Urban land remote sensing image classification method based on evolutionary multi-task optimization

Abstract

The invention discloses a city land remote sensing image classification method based on evolutionary multi-task optimization, and the method comprises the steps: 1, obtaining a city land coverage remote sensing image sample, and carrying out the preprocessing, thereby constructing a sample data set; 2, decomposing the sample data set into a plurality of training subsets, and constructing a corresponding sub-population for each training subset; 3, the probability of executing knowledge migration operation is adjusted on line for the sub-population evaluation state; 4, selecting different source populations according to sub-population states to provide knowledge for the sub-populations; 5, performing interaction by using a model representing global information to obtain a better classifier; 6, selecting a part of classifiers with poor performance to carry out unbalance adjustment; and 7, finally selecting the most excellent individuals from the sub-populations to form a multi-classifier, and identifying different types of land use and coverage conditions of the urban land. According to the method, different urban land types can be quickly classified by constructing and training the optimal multi-classifier, so that the accuracy and precision of a land type classification result are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of urban remote sensing image classification, and in particular to an urban land remote sensing image classification method based on evolutionary multi-task optimization. Background Art

[0002] With the continuous advancement of urbanization, the monitoring and analysis of urban land cover types has become particularly important. Traditional urban land classification methods rely on manual feature extraction and manual adjustment of classification rules, which usually have great uncertainty and low efficiency. With the development of remote sensing technology, more and more remote sensing image classification models are used to monitor and identify the use of urban land. However, these methods are based on machine learning models for classification, which have black box characteristics and are difficult to adjust parameters, resulting in low optimization efficiency and poor interpretability. At the same time, existing urban land remote sensing image samples are often unbalanced. The classification model optimized by unbalanced data sets may ignore a small number of land cover samples, resulting in low classification accuracy and poor generalization ability. Summary of the invention

[0003] In order to overcome the shortcomings of the prior art, the present invention proposes an urban land remote sensing image classification method based on evolutionary multi-task optimization, in order to quickly and accurately classify remote sensing images of different urban land covers, thereby improving the accuracy and precision of land type classification results.

[0004] In order to achieve the above-mentioned purpose, the present invention adopts the following technical scheme:

[0005] The urban land remote sensing image classification method based on evolutionary multi-task optimization of the present invention is characterized in that it is carried out according to the following steps:

[0006] Step 1: Collection remote sensing images of different urban land types, and extract the attribute characteristics of the remote sensing images about each land type and perform standard normalization processing to obtain Remote sensing image sample sets of different urban land types ,in, Indicates Remote sensing image samples of urban land types, Indicates The attribute feature vectors of the remote sensing images of urban land types are: , Indicates Remote sensing images of urban land types attribute feature values, express The land type label, , d represents the dimension of the attribute feature vector, , is the number of land types;

[0007] Step 2: Divide the original sample set based on the one-to-many decomposition strategy ,get A binary classification remote sensing image training subset { , ,…, ,…, },in, Indicates A binary classification remote sensing image training subset, and ,in, represent Middle The attribute feature vector of the remote sensing image of each urban land type, express Land type label; Yes The land type label is A sample set of remote sensing images, Lieutenant General The samples with class labels are considered as positive samples, and the remote sensing image sample sets with other land type labels are considered as negative samples;

[0008] Step 3: Define the current number of iterations as , and initialize ;

[0009] Build sub-classifier groups, among which the The subclassifier group contains sub-classifiers, and encode each sub-classifier as an individual, thus obtaining the Daidi subclassifier population ={ , ,…, ,…, } and initialized randomly; among them, Indicates Daidi subclassifier population Middle individuals, and the dimension of each individual is , ;

[0010] Step 4: Initialization ;

[0011] Step 5: Based on In the corresponding A binary classification remote sensing image training subset Classification on In convergence state or ;

[0012] Step 6: Calculate using formula (1) The migration probability ;

[0013] (1)

[0014] In formula (1), express The migration probability is season is the initial value set; is the scaling factor;

[0015] Step 7: Randomly initialize the Generate the jth random number ,like , then the differential evolution algorithm is used to Perform crossover mutation to obtain Generation first child collection , and execute step 10; otherwise, execute step 8;

[0016] Step 8: Multi-state based source population selection strategy from Select a classifier population from the m-1 classifier populations outside As a source population;

[0017] Step 9: Based on the global population model Source population Carry out knowledge transfer and obtain The first descendant set of as well as The first descendant set of ;

