is realized for each task, and a proper network structure is sought for each CD task to carry out change detection, so that the classification precision and the classification efficiency of the change detection can be improved for each CD task.

The present invention will be described in further detail with reference to the accompanying drawings and examples.

Drawings

Fig. 1 is a schematic flow chart of a hyperspectral change detection method based on collaborative analysis autonomous perception network structure according to an embodiment of the present invention;

FIG. 2 is a schematic diagram of a gene encoding process according to an embodiment of the present invention;

FIG. 3 is a schematic diagram of a non-aligned crossover operation according to an embodiment of the present invention;

FIG. 4 is a schematic diagram illustrating a random embedded variation operation according to an embodiment of the present invention;

FIG. 5 is a schematic diagram of a cross-task knowledge communication mechanism according to an embodiment of the present invention;

FIG. 6 is a Farm hyperspectral dataset used in experiments in accordance with embodiments of the present invention; fig. 6 (a) is an image taken on 3 th month 5 of 2006; fig. 6 (b) is an image taken on day 23 of month 4 of 2007; fig. 6 (c) is a change detection reference diagram of fig. 6 (a) and 6 (b);

FIG. 7 is a test dataset River hyperspectral dataset used in experiments in accordance with embodiments of the present invention; fig. 7 (a) is an image taken on 3 days of 2013, 5; fig. 7 (b) is an image taken on the 31 th 12 th 2013 month; fig. 7 (c) is a change detection reference diagram of fig. 7 (a) and 7 (b);

FIG. 8 is a graph comparing the results of detection of changes in the Farm dataset using different algorithms, FIG. 8 (a) using the CVA algorithm; FIG. 8 (b) employs a DSFA algorithm; FIG. 8 (c) employs the GETNET algorithm; FIG. 8 (d) employs the method of the present invention; FIG. 8 (e) is a change detection reference diagram of the Farm dataset;

FIG. 9 is a graph comparing the results of detection of changes in River datasets using different algorithms, FIG. 9 (a) using the CVA algorithm; FIG. 9 (b) employs a DSFA algorithm; FIG. 9 (c) employs the GETNET algorithm; FIG. 9 (d) employs the method of the present invention; FIG. 9 (e) is a variation detection reference diagram of River dataset;

FIG. 10 is a comparison experiment result of an SPNA-CA method and a single-task SPNA method based on collaborative analysis of an autonomous perceived network structure according to an embodiment of the present invention; FIG. 10 (a) shows fitness evaluation contrast; FIG. 10 (b) shows a classification accuracy evaluation comparison;

FIG. 11 is a graph showing the adaptive experience sharing probability P according to the embodiment of the present invention _es Is determined according to the analysis result of the (a); FIG. 11 (a) shows P _es A change in the optimization process; FIG. 11 (b) shows a different P _es A change in the lower loss function value.

Detailed Description

The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

The embodiment of the invention provides a hyperspectral change detection method based on a collaborative analysis autonomous perception network structure (Collaborative Analysis Framework with Self-Perception Network Architecture, SPNA-CA). The execution subject of the method can be a hyperspectral change detection device based on collaborative analysis autonomous perception network structure, and the corresponding device can be operated in electronic equipment. The electronic device may be a server or a terminal device, but is not limited thereto.

As shown in fig. 1, the hyperspectral change detection method based on collaborative analysis autonomous perception network structure provided by the embodiment of the invention may include the following steps:

s1, acquiring two groups of hyperspectral change detection data sets acquired by the same sensor, and determining a training sample and a sample to be detected in each group of hyperspectral change detection data sets.

In the embodiment of the invention, two groups of hyperspectral change detection data sets { D } acquired by the same sensor are acquired ₁ ,D ₂ Wherein each hyperspectral change detection dataset contains hyperspectral remote sensing image data for the front and back time instants of the same region, can be used as I ₁ And I ₂ And (3) representing.

Since the hyperspectral remote sensing image datasets obtained by the same sensor have the same spectral range and similar spectral-spatial structure, they represent the radiation or reflectivity of the ground target. Therefore, the blindness of searching an optimal network structure by one data set can be reduced by utilizing the public knowledge between two groups of hyperspectral change detection data sets acquired by the same sensor, the balance of global search and local search can be realized to a certain extent, the difficult problem of high-efficiency utilization of priori knowledge and process knowledge is relieved, and the effective resource allocation is realized.

In the embodiment of the invention, a training sample and a sample to be tested are selected from each group of hyperspectral change detection data sets, wherein the training sample can be used for determining a sample label by utilizing a certain change detection means and is used for subsequent network training; and the sample to be tested is used as a detection object formally sent into the network after training is completed and is used for obtaining a change detection result.

In an alternative embodiment, determining training samples and samples to be tested in each set of hyperspectral variation detection data sets includes:

(1) And respectively carrying out pre-change detection on each group of hyperspectral change detection data sets to obtain corresponding pre-change detection results.

Detecting data set D with a set of hyperspectral changes ₁ ＝{I ₁ ,I ₂ For example, for D ₁ I of (2) ₁ And I ₂ D was obtained using the existing K-means (K-means) method ₁ The pre-change detection result of (2) is obtained to obtain a pre-change detection result graph R. It will be appreciated that the pre-change detection of each pixel in R indicates I ₁ And I ₂ Whether two pixels at the same position in the image change.

The specific change detection process belongs to the prior art and is not described here.

(2) According to the pre-change detection result corresponding to each hyperspectral change detection data set, training samples in the hyperspectral change detection data set are selected, and samples except the training samples in the hyperspectral change detection data set are used as samples to be tested.

In an alternative embodiment, selecting the training samples in each set of hyperspectral variation detection data sets according to the pre-variation detection result corresponding to the hyperspectral variation detection data sets includes:

for each pre-change detection result, for each pixel in the image data of the pre-change detection result, taking the pixel as a central pixel of a rectangular window with a preset size, and judging whether the pre-change detection results of all pixels in the rectangular window are consistent.

If so, splicing the two pixels in the corresponding hyperspectral change detection data set by using the positions of the pixels to obtain a training sample in the corresponding hyperspectral change detection data set.