[0018] Step 10: Build a bootstrap dataset and generate bootstrap vectors for individuals with poor performance , used to get The second generation set of ;

[0019] Step 11: Assign to ,like , then execute step 12, otherwise, return to step 5 and execute sequentially;

[0020] Step 12: of and Merge into The jth merged population ,calculate Each individual in the On the training subset of two-class remote sensing images The area under the curve, thus using the elite selection strategy from Select N outstanding individuals as the first Daidi subclassifier population ;

[0021] Step 13: Assign to ,judge Is it true? If so, go to step 14; otherwise, return to step 4 and execute sequentially;

[0022] Step 14: From In each generation, select The individual with the largest area under the curve constitutes the optimal multi-classifier for urban land remote sensing image sample classification ={ , ,…, ,…, }, used to classify different urban land types, where: Indicates Daidi subclassifier population The best individual.

[0023] The method for classifying urban land remote sensing image samples based on evolutionary multi-task optimization described in the present invention is characterized in that if the label of the t-th urban land remote sensing image sample , indicating that the t-th urban land remote sensing image sample is classified as artificial building cover. , indicating that the tth urban land remote sensing image sample is classified as forest cover. , indicating that the t-th urban land remote sensing image sample is classified as water cover. , indicating that the tth urban land remote sensing image sample is classified as agriculture / grassland cover.

[0024] Further, the step 5 is performed as follows:

[0025] Step 5.1: A binary classification remote sensing image training subset All attribute feature vectors in enter Processed in Output of all land type prediction labels , and with True land type labels in the training subset of two-class remote sensing images Compare and calculate In the A training subset of two remote sensing images The true positive rate and the false positive rate And form coordinate points and map them to the target space grid, and then The coordinate points composed of the true positive rate and false positive rate of each individual in are mapped to the target space grid;

[0026] Step 5.2: According to The true positive rate and false positive rate The distribution in the target space is obtained The number of grids occupied by the individuals on the Pareto frontier in the target space grid is taken as Diversity indicators ;

[0027] Step 5.3: Calculation Each individual in the binary classification remote sensing image training subset The classification accuracy on The average precision and as The accuracy index of

[0028] Step 5.4: According to and , distinguish In convergence state or :

[0029] when > and > When In evolutionary state ;in, Indicates -1st generation subclassifier population The accuracy index, Indicates -1st generation subclassifier population Diversity indicators;

[0030] when < or < When In convergence state .

[0031] Further, the step 8 is performed as follows:

[0032] Step 8.1: If In convergence state , then calculate The probability of selecting the i-th random task = , thus obtaining The random task selection vector ={ , ,…, ,…, },in,

[0033] like In evolutionary state , then calculate The probability of selecting the i-th similar task , thus obtaining Similar task selection vector ={ , ,…, ,…, },in, Indicates Daidi subclassifier population, represents the cosine similarity,

[0034] Step 8.2: According to or , using the roulette algorithm from the remaining Select a classifier population from the m-1 classifier populations as the source population.

[0035] Further, step 9 is performed as follows:

[0036] Step 9.1: Order The element in row a and column b is recorded as ,but Middle Dimensional column vector Recorded as , and obeys Gaussian distribution , thus constructing Independent multivariate Gaussian model ;in, , ;

[0037] make The element in row a and column b is denoted by ,but Middle Dimensional column vector Recorded as , and obeys Gaussian distribution , thus constructing Independent multivariate Gaussian model ;

[0038] Step 9.2: The mean encoding of , ;in, express Middle The mean of the dimensional column vector, and , and The variance of is encoded as the variance vector , ;in, express Middle The variance of the dimensional column vector, and ;

[0039] Step 9.3: According to and The KL divergence between is obtained by formula (2): and The crossover probability between , thus obtaining and The crossover probability vector between ={ };

[0040] (2)

[0041] Step 9.4: Initialize b=1;

[0042] Step 9.5: Randomly initialize the Generate the jth random number of the bth dimension ,like , then In and In After the crossover, each of them mutates, and the two b-th dimension means after the crossover mutation are obtained accordingly. and ; Meanwhile, for in and in perform cross-over and then mutate them respectively, and correspondingly obtain two cross-mutated variances of the b-th dimension and ; Otherwise, directly mutate and respectively, to obtain two mutated means of the b-th dimension and , then mutate and respectively, to obtain two mutated variances and ;

[0043] Step 9.6: If b < d, then assign b + 1 to b, and return to Step 9.5; Otherwise, it means that the cross-mutated mean vector and variance vector are obtained, and execute Step 9.7;

[0044] Step 9.7: Decode and of the cross-mutated mean vector and variance vector , to obtain the new independent multivariate Gaussian model of and the new independent multivariate Gaussian model of ;

[0045] Step 9.8: Randomly sample and respectively to obtain / 2 individuals and form the g-th generation temporary population = , where represents the n-th temporary individual of the g-th generation;

[0046] Step 9.9: If and 's Euclidean distance is less than and 's Euclidean distance , then add to 's first offspring set , otherwise, add to 's first offspring set ; Thus, add N temporary individuals to the first offspring sets of the corresponding populations respectively.