Specifically, data set D is detected as a set of hyperspectral changes ₁ ＝{I ₁ ,I ₂ For example, for D ₁ And (3) sliding and traversing each pixel in R by using a rectangular window with a preset size, so that the central pixel of the rectangular window is sequentially traversed each pixel. The preset size may be 3×3, 5×5, etc., and may be specifically selected as required. For ease of understanding, the following description is given in terms of 3×3.

For one pixel (i, j), a pixel composed of 9 pixels within a rectangular window of 3×3 size having the pixel as a center is determined Block M _ij If the pre-change detection results of the 9 pixels are consistent, i.e. if the pre-change detection results of the 9 pixels are all "changed", or if the pre-change detection results of the 9 pixels are all "unchanged", the pixel (i, j) is selected as a training sample, and the sample label of the training sample, i.e. the classification label, is consistent with the pre-change detection results of the 9 pixels.

Wherein I is ₁ And I ₂ Is set to be r×c×b, each pixel is extracted separately, and the pixel (I, j) in R corresponds to I ₁ The pixels in (a) are marked asThe size is 1×1×b. Pixel (I, j) in R corresponds to I ₂ The pixels in (a) are marked->The size is 1×1×b. The two pixels are +.>And->Splicing to obtain training sample m _ij The size is 1×1×2b.

It will be appreciated that hyperspectral change detection dataset D ₁ The samples other than the training samples are samples to be tested, that is, if a pixel (i, j) in R is determined to be within a rectangular window of 3×3 size with the sample as the center pixel, a pixel block M of 9 pixels _ij If the pre-change detection results of (i) and (j) are inconsistent, the pixel (i, j) is selected as a sample to be detected, and the sample to be detected also needs to be spliced by the two pixels, which is not repeated here.

It should be noted that, for the edge pixel (i, j) in R, there are no 9 pixels in a rectangular window with a size of 3×3, which uses the edge pixel as a center pixel, and the embodiment of the present invention directly uses the pixel as a sample to be tested.

S2, respectively establishing corresponding change detection network structure search tasks aiming at each group of hyperspectral change detection data sets; and aiming at each change detection network structure searching task, generating a corresponding initial network structure population by utilizing a preset individual gene coding mode.

Embodiments of the present invention detect data sets { D for two sets of hyperspectral changes ₁ ,D ₂ Respectively establishing corresponding change detection network structure search tasks { T }, respectively ₁ ,T ₂ }, wherein D ₁ And T ₁ Corresponding to the above; d (D) ₂ And T ₂ Corresponding to the above. Each change detection network structure search task aims to seek an optimal network structure for the set of hyperspectral change detection datasets for subsequent change detection.

For easy understanding, the inventive concept of the embodiments of the present invention will be described first.

The conventional hyperspectral change detection method detects the change condition of all hyperspectral data sets by using a fixed network structure, so that the detection accuracy of the conventional hyperspectral change detection method on part of the data sets is limited. And the traditional change detection method only analyzes one specific data set at a time, performs a change detection task, and performs the change detection task from a knowledge zero point, so that useful information between hyperspectral data sets cannot be effectively mined.

However, the inventor of the embodiment of the invention researches that HSIs spectrum information obtained by the same sensor has similar physical meaning (brightness or reflectivity), so that the possibility of knowledge sharing exists between different data sets of the same sensor.

Thus, the inventors studied to obtain: multiple data set collaborative analysis may better explore the inherent relevance of similar tasks than single data set analysis. The collaborative analysis framework facilitates the flow of information through knowledge sharing among tasks. In a single task, individuals gradually improve their quality through heuristic searches accumulated by their own experience during the internal task optimization process. Under the collaborative analysis framework, the network structure search is not realized by means of individual heuristic search, but a cross-task knowledge communication mechanism is adopted to promote collaborative cognition among groups, and local autonomous learning is utilized to guide global optimization. The collaborative analysis framework effectively utilizes knowledge to guide the structure searching process, reduces blindness of single task searching, can realize balance of global searching and local searching to a certain extent, relieves the difficult problem of high-efficiency utilization of priori knowledge and process knowledge, and realizes effective resource allocation.

The embodiment of the invention provides a feasible solution for solving the construction problem of a plurality of hyperspectral image change detection structures. The specific formula is as follows:

where x is the feasible network structure solution in the unified search space Ω, T _i Representing a corresponding hyperspectral dataset D _i Is a hyperspectral change detection network structure search task. In the collaboration process, a "knowledge bridge" may be established that allows knowledge communication across data sets, thereby improving the performance of the change detection network structure search task for each data set. In the embodiment of the invention, two groups of hyperspectral change detection data sets are used for constructing corresponding change detection network structure search tasks to implement.

Since the search task of the change detection network structure is to be performed for each set of hyperspectral change detection data, it is necessary to determine a starting point of the search range, i.e. the initial network structure population.

The initial network structure population contains a preset number of individuals used for representing different neural network structures. The preset number can be N _p Representation, N _p Is a natural number greater than 0.

In an alternative embodiment, for each change detection network structure search task, a corresponding initial network structure population is generated by using a preset individual gene coding mode, including:

Detecting a network structure search task aiming at each change, and randomly generating a positive integer n within a preset network layer number range; and randomly generating n positive integers in a preset intra-layer neuron number range to form a one-dimensional vector serving as an initial network structure, and forming an initial network structure population corresponding to the change detection network structure searching task by the obtained preset number of initial network structures.

The preset individual gene coding mode comprises coding the depth of a hidden layer and the number of neurons of each layer in the neural network.

It will be appreciated by those skilled in the art that for a neural network structure, the layers other than the input and output layers are considered hidden layers.

In the embodiment of the present invention, the number of neurons in the input layer of the neural network has been automatically set to adapt to different data sets, for example, the number of neurons in the input layer is 396 when the "River" (River) data set photographed by an Earth observation No. 1 (Earth serving-1, EO-1) sensor has 198 spectral bands. That is, for each data set, the corresponding number of neurons in the input layer may be automatically configured for the current data set according to the predetermined correspondence between the data set and the number of neurons in the input layer, so that the corresponding number of neurons in the input layer is a determined value in the implementation process of the embodiment of the present invention.

In the embodiment of the present invention, the number of neurons of the output layer is set to the classification number of the change detection, for example, for the classification problem of the change detection, the classification number is 2, and the number of neurons of the output layer is set to 2.

The embodiment of the invention specifically designs a gene coding method to express the uncertain depth and the neuron number of each hidden layer in the neural network. The preset individual gene coding mode only codes the hidden layers, and the number of layers (namely depth) of the network formed by all the hidden layers and the number of neurons of each layer are reflected.

Specifically, a search task T is performed for a change detection network structure ₁ And T ₂ Firstly, randomly generating a positive integer n within a preset network layer number range; randomly generating n positive integers x within the preset number range of the neurons in the layer ₁ ,x ₂ ,...,x _n Form a one-dimensional vector x= (X) ₁ ,x ₂ ,...,x _n ) As a oneInitial network structure. That is, the embodiment of the invention takes the number sequence corresponding to the one-dimensional vector as the coding sequence of a network structure, namely, one individual is one coding sequence. If an individual coding sequence is decoded, a network structure related to all hidden layers can be restored, and an input layer and an output layer are further added, so that a complete network structure can be obtained; repeating the above process N _p Next, N can be obtained _p And the initial network structures jointly form an initial network structure population corresponding to the change detection network structure searching task.

By the above treatment, T can be obtained ₁ And T ₂ The initial network structure population of (a) is P respectively _{0_1} And P _{0_2} 。

The preset network layer number range may be [1,6] or the like, for example. The preset number range of neurons in the layer may be [7,1024], for example. Can be specifically designed according to the needs.

The individual gene coding mode of the embodiment of the invention can reduce a part of coding data volume because the coding process of an input layer and an output layer is omitted. For a preset individual gene encoding mode, please refer to fig. 2, fig. 2 is a schematic diagram illustrating a gene encoding process according to an embodiment of the invention. Individual, i.e. coding sequence H ₁ ,H ₂ ,H ₃ ,H ₄ ,...,H _N The encoded length N represents the number of layers of the hidden layer, each encoding position H _i ；i∈[1,N]The number of neurons on the corresponding hidden layer. It will be appreciated that the lengths of the coding sequences may not be the same for different individuals.

S3, aiming at each current network structure population, carrying out intra-population evolution by utilizing a genetic algorithm to obtain a corresponding intra-evolved network structure population.

Wherein, when iterating for the first time, the current network structure population is the initial network structure population, namely P _{0_1} And P _{0_2} 。

In an alternative embodiment, S3 may include the steps of:

s31, aiming at each current network structure population, carrying out network structure performance evaluation on each individual by utilizing a training sample of the current network structure population to obtain an evaluation value corresponding to the individual.

For ease of understanding, a current network structure population P is illustrated, which contains N _p Individual. In order to select excellent individuals as parents for genetic manipulation, N in P is required to be selected _p And evaluating the network structure performance corresponding to each individual.

The embodiment of the invention can adopt any existing network performance evaluation index, such as network loss function value based, verification accuracy based and the like. The calculation process of the specific evaluation index belongs to the prior art, and is not described in detail here.

In an alternative implementation, this step may be implemented using a hybrid criterion network structure evaluation strategy designed by an embodiment of the present invention. The evaluation strategy comprehensively considers the change detection performance and the structural complexity of the network structure in the network structure evaluation link, and is used for excavating a high-performance and light-weight network structure.

Specifically, when the mixed criterion network structure evaluation strategy designed by the embodiment of the invention is adopted, the process of evaluating the network structure performance of an individual to obtain a corresponding evaluation value comprises the following steps:

Calculating a corresponding evaluation value for an individual by using a preset fitness evaluation function, wherein the fitness evaluation function is as follows:

F(x _i )＝λ ₁ L _t (x _i )+λ ₂ L _p (x _i ) (2)

wherein F (x) _i ) Representing an fitness evaluation function; x is x _i Representing an individual; l (L) _t (x _i ) The representation is based on individuals x _i The constructed network trains and loses; l (L) _p (x _i ) Representing network parameter loss; lambda (lambda) ₁ Represents L _t (x _i ) Weights of (2); lambda (lambda) ₂ Represents L _p (x _i ) Is a weight of (2). Lambda (lambda) ₁ And lambda (lambda) ₂ In one alternative embodiment, the sum of (2) is 1, and the sum of (2) and (2) can be 0.8 and 0.2 respectively, although they can be as desiredThe remaining values are set.

N represents the number of training samples; b _j A true class label representing the jth training sample; a, a _j Representing a j-th training sample in the training sample set; w represents the weight parameter set of the tested network structure and can be understood as an array; f (a) _j W) represents the network structure under test versus sample a _j Is predicted by the computer; i.e., predict whether it is "changed" or "unchanged"; lambda represents a regularization parameter, and the value range of lambda is (0, 1); b andrepresenting the true value and the predicted value, respectively.

Wherein,is an estimate of the currently tested network trainable parameters; n is n _i Representing the number of neurons in the ith layer; k represents the number of layers of the current network under test; p (P) _max Is the trainable parameter number of the network with the maximum neuron number and the maximum layer number; n is n _m Representing the maximum number of neurons, which may be 1024; k represents the maximum number of layers and may be 6.

In the embodiment of the invention, the network structure performance evaluation of an individual needs to configure network structure search parameters for the individual. And decoding the coding sequences corresponding to the individuals to obtain a complete network structure comprising an input layer and an output layer. And then according to the regulation for each obtained complete network structureTraining with training sample, and calculating fitness evaluation function F (x) according to mixed criterion network structure evaluation strategy _i ) I.e. the evaluation value.

As an example, a network structure search parameter configured for an embodiment of the present invention may be shown in table 1.

TABLE 1

The specific process of obtaining the evaluation value of the individual by training and predicting the complete network structure will not be described in detail here.

It will be appreciated that, through S31, N in the current network structure population P _p Individual individuals obtained respective evaluation values. The higher the evaluation value, the better the performance of the network structure corresponding to the individual.

S32, selecting partial individuals with the evaluation value ranking in front from the current network structure population as parent population, and obtaining a plurality of child individuals by utilizing the pairwise cross mutation operation.