[0047] Further, the step 10 is performed as follows:

[0048] Step 10.1: Calculation Each individual in A binary classification remote sensing image training subset on The area under the curve, and select The area under the curve is minimal Individuals form the jth population to be guided in the gth generation ={ , ,…, ,…, },in, express The zth individual to be guided;

[0049] Step 10.2: Find In the A training subset of two remote sensing images The wrongly classified remote sensing image samples and the remote sensing image samples with minority class labels are combined to form the individuals to be guided. The bootstrap dataset ;

[0050] Step 10.3: Calculate using formula (3) The guiding vector on ;

[0051] (3)

[0052] In formula (3), express The distance matrix between remote sensing image samples of different land categories, express The distance matrix of remote sensing image samples of the same land category in;

[0053] Step 10.4: Using the crossover operator pair of differential evolution and Interact to obtain the zth new guided individual And join The second generation set of middle.

[0054] The electronic device of the present invention includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the urban land remote sensing image classification method, and the processor is configured to execute the program stored in the memory.

[0055] The present invention provides a computer-readable storage medium, wherein a computer program is stored on the computer-readable storage medium, and the computer program executes the steps of the urban land remote sensing image classification method when the computer program is run by a processor.

[0056] Compared with the prior art, the present invention has the following beneficial effects:

[0057] 1. The present invention is different from the traditional urban land remote sensing image classification method. It proposes an urban land remote sensing image classification method based on the combination of evolutionary multi-task optimization and decomposition multi-classification. By extracting attribute features that can effectively distinguish different types of urban land from urban land remote sensing image samples, and using evolutionary multi-task optimization to train the classification model, it is possible to quickly, efficiently and accurately classify the sample land coverage from a large number of urban land remote sensing image features. This method decomposes the multi-classification problem of urban land remote sensing images into multiple binary classification problems of land remote sensing images, so that each binary classification model only needs to distinguish the characteristic differences between one type of land cover and other land covers; at the same time, when a new land cover category appears, only a new binary classification model needs to be trained to identify the difference between the new land cover and the original land cover, and the remaining binary classification models can be fine-tuned, avoiding the complexity of redesigning the entire land remote sensing image multi-classifier; in addition, the method of the present invention can independently optimize the binary classification model for specific land cover types that are difficult to optimize, and can train special models for certain land cover categories (such as artificial building cover) according to the priority of recognition importance, thereby improving the recognition rate of specific land cover categories; if a binary classification model performs poorly, it can also be clearly mapped to a specific binary classification model, which helps to locate the source of the problem and optimize the training process, thereby saving manpower and time on a large scale, and in most cases, it can also save money, and can achieve the same or higher accuracy as manual interpretation and the combination of traditional statistical methods, thereby improving the recognition efficiency of land cover categories.

[0058] 2. In the process of training urban land remote sensing image prediction parameters, the present invention adopts an evolutionary multi-task method. Each binary land remote sensing image classification model uses Area under the curve ( ) as the optimization goal, it can well measure the overall performance and classification accuracy of unbalanced data processing, thereby overcoming the problems that the traditional land remote sensing image classification algorithm cannot effectively solve the imbalance of existing land remote sensing image sample categories and the difficulty of black-box optimization and parameter adjustment of land remote sensing image classification models. The trained classification model is superior to the traditional land remote sensing image classification model, thereby improving the classification performance and accuracy of urban land remote sensing images.

[0059] 3. In the process of training urban land remote sensing image samples, the present invention judges the state of the land remote sensing image classifier population corresponding to each binary classification remote sensing image sub-problem online, and timely adjusts when the population migrates the land cover feature knowledge and migrates the appropriate binary classifier population. At the same time, by establishing a distribution model for each population that can characterize the global urban land image feature information, the optimization is performed according to the principle of interacting with similar land cover feature dimensions and independently optimizing dissimilar land cover feature dimensions, so that each sub-population evolves to obtain the most beneficial optimization operation, thereby improving the classification effect of each binary classification sub-problem, combining to obtain multiple classifiers, and ultimately improving the classification performance of urban land remote sensing images. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 This is a diagram of the solution based on one-to-many decomposition and multi-classification of the present invention

[0061] Figure 2 The species of the present invention Independent multivariate Gaussian model 's illustration;

[0062] Figure 3 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0063] In this embodiment, a method for classifying urban land remote sensing images based on evolutionary multi-task optimization includes a step of training an urban land remote sensing image classification model and a step of detecting an urban land remote sensing image. Specifically, Figure 3 As shown, the steps are as follows:

[0064] Step 1: Urban land remote sensing image classification model training:

[0065] Step 1.1: Collection remote sensing images of different urban land types, and extract the attribute characteristics of the remote sensing images about each land type and perform standard normalization processing to obtain Remote sensing image sample sets of different urban land types ,in, Indicates Remote sensing image samples of urban land types, Indicates The attribute feature vectors of the remote sensing images of urban land types are: , Indicates Remote sensing images of urban land types attribute feature values, express The corresponding land type label, , d represents the dimension of the attribute feature vector, , is the number of land types.

[0066] Step 1.2: Divide the original sample set based on the one-to-many decomposition strategy ,get A binary classification remote sensing image training subset { , ,…, ,…, },in, Indicates A binary classification remote sensing image training subset, and ,in, represent Middle The attribute feature vector of the remote sensing image of each urban land type, express Land type label; Yes The land type label is The remote sensing image sample sets with the labels of land types are regarded as positive samples, and the remote sensing image sample sets with the labels of other land types are regarded as negative samples.

[0067] Figure 1 Taking a four-class urban land remote sensing image classification problem as an example, the basic processing flow of one-to-many decomposition is demonstrated, where the binary classifier trained by the i-th binary remote sensing image training subset is used to distinguish the i-th type of urban land remote sensing image samples from urban land remote sensing image samples of all other types.

[0068] Step 1.3: Define the current number of iterations as , and initialize ;

[0069] Build sub-classifier groups, among which the The subclassifier group contains sub-classifiers, and encode each sub-classifier as an individual, thus obtaining the Daidi subclassifier population ={ , ,…, ,…, } and initialized randomly; among them, Indicates Daidi subclassifier population Middle individuals, and the dimension of each individual is , .

[0070] Step 1.4: Initialization ;

[0071] Step 1.5: Based on In the corresponding A binary classification remote sensing image training subset The classification situation on In convergence state or ; Consider whether it is necessary to introduce external knowledge transfer or focus on self-evolution and iteration based on the state of the land remote sensing image classifier;

[0072] Step 1.5.1: A binary classification remote sensing image training subset All attribute feature vectors in enter Processed in Output of all land type prediction labels , and with True land type labels in the training subset of two-class remote sensing images Compare and calculate In the A training subset of two remote sensing images The true positive rate and the false positive rate And form coordinate points and map them to the target space grid, and then The true positive rate and false positive rate of each individual in the target space are mapped to the target space grid.

[0073] Step 1.5.2: According to The true positive rate and false positive rate ,get The number of grids occupied by the individuals on the Pareto frontier in the target space grid is taken as Diversity indicators ;

[0074] Step 1.5.3: Calculation The classification accuracy of each individual in The average precision and as The accuracy index of

[0075] Step 1.5.4: According to and , distinguish In convergence state or :

[0076] when > and > When In evolutionary state , which indicates that the optimization state of the land remote sensing image classifier is good and has made progress compared with the previous generation. Consider reducing the probability of introducing external knowledge to make it focus on self-evolution; among them, Indicates -1st generation subclassifier population The accuracy index, Indicates -1st generation subclassifier population Diversity indicators;

[0077] when < or < When In convergence state At this time, it means that the optimization state of the land remote sensing image classifier has encountered a bottleneck in a certain indicator. Compared with the previous generation, no progress has been made. Consider increasing the probability of introducing external knowledge to make it jump out of the local optimum.

[0078] Step 1.6: Calculate using formula (1) The migration probability , and according to Determine whether knowledge transfer is needed

[0079] (1)

[0080] In formula (1), express The migration probability is season is the initial value set; is the scaling factor;

[0081] Step 1.7: Randomly initialize Generate the jth random number ,like , then the differential evolution algorithm is used to Perform crossover mutation to obtain Generation first child collection , proceed to step 1.10; otherwise, proceed to step 1.8.