First, N in the current network structure population P _p Individuals are ranked from high to low according to the evaluation value, and N ranked at the top is selected _mp Placing individual into mating pool MP, wherein N _mp ＜N _p . To improve the diversity of the population to a certain extent, two individuals are randomly selected from the mating pool MP as parent individuals p ₁ ,p ₂ And performing cross mutation operation. Specifically, for two randomly selected individuals p ₁ ,p ₂ The following treatments were performed:

1) Will p ₁ ,p ₂ Obtaining two new individuals s by using non-aligned crossover operation ₁ ,s ₂ 。

The method comprises the following steps:

for p ₁ ,p ₂ The corresponding coding sequences of the network structures are respectively X _i ,X _j ,(i≠j)。

First, a random number n is generated _align ∈[1,L _max -L _min ]Wherein L is _max Is two individuals p ₁ ,p ₂ The coding length L of an individual with longer coding sequence length _min Then the code length of another individual.

Second, this random number n will be selected _align As alignment positions, two individuals p ₁ ,p ₂ Aligned in an aligned position.

Next, a new random number n is generated _cnum ∈[1,L _min ]As the number of individuals crossed.

Then, a new random number n is generated _cross E (0, 1) is taken as the cross probability to be measured, when n _cross ＞P _c In this case, individual crossing is performed according to the alignment position and the number of crossed individuals to generate a new individual s ₁ ,s ₂ . Wherein P is _c The crossover probability set when searching the parameter configuration for the network structure is, for example, 0.3 in table 1.

Referring to fig. 3 for a non-aligned cross operation, fig. 3 is a schematic diagram illustrating a non-aligned cross operation according to an embodiment of the present invention. In figure 3 p is distinguished by different colours ₁ And p ₂ Is a coded number of (a). As can be seen from fig. 3, p after the misalignment crossing ₁ H in (1) ₃ And p is as follows ₂ H in (1) ₄ Exchanging; p is p ₁ H in (1) ₄ And p is as follows ₂ H in (1) ₅ Exchange to finally obtain new individual s ₁ Sum s ₂ 。

It will be appreciated that the operation of step 1) is performed for both individuals in the pool MP who randomly take out pairs until all individuals in the pool participate in completing step 1), a plurality of new individuals s are obtained ₁ ,s ₂ ...。

It should be noted that when n _cross ≤P _c When the two individuals do not cross treatment, the two individuals directly enter the subsequent mutation operation.

2) And respectively carrying out random embedding mutation operation on all new individuals obtained through the cross operation to obtain a plurality of offspring individuals.

The method comprises the following steps:

for a plurality of new individuals s obtained in step 1) ₁ ,s ₂ ...。

First, a random number n is generated _{m_ind} ∈[1,N _mp ]By using the random number n _{m_ind} One new individual is selected from the plurality of new individuals at a time as a variant individual. Wherein, the selection mode is that the numerical value of the random number is the serial number of the new individual.

Next, a random number n is generated for the selected variant _{m_point} ∈[1,L _{m_ind} ]Wherein L is _{m_ind} Selecting n for the length of the gene code of the variant _{m_point} The corresponding node serves as a location where a mutation may occur.

Then, for the selected position with variation, a random number between (0, 1) is generated to judge whether it is greater than the preset variation probability P _m If yes, judging that mutation occurs; if not, judging that the mutation does not occur, stopping the current operation on the individuals, and performing the cross mutation operation of a new pair of individuals. If variation occurs, a random number n is generated _{m_method} ∈[1,3]And determining the mutation mode, and executing corresponding mutation operation to generate the child individuals of the new individual.

There are three variation modes of random embedding variation operation. When n is _{m_method} When the variation mode is modification of the selected unit in time 1, modifying the number of neurons of the node; when n is _{m_method} When the variation mode is to add the selected unit in the time of=2, a node is added after the node; when n is _{m_method} When the expression of variation is deletion of the selected unit in time=3, the node is deleted.

For the random embedding mutation operation, please refer to fig. 4, fig. 4 is a schematic diagram of the random embedding mutation operation according to an embodiment of the present invention. The three variation modes are respectively: (1) modifying the node, i.e. modifying the selected cell; (2) adding nodes, i.e., adding selected cells; (3) deleting a node, i.e., deleting a selected cell. In fig. 4, different variation nodes are distinguished by a certain gray scale.

It will be appreciated that the number of components,for a plurality of new individuals s obtained in step 1) ₁ ,s ₂ .... Each new individual can be mutated to generate a child-generation individual according to the random embedded mutation operation. At this time, all the resulting offspring individuals can be represented as o ₁ ,o ₂ ...。

S33, selecting a preset number of excellent individuals from a set obtained by adding a plurality of child individuals into the current network structure population, and obtaining an internal evolved network structure population corresponding to the current network structure population.

Specifically, the current network structure population P is added into all child individuals o obtained by random embedded mutation operation ₁ ,o ₂ .., a collection of individuals of the population is obtained. It can be understood that, compared with the original population size, the number of individuals in the set is increased, in order to keep the population size unchanged, each individual in the set can be subjected to network structure performance evaluation to obtain a corresponding evaluation value, and N with higher evaluation value is reserved _p And obtaining the internal evolved network structure population corresponding to the current network structure population P by excellent individuals.

For example, in an alternative embodiment, S33 may include:

and (3) performing network structure performance evaluation on each individual in a set obtained by adding a plurality of child individuals into the current network structure population to obtain a corresponding evaluation value.

All evaluation values are ranked from high to low, and N with the top ranking is selected _p The individuals constitute an internally evolved network structure population corresponding to the current network structure population P.

For the process of evaluating the performance of the network structure for an individual to obtain the corresponding evaluation value, please refer to the related description above, and the description is not repeated here.

Alternatively, in another embodiment, S33 may include:

s331, generating a current elimination random number, and judging whether the current elimination random number is smaller than a preset elimination probability.

Specifically, the implementation mode adopts an iterative elimination mode, and each timeIteration is carried out, and a obsolete random number r, r epsilon (0, 1) aiming at the current time is generated. And judging whether the current elimination random number r is smaller than the preset elimination probability P _D . Wherein P is _D The configuration may be set at the time of network structure search parameter configuration, for example, may be 0.5 or the like.