[0082] Step 1.8: Multi-state based source population selection strategy from Select a classifier population from the m-1 classifier populations outside As a source population;

[0083] Step 1.8.1: If In convergence state , more random and diverse knowledge sources may help Jump out of the local optimum and calculate The probability of selecting the i-th random task = , thus obtaining The random task selection vector ={ , ,…, ,…, },in,

[0084] like In evolutionary state , It has not converged yet. Choosing more similar knowledge sources will help Accelerate the optimization process to converge faster and calculate The probability of selecting the i-th similar task , thus obtaining Similar task selection vector ={ , ,…, ,…, },in, Indicates Daidi subclassifier population, represents the cosine similarity,

[0085] Step 1.8.2: According to or , using the roulette algorithm from the remaining Select a classifier population from the m-1 classifier populations as the source population.

[0086] Step 1.9: Population based on the global population model Source population Carry out knowledge transfer and obtain The first descendant set of as well as The first descendant set of ,Different from using knowledge transfer between individuals, the global population model can better ,represent all the land remote sensing image feature knowledge learned by the ,current classifier population, and can perform knowledge transfer more comprehensively and ,efficiently;

[0087] Step 1.9.1: Order The element in row a and column b is recorded as ,but Middle Dimensional column vector Recorded as , and obeys Gaussian distribution , which actually represents the distribution of classification decision variables learned from the b-th dimension features of land remote sensing image samples, and can characterize All classifiers are trained on the b-th dimension features to obtain knowledge, and then construct Independent multivariate Gaussian model ; Represents the land remote sensing image classifier population The classification knowledge learned by training on the land features of all dimensions, , ;

[0088] make The element in row a and column b is denoted by ,but Middle Dimensional column vector Recorded as , and obeys Gaussian distribution , thus constructing Independent multivariate Gaussian model ;

[0089] Figure 2 The independent multivariate Gaussian model and the Gaussian model diagram of each dimensional decision variable are shown. Each dimension is independent of each other, so each dimensional decision variable (column vector) of the population obeys a Gaussian model. All Gaussian models together constitute a d-dimensional independent multivariate Gaussian model, which can characterize the parameter distribution range, peak or mean dispersion degree and other information of each dimensional land remote sensing image classifier.

[0090] Step 1.9.2: The mean encoding of , ;in, Represents population Middle The mean of the dimensional column vector, and , and The variance of is encoded as the variance vector , ;in, Represents population Middle The variance of the dimensional column vector, and .

[0091] Step 1.9.3: According to and The KL divergence between and The correlation between each dimension of decision variables. The more similar each dimension of decision variables is, the easier it is to benefit from similar remote sensing image features. Similarly, if they are not similar, the crossover probability of this dimension is reduced to prevent the phenomenon of negative migration caused by irrelevant land remote sensing image features. Through formula (2), we can get and The crossover probability between , thus obtaining and The crossover probability vector between ={ };

[0092] (2)

[0093] Step 1.9.4: Initialize b=1;

[0094] Step 1.9.5: Randomly initialize the Generate the jth random number of the bth dimension ,like , then In and In After the crossover, each of them mutates, and the two b-th dimension means after the crossover mutation are obtained accordingly. and At the same time, In and middle Cross and then mutate each, and get the two b-dimensional variances after cross mutation. and Otherwise, directly and Mutate them separately to get the two b-th dimension means after mutation and , then and Mutate separately and get two variances after mutation and ;

[0095] Step 1.9.6: If b < d, then assign b + 1 to b and return to Step 9.5; otherwise, it means that the mean vector after crossover mutation is obtained and the variance vector , and execute Step 1.9.7.

[0096] Step 1.9.7: Decode and the mean vector after crossover mutation of and the variance vector , to obtain the new independent multivariate Gaussian model of , and the new independent multivariate Gaussian model of ;

[0097] Step 1.9.8: Randomly sample and / 2 individuals respectively from and form the g-th generation temporary population = , where represents the n-th temporary individual in the g-th generation;

[0098] Step 1.9.9: If and the Euclidean distance is less than and the Euclidean distance , then add to the first offspring set of , otherwise, add to the first offspring set of ; thus adding N temporary individuals to the first offspring sets of the corresponding populations respectively. When knowledge transfer is performed between the source population and in this step, may also benefit from this transfer and obtain effective information. Similarly, for the sake of optimizing efficiency, the g-th generation temporary population can be directly added to the first offspring set of .

[0099] Step 1.10: Considering that the urban land categories themselves have unbalanced attributes, for example, the forest cover samples in the city are less, while the housing cover and artificial green land cover samples are more. To prevent the unbalanced land cover categories from interfering with the optimization of the land remote sensing image classifier, construct a guiding dataset , generate guidance vectors for individuals with poor performance , used to get The second generation set of ;

[0100] Step 1.10.1: Calculation Each individual in A binary classification remote sensing image training subset on The area under the curve, and select The area under the curve is minimal Individuals form the jth population to be guided in the gth generation ={ , ,…, ,…, },in, express The zth individual to be guided;

[0101] Step 1.10.2: Find In the A training subset of two remote sensing images The wrongly classified remote sensing image samples and the remote sensing image samples with minority class labels are combined to form the individuals to be guided. The bootstrap dataset ;

[0102] Step 1.10.3: Calculate using formula (3) The guiding vector on ;

[0103] (3)

[0104] In formula (3), express The distance matrix between remote sensing image samples of different land categories, express The distance matrix of remote sensing image samples of the same land category in;

[0105] Step 1.10.4: Using the crossover operator pair of differential evolution and Interact to obtain the zth new guided individual And join The second generation set of middle.

[0106] Step 1.11: Assign to ,like , then execute step 1.12, otherwise, return to step 1.5 and execute sequentially;

[0107] Step 1.12: of and Merge into The jth merged population ,calculate Each individual in the A training subset of two remote sensing images on The area under the curve, thus using the elite selection strategy from Select N outstanding individuals as the first Daidi subclassifier population ;

[0108] Step 1.13: Assign to ,judge Is it true? If so, execute step 1.14; otherwise, return to step 1.4 and execute sequentially;

[0109] Step 1.14: From In each generation, select The individual with the largest area under the curve constitutes the optimal multi-classifier for urban land remote sensing image sample classification ={ , ,…, ,…, }, used to classify different urban land types, where: Indicates Daidi subclassifier population The best individual.

[0110] Step 2: Urban land remote sensing image prediction, using the urban land remote sensing image classification model obtained in step 1 and the corresponding urban land remote sensing image attribute characteristics to be tested to classify urban land cover:

[0111] Step 2.1: Input the urban remote sensing image data to be tested as the sample to be tested and extract the attribute features of urban land remote sensing image, recorded as ,in, Indicates the number of samples to be tested Attribute characteristics of urban land remote sensing images;

[0112] Step 2.2: Attribute characteristics of the sample to be tested Perform standard normalization to obtain the feature vector of the sample to be tested ;in, Represents the normalized first Attribute characteristics of urban land remote sensing images;

[0113] Step 2.3: Input the feature vector of the sample to be tested into the urban land remote sensing image classification model obtained in step 1.14 to obtain the classification result. If the output result is equal to 1, it means that the t-th urban land remote sensing image sample is classified as artificial building cover; if the output result is equal to 2, it means that the t-th urban land remote sensing image sample is classified as forest cover; if the output result is equal to 3, it means that the t-th urban land remote sensing image sample is classified as water cover; if the output result is equal to 4, it means that the t-th urban land remote sensing image sample is classified as agriculture / grassland cover.

[0114] In this embodiment, an electronic device includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the above method, and the processor is configured to execute the program stored in the memory.

[0115] In this embodiment, a computer-readable storage medium stores a computer program on the computer-readable storage medium, and the computer program executes the steps of the above method when executed by a processor.

[0116] The specific implementation method of the present invention is illustrated and the effect of the method of the present invention is verified by using an example of a simulation data set.

[0117] 1) Prepare standard data set:

[0118] The present invention uses the Urban Land Cover dataset as a standard dataset to verify the effectiveness of the urban land remote sensing image classification model. The Urban Land Cover dataset is a multi-classification benchmark dataset widely used in imbalanced classification. The Ecoli dataset contains 307 data, which correspond to the collected remote sensing image samples of land cover in different areas of 307 cities. All samples are used to verify the performance of the present invention for urban land remote sensing image classification. For these 307 training samples, they are divided into training samples and test sample data.

[0119] 2) Evaluation indicators:

[0120] The area under the ROC curve (AUC) is used as the evaluation index of this example to evaluate the classification performance of the present invention on different breast tissue CT images. The higher the value, the better the classification performance and the stronger the generalization when facing unbalanced samples.

[0121] 3) Experiment on standard datasets:

[0122] In order to verify the effectiveness of the proposed method, the classification algorithm of the present invention is used to model and predict the Urban Land Cover dataset with the classic EMT-MC algorithm and EMaTO algorithm, and the prediction results of the three are compared. The experimental results are shown in Table 1.