If so, S332 is performed to eliminate the longest-existing individuals from the current population of network structures.

Specifically, if r < P _D And eliminating one body with the longest existing time from the current network structure population P so as to prevent the body from falling into local optimum and promote the global property of evolution.

If not, S333 is executed to eliminate the individual with the lowest evaluation value from the plurality of offspring individuals.

Specifically, if r.gtoreq.P _D From a plurality of offspring individuals o ₁ ,o ₂ .. the one with the lowest evaluation value is eliminated.

S334, judging whether the number of the individuals reserved currently reaches the preset number.

Whether the elimination process of S332 or S333 is performed, one individual in the set is eliminated after one iteration is finished, and at this time, it is required to determine whether the number of currently reserved individuals reaches N _p If the instruction population size is restored, continuing iteration is not needed; if not, the subject needs to be eliminated continuously for further iteration.

If not, returning to the step of generating the current obsolete random number.

Specifically, if the number of individuals currently remaining does not reach N _p If yes, the process returns to step S331.

If so, S335 is performed, wherein the current retained individuals form an internally evolved network structure population corresponding to the current network structure population.

Specifically, if the number of individuals currently remaining reaches N _p And if so, the individuals form an internal evolutionary network structure population corresponding to the current network structure population P.

It can be appreciated that, through step S3, both current network structure populations undergo an intra-population evolution to obtain a corresponding intra-evolved network structure population, i.e., a pair of intra-evolved network structure populations.

S4, aiming at the pair of internal evolutionary network structure populations, when the cross-task knowledge communication condition is met, excellent individual information is shared based on the cross-task knowledge communication, and the updated network structure population corresponding to each internal evolutionary network structure population is obtained.

This step is based on network structure searching across task knowledge exchange mechanisms. In the embodiment of the invention, the cross-task knowledge exchange mechanism shares the probability P according to self-adaptive experience _es The method is carried out, and excellent individuals with good fitness in different tasks are selected to share in other tasks. In order to solve the problem of network structure design tasks on multiple data sets at the same time, the embodiment of the invention establishes a collaborative analysis framework to model related change detection tasks, and performs joint analysis by utilizing the correlation among the data. Different network structure search tasks are established in a targeted manner for each data set, wherein an evolutionary multitasking self-aware network structure method is designed for exploring effective and reasonable network structures among a plurality of data sets.

For convenience of subsequent description, the current pair of internally evolved network structure populations is denoted by P ¹ ,P ² And (3) representing.

The judging process of whether the network structure population after internal evolution meets the cross-task knowledge communication condition comprises the following steps:

And calculating an adaptive experience sharing probability value for the pair of internally evolved network structure populations.

It is determined whether the adaptive experience sharing probability value is less than a presently generated one of the judgment random numbers.

If yes, determining that the pair of network structure populations after internal evolution meet cross-task knowledge communication conditions; if not, determining that the pair of network structure populations after internal evolution does not meet the cross-task knowledge communication condition.

The calculation formula of the self-adaptive experience sharing probability value is as follows:

P _es ＝1-min{F(I)}/μ _m {F(I)} (6)

wherein I is any one of the pair of internally evolved network structure populations; f (I) represents an evaluation value for each individual within the any one population; min { F (I) } represents the minimum value of all individual evaluation values calculated; mu (mu) _m { F (I) } represents the median of all individual evaluation values calculated.

For the current pair of internal post-evolution network structure populations P ¹ ,P ² Firstly, judging whether a cross-task knowledge communication condition is met, if so, P ¹ ,P ² Cross-task knowledge communication is enabled.

Specifically, for the pair of internally evolved network structure populations P ¹ ,P ² Optionally, one of the adaptive empirical sharing probability values may be calculated according to equation (6). For example, in P ¹ Calculating, I in formula (6) is P ¹ . F (I) represents P ¹ An evaluation value for each individual within the range; min { F (I) } represents the minimum value of all individual evaluation values calculated; mu (mu) _m { F (I) } represents P ¹ The evaluation values of all individuals in the range are arranged according to the size and then are positioned at one evaluation value in the middle position, namely the median.

If P ¹ The obtained adaptive experience shares the probability value P _es A judgment random number smaller than the current one in the range of (0, 1) means that knowledge transfer has positive influence on the current generation, then the next group should make full use of information of other tasks, which means that cross-task communication at this time can improve the performance of two groups, and P is determined ¹ ,P ² The cross-task knowledge communication condition is satisfied, and the subsequent cross-task knowledge communication processing can be performed.

Meanwhile, the method of the embodiment of the invention further comprises the following steps:

and returning to the step of carrying out intra-population evolution by utilizing a genetic algorithm for each current network structure population when the inter-task knowledge communication condition is not satisfied for the pair of internally evolved network structure populations.

Still as exemplified above, if P ¹ The obtained adaptive experience shares the probability value P _es A judgment random number which is larger than or equal to the current generated in the range of (0, 1) indicates that the performance of two populations cannot be improved by cross-task communication at the moment, invalid knowledge migration or knowledge negative migration is generated between the two populations, and P is determined ¹ ,P ² And (3) returning to the step (S3) to carry out the intra-population evolution again respectively if the inter-task knowledge communication condition is not met.

In an alternative embodiment, for the pair of internal post-evolution network structure populations, when the cross-task knowledge communication condition is satisfied, the method shares the excellent individual information based on the cross-task knowledge communication to obtain an updated network structure population corresponding to each internal post-evolution network structure population, including:

a1, aiming at each internal evolutionary network structure population meeting cross-task knowledge communication conditions, acquiring a target individual with the highest evaluation value in the internal evolutionary network structure population.

Let P be ¹ ,P ² Meets the cross-task knowledge communication condition, aims at P ¹ All individuals in the network structure are respectively subjected to network structure performance evaluation to obtain corresponding evaluation values, and one individual with the highest evaluation value is selected, and is assumed to be q ₁ The method comprises the steps of carrying out a first treatment on the surface of the Likewise, for P ² All individuals in the network structure are respectively subjected to network structure performance evaluation to obtain corresponding evaluation values, and one individual with the highest evaluation value is selected, and is assumed to be q ₂ 。

A2, performing non-alignment crossing operation on the obtained two target individuals to obtain a pair of new target individuals.