[0123] Table 1 Comparative experimental results

[0124]

[0125] It can be seen from Table 1 that in the multi-class dataset Urban Land Cover, the classification algorithm of the present invention obtains better classification performance than the other two, which further reflects the effectiveness of the present invention. With the continuous advancement of urbanization, the monitoring of urban land cover types is of great significance for urban planning and disaster management. Therefore, more and more remote sensing images are used to monitor and identify the use of urban land. This method can cover a wider range and improve accuracy. Based on this, the present invention can quickly and effectively classify the sample status from a large number of urban land remote sensing image features, thereby improving the efficiency of urban land cover type classification.

Claims

1. A method for urban land remote sensing image classification based on evolutionary multi-task optimization, characterized in that: The steps are as follows: Step 1: Collection remote sensing images of different urban land types, and extract the attribute characteristics of the remote sensing images about each land type and perform standard normalization processing to obtain Remote sensing image sample sets of different urban land types ,in, Indicates Remote sensing image samples of urban land types, Indicates The attribute feature vectors of the remote sensing images of urban land types are: , Indicates Remote sensing images of urban land types attribute feature values, express The land type label, , d represents the dimension of the attribute feature vector, , is the number of land types; Step 2: Divide the original sample set based on the one-to-many decomposition strategy ,get A binary classification remote sensing image training subset { , ,…, ,…, },in, Indicates A binary classification remote sensing image training subset, and ,in, represent Middle The attribute feature vector of the remote sensing image of each urban land type, express Land type label; Yes The land type label is A sample set of remote sensing images, Lieutenant General The samples with class labels are considered as positive samples, and the remote sensing image sample sets with other land type labels are considered as negative samples; Step 3: Define the current number of iterations as , and initialize ; Build sub-classifier groups, among which the The subclassifier group contains sub-classifiers, and encode each sub-classifier as an individual, thus obtaining the Daidi subclassifier population ={ , ,…, ,…, } and initialized randomly; among them, Indicates Daidi subclassifier population Middle individuals, and the dimension of each individual is , ; Step 4: Initialization ; Step 5: Based on In the corresponding A binary classification remote sensing image training subset Classification on In convergence state or ; Step 6: Calculate using formula (1) The migration probability ; (1) In formula (1), express The migration probability is season is the initial value set; is the scaling factor; Step 7: Randomly initialize the Generate the jth random number ,like , then the differential evolution algorithm is used to Perform crossover mutation to obtain Generation first child collection , and execute step 10; otherwise, execute step 8; Step 8: Multi-state based source population selection strategy from Select a classifier population from the m-1 classifier populations outside As a source population; Step 9: Based on the global population model Source population Carry out knowledge transfer and obtain The first descendant set of as well as The first descendant set of ; Step 10: Build a bootstrap dataset and generate bootstrap vectors for individuals with poor performance , used to obtain The second generation set of ; Step 11: Assign to ,like , then execute step 12, otherwise, return to step 5 and execute sequentially; Step 12: of and Merge into The jth merged population ,calculate Each individual in the On the training subset of two-class remote sensing images The area under the curve, thus using the elite selection strategy from Select N outstanding individuals as the first Daidi subclassifier population ; Step 13: Assign to ,judge Is it true? If so, go to step 14; otherwise, return to step 4 and execute sequentially; Step 14: From In each generation, select The individual with the largest area under the curve constitutes the optimal multi-classifier for urban land remote sensing image sample classification ={ , ,…, ,…, }, used to classify different urban land types, where: Indicates Daidi subclassifier population The best individual.

2. The urban land remote sensing image sample classification method based on evolutionary multi-task optimization according to claim 1 is characterized in that: If the label of the t-th urban land remote sensing image sample , indicating that the t-th urban land remote sensing image sample is classified as artificial building cover. , indicating that the tth urban land remote sensing image sample is classified as forest cover. , indicating that the t-th urban land remote sensing image sample is classified as water cover. , indicating that the tth urban land remote sensing image sample is classified as agriculture / grassland cover.

3. The urban land remote sensing image classification method based on evolutionary multi-task optimization according to claim 1 is characterized in that: The step 5 is carried out as follows: Step 5.1: A binary classification remote sensing image training subset All attribute feature vectors in enter Processed in Output of all land type prediction labels , and with True land type labels in the training subset of two-class remote sensing images Compare and calculate In the A training subset of two remote sensing images The true positive rate and the false positive rate And form coordinate points and map them to the target space grid, and then The coordinate points composed of the true positive rate and false positive rate of each individual in are mapped to the target space grid; Step 5.2: According to The true positive rate and false positive rate The distribution in the target space is obtained The number of grids occupied by the individuals on the Pareto front in the target space grid is taken as Diversity indicators ; Step 5.3: Calculation Each individual in the binary classification remote sensing image training subset The classification accuracy on The average precision and as The accuracy index of Step 5.4: According to and , distinguish In convergence state or : when > and > When In evolutionary state ;in, Indicates -1st generation subclassifier population The accuracy index, Indicates -1st generation subclassifier population Diversity indicators; when < or < When In convergence state .

4. The urban land remote sensing image classification method based on evolutionary multi-task optimization according to claim 3 is characterized in that: The step 8 is carried out as follows: Step 8.1: If In convergence state , then calculate The probability of selecting the i-th random task = , thus obtaining The random task selection vector ={ , ,…, ,…, },in, ; like In evolutionary state , then calculate The probability of selecting the i-th similar task , thus obtaining Similar task selection vector ={ , ,…, ,…, },in, Indicates Daidi subclassifier population, represents the cosine similarity, ; Step 8.2: According to or , using the roulette algorithm from the remaining Select a classifier population from the m-1 classifier populations as the source population.

5. The urban land remote sensing image classification method based on evolutionary multi-task optimization according to claim 4 is characterized in that: The step 9 is carried out as follows: Step 9.1: Order The element in row a and column b is recorded as ,but Middle Dimensional column vector Recorded as , and obeys Gaussian distribution , thus constructing Independent multivariate Gaussian model ;in, , ; make The element in row a and column b is denoted by ,but Middle Dimensional column vector Recorded as , and obeys Gaussian distribution , thus constructing Independent multivariate Gaussian model ; Step 9.2: The mean encoding of , ;in, express Middle The mean of the dimensional column vector, and , and The variance of is encoded as the variance vector , ;in, express Middle The variance of the dimensional column vector, and ; Step 9.3: According to and The KL divergence between is obtained by formula (2): and The crossover probability between , thus obtaining and The crossover probability vector between ={ }; (2) Step 9.4: Initialize b=1; Step 9.5: Randomly initialize the Generate the jth random number of the bth dimension ,like , then In and In After the crossover, each of them mutates, and the two b-th dimension means after the crossover mutation are obtained accordingly. and At the same time, In and middle Cross and then mutate each, and get the two b-dimensional variances after cross mutation. and Otherwise, directly and Mutate them separately to get the two b-th dimension means after mutation and , then and Mutate separately and get two variances after mutation and ; Step 9.6: If b < d, then assign b + 1 to b and return to Step 9.5; otherwise, it means that the mean vector after crossover and mutation is obtained. and the variance vector , and execute Step 9.7; Step 9.7: Decoding and The mean vector after cross-mutation With the variance vector ,get New independent multivariate Gaussian model and New independent multivariate Gaussian model ; Step 9.8: From and Randomly sample out / 2 individuals and form the g-th generation temporary population = ,in, represents the nth temporary individual of the gth generation; Step 9.9: If and Euclidean distance Less than and Euclidean distance , then join in The first descendant set of Otherwise, join in The first descendant set of Thus, the N temporary individuals are added to the first offspring set of the corresponding population respectively.

6. The urban land remote sensing image classification method based on evolutionary multi-task optimization according to claim 5 is characterized in that: The step 10 is performed as follows: Step 10.1: Calculation Each individual in A binary classification remote sensing image training subset on The area under the curve, and select The area under the curve is minimal Individuals form the jth population to be guided in the gth generation ={ , ,…, ,…, },in, express The zth individual to be guided; Step 10.2: Find In the A training subset of two remote sensing images The wrongly classified remote sensing image samples and the remote sensing image samples with minority class labels are combined to form the individuals to be guided. The bootstrap dataset ; Step 10.3: Calculate using formula (3) The guiding vector on ; (3) In formula (3), express The distance matrix between remote sensing image samples of different land categories, express The distance matrix of remote sensing image samples of the same land category in; Step 10.4: Using the crossover operator pair of differential evolution and Interact to obtain the zth new guided individual And join The second generation set of middle.

7. An electronic device, comprising a memory and a processor, characterized in that: The memory is used to store a program that supports the processor to execute the urban land remote sensing image classification method described in any one of claims 1-6, and the processor is configured to execute the program stored in the memory.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the urban land remote sensing image classification method described in any one of claims 1 to 6 are executed.

Images