Will q ₁ And q ₂ Performing non-alignment crossover operation to obtain s ₁ 'and s' ₂ . For specific non-aligned cross operations, please refer to the above related descriptions, and the detailed description is omitted herein.

A3, adding a pair of new target individuals into each network structure population after internal evolution to obtain respective summation populations.

Will s ₁ 'and s' ₂ Adding P ¹ Obtaining P ¹ Is a sum population of (a). And will s ₁ 'and s' ₂ Adding P ² Obtaining P ² Is a sum population of (a)

And A4, selecting a preset number of excellent individuals from each added population by using the evaluation values of the individuals in the added population to obtain an updated network structure population corresponding to each network structure population after internal evolution.

For P ¹ Respectively evaluating the network structure performance of all individuals in the population, obtaining corresponding evaluation values, sorting the evaluation values in a mode from high to low, and selecting N with the front sorting _p Individual, get P ¹ A corresponding updated network fabric population.

For P ² Respectively evaluating the network structure performance of all individuals in the population, obtaining corresponding evaluation values, sorting the evaluation values in a mode from high to low, and selecting N with the front sorting _p Individual, get P ² A corresponding updated network fabric population.

It can be appreciated that the updated network structure population is obtained by inter-population evolution and cross-task knowledge exchange of the two populations to share the excellent individual information.

Regarding the cross-task knowledge exchange mechanism, please refer to fig. 5, fig. 5 is a schematic diagram of the cross-task knowledge exchange mechanism provided by the embodiment of the present invention.

S5, judging whether the current iteration number reaches the preset maximum iteration number.

The preset maximum iteration number is set when the network structure search parameter is configured, for example, 20 times in a table.

If not, returning to the step of utilizing a genetic algorithm to carry out intra-population evolution aiming at each current network structure population.

That is, if the current iteration number does not reach the preset maximum iteration number, S3 is returned. The updated network structure population obtained at this time S4 becomes the "current network structure population" in S3 after returning to S3.

And S6, if yes, stopping iteration, acquiring an optimal network structure from each updated network structure population, and obtaining a change detection result by using a corresponding sample to be detected.

If the current iteration number reaches the preset maximum iteration number, an optimal network structure with the highest evaluation value is obtained from each updated network structure population at the moment, individual decoding is carried out, the number of neurons of an input layer and an output layer is determined according to the characteristic dimension and the classification number of the input network, and the input layer and the output layer are added to obtain a real network structure. And (3) determining the real network structure as a network model which is most suitable for the current task, and then inputting each sample to be tested of the task in the step S1 into the network model respectively to obtain a corresponding classification result, namely obtaining a change detection result of whether hyperspectral remote sensing image data at the front moment and the rear moment corresponding to the sample to be tested change.

According to the scheme provided by the embodiment of the invention, the respective CD tasks can be executed for two hyperspectral change detection data sets at the same time, on the basis that each task performs intra-population evolution, collaborative analysis among the populations is performed by means of cross-task knowledge communication to share excellent individual information, useful information among different data sets can be effectively mined, an autonomous perception optimal network structure is realized for each task, and a proper network structure is sought for each CD task to perform change detection, so that the classification precision and the classification efficiency of the change detection can be improved for each CD task.

In order to verify the effectiveness of the methods provided by the embodiments of the present invention, experimental data are described below. The method of the embodiment of the invention is compared with other three change detection algorithms, namely a change vector analysis (Change Vector Analysis, CVA), a depth network and slow feature analysis (Deep Network and Slow Feature Analysis, DSFA) based and a General End-to-End Two-Dimensional convolutional neural network (GETNET) change detection algorithm, so that the practicability of the method of the embodiment of the invention is further verified by comparing the detection result images and detection quantization indexes of the three other change detection algorithms on a Farm hyperspectral remote sensing dataset and a River hyperspectral dataset.

FIG. 6 is a Farm hyperspectral dataset used in experiments in accordance with embodiments of the present invention; fig. 6 (a) is an image taken on 3 th month 5 of 2006; fig. 6 (b) is an image taken on day 23 of month 4 of 2007; fig. 6 (c) is a change detection reference diagram of fig. 6 (a) and 6 (b). The data set has a size of 450 x 140 pixels and has 155 spectral bands after removal of the noise spectral band. The Farm hyperspectral dataset presents the change of the farmland coverage area near the salt city of Jiangsu province of China, and the change content presented in the farmland dataset is the farmland size and the regional transition from visual view.

FIG. 7 is a test dataset River hyperspectral dataset used in experiments in accordance with embodiments of the present invention; fig. 7 (a) is an image taken on 3 days of 2013, 5; fig. 7 (b) is an image taken on the 31 th 12 th 2013 month; fig. 7 (c) is a change detection reference diagram of fig. 7 (a) and 7 (b). The data set has a size of 463×241 pixels, and has 198 spectral bands after removing the noise spectral band. The River hyperspectral dataset is photographed by Jiangsu province in China, and the main change type is the change of the ground features of the River bed and the River coast.

In the embodiment of the invention, the parameter setting of the adaptive search of the network structure for the hyperspectral remote sensing image dataset is shown in the table 1.

The quantitative indicators of the performance of the experimental selection evaluation algorithm are the Overall detection Accuracy (OA), kappa coefficient, F1 score (First Error Measure, F1) and the area under the curve (Area Under ROC Curve, AUC) of the subject's working characteristics (Receiver Operating Characteristic, ROC). OA can be calculated by the following formula:

the correct number of changes (TP) is the number of correctly detected changed pixels. The False Positive (FP) is the number of pixels that are actually unchanged but are erroneously detected at the time of change. The correct invariance (TN) is the number of unchanged pixels that are correctly detected. The missing detection number (FN) is the number of pixels that actually change but are erroneously detected as unchanged.

To further evaluate the change detection result, a kappa coefficient is introduced, which can be calculated by the following formula:

and PRE can be calculated by the following formula:

kappa coefficients are one way to evaluate consistency in statistics, in the change detection problem, the meaning of consistency is whether the model prediction result and the actual classification result are the same, and the higher the value of this coefficient, the higher the classification accuracy that represents the model implementation.

The F1 score (First Error Measure, F1) is a decision criterion that integrates the accuracy and recall together, the value of which is a weighted harmonic average of the accuracy and recall. F1 can measure the performance of the model, and the higher the value is, the better the performance of the model is, the maximum value is 1, the minimum value is 0, and the calculation mode is as follows:

Wherein,

for a classification problem, a subject work feature (Receiver Operating Characteristic, ROC) curve is commonly used to evaluate the predictive effect of the model. The graph is a graph drawn by FP on the horizontal axis and TP on the vertical axis, and these two coordinates are mutually restricted. For the change detection problem, it is optimal to have a lower FP and a higher TP as a result, so the ROC curve always takes on an upward convex shape, rather than a downward concave one. After the curve is plotted, if further quantitative analysis of the model performance is required, the concept of ROC area under the curve (Area Under ROC Curve, AUC) needs to be introduced. AUC area is actually the size of the area under the ROC curve calculated by integrating along the ROC horizontal axis. The larger the AUC value, the better the performance of the model is, and the higher the application value is.

Table 2 shows the quantization index for change detection using the method of the present invention embodiment with the other three algorithms on both the Farm hyperspectral dataset and the River hyperspectral dataset.

TABLE 2

Fig. 8 (a) - (d) show the results of the change detection on the Farm dataset using the CVA, DSFA, GETNET algorithm and the method of the present invention, respectively, and fig. 8 (e) is a reference diagram of the change detection of the dataset. The data set mainly shows the change condition of farmland areas, has the characteristics of concentrated change areas, single change types and the like, but presents the phenomenon that a plurality of lines are crossed at the juncture of the farmland, and is easy to be interfered by noise. As can be seen from the examination of the changes in the norm data set, GETNET and the method of the invention have similar good visual properties. The detection results of the changes of CVA, DSFA and GETNET all have quite speckle noise, so that a large number of false detections are caused, and the CVA performance is seriously affected by the speckle noise. Wherein the boxes in fig. 8 represent regions that are falsely detected as more distinct in the class of variation.

Fig. 9 (a) - (d) show the results of the change detection on the River dataset using the CVA, DSFA, GETNET algorithm and the method of the present invention, respectively, and fig. 9 (e) shows a reference view of the change detection of the dataset. Because River data sets mainly show the change of riverbed sediments, change areas are in an elongated discontinuous shape, a plurality of scattered small areas are accompanied, and almost no large areas appear, so that the noise influence effect is relatively more prominent in the change detection process, and the change detection difficulty of some small areas is improved greatly. As can be seen from comparison of the change detection results on the River data set, the change detection results obtained by the method are less affected by noise interference than CVA and GETNET, and a more complete change area is presented than CVA, DSFA, GETNET. In fig. 9, the solid line box indicates a region where the change class is detected erroneously, and the broken line box indicates a region where the unchanged class is detected erroneously.

In order to verify the role of the collaborative analysis framework, a comparison experiment was performed on the SPNA-CA method under the collaborative analysis framework of the present invention with the SPNA-single method (SPNA-single) based on a single task on one dataset. As shown in fig. 10, fig. 10 is a comparison experiment result of an SPNA-CA method and a single-task SPNA method based on collaborative analysis of an autonomous perceived network structure according to an embodiment of the present invention; FIG. 10 (a) shows fitness evaluation contrast; FIG. 10 (b) shows a classification accuracy evaluation comparison; in fig. 10, the Generation of the horizontal axis represents the number of iterations; fitness represents Fitness; OA represents the overall detection accuracy, i.e. classification accuracy.

As can be seen from fig. 10, in the optimization process, compared with the single-task SPNA, the convergence speed of the collaborative SPNA-CA is faster, the population fitness is faster and tends to be in a stable state, and meanwhile, the population fitness value is better (the lower the fitness value is, the better). In addition, in the iterative process, the fluctuation of the multi-task SPNA-CA method is small, the speed that the OA value of the classification precision tends to be stable is high, and the final OA value is higher than that of the single-task SPNA-single method.

FIG. 11 is a graph showing the adaptive experience sharing probability P according to the embodiment of the present invention _es Is determined according to the analysis result of the (a); FIG. 11 (a) shows P _es A change in the optimization process; FIG. 11 (b) shows a different P _es A change in the lower loss function value. FIG. 11 (a) is an adaptive empirical sharing probability P _es In the initial stage of the population, the individuals in the population are randomly initialized, so that the distribution of fitness in the population is more dispersed, and the adaptive experience sharing probability P is calculated by a formula _es A large value means that the communication between the different tasks is frequent at this time. With SPNA task optimizedThe process is gradually advanced, the populations corresponding to the tasks of different SPNA are optimized to a certain extent, and some individuals with poor performance are eliminated while excellent individuals are generated, so that the average fitness value among the populations is obviously improved. At this time, P calculated according to the formula _es The values also become progressively smaller, and each SPNA task is more concerned with optimizing inside its own task. As shown in FIG. 11 (b), when P _es As the number of iterations increases and decreases, the Loss value Loss decreases until the steady state is reached. As can be seen by those skilled in the art from the above-mentioned adaptive empirical sharing probability analysis results, the communication between different tasks is important during the initial iteration, but as the number of iterations increases, the communication between different tasks becomes more and more unnecessary, which illustrates the necessity of designing the sharing probability in the method provided by the embodiment of the present invention.

In summary, it can be seen that, according to the hyperspectral change detection method based on the collaborative analysis autonomous perception network structure provided by the embodiment of the invention, change detection of two groups of hyperspectral data sets can be performed simultaneously, and compared with the prior art, the method has good visual performance, less influence degree of noise interference, better presentation of a complete change region, faster convergence speed and higher classification precision.

The foregoing is merely illustrative of the preferred embodiments of the present invention, and is not intended to limit the scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.

Claims

1. The hyperspectral change detection method based on collaborative analysis of the autonomous perception network structure is characterized by comprising the following steps of:

2. The method for detecting hyperspectral changes based on collaborative analysis autonomous perceptual network structure according to claim 1, wherein the determining training samples and samples to be detected in each set of hyperspectral change detection data sets comprises:

respectively performing pre-change detection on each group of hyperspectral change detection data sets to obtain corresponding pre-change detection results;

according to the pre-change detection result corresponding to each hyperspectral change detection data set, training samples in the hyperspectral change detection data set are selected, and samples except the training samples in the hyperspectral change detection data set are used as samples to be tested.

3. The hyperspectral variation detection method based on collaborative analysis autonomous perceptual network structure according to claim 2, wherein the selecting training samples in each set of hyperspectral variation detection data sets according to the pre-variation detection result corresponding to the hyperspectral variation detection data sets comprises:

For each pre-change detection result, regarding each pixel in the image data of the pre-change detection result, taking the pixel as a central pixel of a rectangular window with a preset size, and judging whether the pre-change detection results of all pixels in the rectangular window are consistent;

4. The hyperspectral variation detection method based on collaborative analysis autonomous perception network structure according to claim 1, wherein the detecting network structure search task for each variation, generating a corresponding initial network structure population by using a preset individual gene coding mode, includes:

detecting a network structure search task aiming at each change, and randomly generating a positive integer n within a preset network layer number range; randomly generating n positive integers in a preset intra-layer neuron number range to form a one-dimensional vector serving as an initial network structure, and forming an initial network structure population corresponding to the change detection network structure searching task by the obtained preset number of initial network structures; the preset individual gene coding mode comprises coding of hidden layer depth and neuron number of each layer in the neural network.

5. The hyperspectral variation detection method based on collaborative analysis autonomous awareness networking architecture according to claim 1 or 4, wherein the performing intra-population evolution by genetic algorithm for each current networking architecture population to obtain a corresponding intra-evolved networking architecture population comprises:

aiming at each current network structure population, carrying out network structure performance evaluation on each individual by utilizing a training sample of the current network structure population to obtain an evaluation value corresponding to the individual;

selecting partial individuals with the evaluation values ranked in the front from the current network structure population as a parent population, and obtaining a plurality of offspring individuals by utilizing the pairwise cross mutation operation;

and selecting the preset number of excellent individuals from the set obtained by adding the current network structure population into the plurality of child individuals to obtain an internal evolved network structure population corresponding to the current network structure population.

6. The hyperspectral variation detection method based on collaborative analysis autonomous awareness networking architecture according to claim 5, wherein the process of evaluating the performance of the networking architecture for an individual to obtain a corresponding evaluation value includes:

F(x _i )＝λ ₁ L _t (x _i )+λ ₂ L _p (x _i )

wherein F (x) _i ) Representing an fitness evaluation function; x is x _i Representing an individual; l (L) _t (x _i ) The representation is based on individuals x _i The constructed network trains and loses; l (L) _p (x _i ) Representing network parameter loss; lambda (lambda) ₁ Represents L _t (x _i ) Weights of (2); lambda (lambda) ₂ Represents L _p (x _i ) Weights of (2);

n represents the number of training samples; b _j A true class label representing the jth training sample; a, a _j Representing a j-th training sample in the training sample set; w represents a weight parameter set of the tested network structure; f (a) _j W) represents the network structure under test versus sample a _j Is predicted by the computer; lambda represents a regularization parameter, and the value range of lambda is (0, 1); b andrespectively representing a true value and a predicted value;

wherein,is an estimate of the currently tested network trainable parameters; n is n _i Representing the number of neurons in the ith layer; k represents the number of layers of the current network under test; p (P) _max Is the trainable parameter number of the network with the maximum neuron number and the maximum layer number; n is n _m Represents the maximum number of neurons, and K represents the maximum number of layers.

7. The method for detecting hyperspectral variation of an autonomous perceived network structure based on collaborative analysis according to claim 5, wherein the selecting the predetermined number of excellent individuals from the set of the current network structure population added to the plurality of child individuals to obtain the internal post-evolutionary network structure population corresponding to the current network structure population includes:

Generating a current elimination random number, and judging whether the current elimination random number is smaller than a preset elimination probability;

if yes, eliminating the individuals with the longest existence time from the current network structure population; if not, eliminating the individual with the lowest evaluation value from the plurality of offspring individuals;

judging whether the number of the individuals currently reserved reaches the preset number or not;

if not, returning to the step of generating the obsolete random number of the current time;

if so, forming an internal evolved network structure population corresponding to the current network structure population by the currently reserved individuals.

8. The hyperspectral variation detection method based on collaborative analysis autonomous awareness networking architecture according to claim 6, wherein the determining whether the inter-evolutionary network architecture population satisfies the cross-task knowledge communication condition includes:

calculating a self-adaptive experience sharing probability value for the pair of internally evolved network structure populations;

judging whether the self-adaptive experience sharing probability value is smaller than a judgment random number generated currently;

if yes, determining that the pair of network structure populations after internal evolution meet cross-task knowledge communication conditions; if not, determining that the pair of network structure populations after internal evolution do not meet the cross-task knowledge communication condition;

P _es ＝1-min{F(I)}/μ _m {F(I)}

9. The method for detecting hyperspectral changes based on collaborative analysis autonomous perceived network structures according to claim 8, further comprising:

and returning to the step of carrying out intra-population evolution by utilizing a genetic algorithm for each current network structure population when the inter-task knowledge communication condition is not satisfied for the pair of the internally evolved network structure populations.

10. The hyperspectral variation detection method based on collaborative analysis autonomous awareness networking architecture according to claim 8, wherein the obtaining an updated networking architecture population corresponding to each internal post-evolution networking architecture population based on cross-task knowledge communication sharing excellent individual information when the cross-task knowledge communication condition is satisfied for the pair of internal post-evolution networking architecture populations includes:

Aiming at each internal evolutionary network structure population meeting cross-task knowledge communication conditions, acquiring a target individual with the highest evaluation value in the internal evolutionary network structure population;

performing non-alignment crossing operation on the two obtained target individuals to obtain a pair of new target individuals;

adding the pair of new target individuals into each internal evolved network structure population to obtain respective added populations;

and selecting the preset number of excellent individuals from each summation population by using the evaluation values of the individuals in the summation population to obtain an updated network structure population corresponding to each network structure population after internal evolution.