Meteorological cloud chart cloud shape recognition method based on large-scale evolutionary multi-task algorithm

Abstract

The application discloses a weather cloud map cloud shape recognition method based on a large-scale evolution multi-task algorithm, and comprises the following steps: 1, collecting ground-based weather cloud map original observation data, and performing feature extraction on the data; 2, generating a plurality of low-dimensional weather cloud map feature selection tasks; 3, initializing a weather cloud map feature selection scheme population corresponding to each task; 4, generating a child weather cloud map feature selection scheme of each task by knowledge transfer on an individual of each task; 5, merging the child and parent weather cloud map feature selection scheme populations to perform environment selection, and iteratively performing the above process to finally obtain an optimal weather cloud map feature selection scheme, and using the features in the scheme to perform weather cloud map cloud shape recognition. The application can efficiently process the cloud shape recognition classification problem based on large-scale weather cloud map data, and greatly improves the accuracy of cloud shape recognition.

Description

Technical Field

[0001] This invention relates to the field of feature selection technology for meteorological cloud image cloud shape recognition in large-scale multi-objective problems, specifically a meteorological cloud image cloud shape recognition method based on a large-scale evolutionary multi-task algorithm. Background Technology

[0002] Clouds play a crucial role in indicating weather and climate changes, and accurate cloud information is of great significance to various aspects of political, economic, and social life. Currently, cloud observation is mainly conducted through two methods: satellite cloud observation and ground-based cloud observation. Among them, ground-based meteorological cloud imagery observation has attracted much attention due to its timeliness and accuracy in local cloud observation. In particular, a key parameter in ground-based meteorological cloud imagery observation is cloud type, and the correct identification of cloud type plays a vital role in understanding numerical weather prediction, analyzing climate conditions, and atmospheric circulation models.

[0003] In the task of cloud shape recognition in ground-based meteorological cloud images, useful features are typically extracted from the cloud images first, and then a classifier is used to identify cloud shapes based on these features. Among these methods, the Local Binary Pattern (LBP) feature extraction method has received widespread attention and application. However, the number of features extracted by the LBP method from cloud images is enormous, far exceeding the number of samples, thus facing a severe challenge in high-dimensional feature selection. If all features are considered for modeling or classification, not only is the computational cost high, but overfitting is also prone to occur, resulting in a significant decrease in prediction accuracy. At the same time, for some meteorological cloud image observations, the acquisition cost of certain feature observation data is very high. If these features themselves have little or no impact on the categorical variable, and we include these features indiscriminately in the model, it will further increase the cost of observation data collection and model application. Therefore, how to remove redundant and irrelevant features from the extracted large-scale meteorological cloud image feature data and further select high-quality meteorological cloud image features has become the focus and difficulty of meteorological cloud image cloud shape recognition and classification technology.

[0004] However, most current meteorological cloud image feature selection methods suffer from the "curse of dimensionality" problem due to the high number of features on large-scale meteorological cloud image feature sets. These methods are computationally intensive, time-consuming, and memory-intensive, ultimately failing to find a subset of meteorological cloud image features with good cloud shape recognition capabilities. Many existing ranking and filtering meteorological cloud image feature selection methods, while having the advantage of low computation time, only consider the correlation between a single feature and the label, or only the correlation between paired features. Therefore, the cloud shape recognition ability of the resulting meteorological cloud image feature subset is not ideal. On the other hand, many wrapper-style meteorological cloud image feature selection methods utilize the "divide and conquer" approach, using clustering algorithms to isolate the entire high-dimensional meteorological cloud image feature set into several low-dimensional meteorological cloud image feature groups, hoping to solve the "curse of dimensionality" problem in meteorological cloud image feature selection. However, this method ignores the complex relationships between different groups of meteorological cloud image features, so the cloud shape recognition ability of the resulting meteorological cloud image feature subset is often weak, and may even produce significant bias or errors in the identification of meteorological cloud image cloud shapes. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, this invention proposes a meteorological cloud image cloud shape recognition method based on a large-scale evolutionary multi-task algorithm. The aim is to efficiently process meteorological cloud image data and obtain a subset of meteorological cloud image features with the fewest possible features and the strongest possible classification ability, thereby enabling rapid and accurate identification of meteorological cloud image cloud shapes from large-scale meteorological cloud image data.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0007] The meteorological cloud image feature selection method based on a large-scale evolutionary multi-task algorithm of the present invention is characterized by the following steps:

[0008] Step 1: Obtain the raw observation data from the ground-based meteorological cloud map, and use the Local Binary Pattern Feature (LBP) method to extract features from the raw observation data to obtain an X×D-dimensional meteorological cloud map feature set and an X-dimensional cloud label vector, where X represents the number of cloud map samples and D represents the number of meteorological cloud map features corresponding to each cloud map sample.

[0009] Step 2: Generate multiple low-dimensional meteorological cloud image feature selection tasks;

[0010] Step 2.1: Calculate the k-th weight w using equation (1). k Thus, the weight set W = {w1, w2, ..., w...} is obtained. k , ..., w T};

[0011]

[0012] In equation (1), T is the weight number;

[0013] Step 2.2: Calculate the importance index m of the kth feature using equation (2). k Thus, a list of feature importance indices M = {m1, m2, ..., m} is obtained. k , ..., m T};

[0014] m k =w k *ReliefF+(1-m k )*SU (2)

[0015] In equation (2), ReliefF represents the D-dimensional ReliefF index vector; SU represents the D-dimensional SU index vector; m k It is a D-dimensional vector, where the value of each dimension represents the importance of the corresponding meteorological cloud map feature;

[0016] Step 2.3: Generate T low-dimensional meteorological cloud image feature selection tasks based on the feature importance index set M, and construct a meteorological cloud image feature selection task set G = {g1, g2, ..., g...} k , ..., g T}, where g k This represents the subset of meteorological cloud image features searched by the k-th task; and each task is restricted to a low-dimensional feature space for searching.

[0017] Step 2.3.1: Based on the importance index m of the k-th feature k The D meteorological cloud image features of all cloud image samples are sorted in descending order to obtain the sorted feature vector, where k∈[1,T];

[0018] Step 2.3.2: Select the top r features from the sorted feature vectors to form the feature subset g of the meteorological cloud image searched for the k-th task. k Let D k Represents a subset g of meteorological cloud image features k Meteorological cloud image feature numbers;

[0019] Step 3: Set the total population size to N, then the population size for the k-th task is... Initialize k = 1;

[0020] Step 3.1: Define the population set of the k-th task as denoted as . p ki Let represent the i-th individual in the population for the k-th task, where each individual represents a subset g of features from the meteorological cloud image. kChoose a meteorological cloud image feature selection scheme composed of different meteorological cloud image features, and p ki ={p ki / 1 p ki / 2 , ..., p ki / j , ..., p ki / D}, p ki / j Represents the i-th individual p ki Whether the j-th meteorological cloud image feature should be included in the meteorological cloud image feature selection scheme, if p ki / j =1, which means that the j-th meteorological cloud image feature is selected into the i-th individual. If p ki / j =0, which means that the j-th meteorological cloud map feature was not selected for the i-th individual; where k∈[1,T];

[0021] Step 3.1.1: Define and initialize the number of iterations t = 1, and the maximum number of iterations is maxT;

[0022] Step 3.1.2: Define the t-th generation population for the k-th task. in, This represents the t-th generation population for the k-th task. The i-th individual, and initialized For a length of D k A vector consisting entirely of zeros;

[0023] Step 3.1.3: Initialize i = 1;

[0024] Step 3.1.4: Generate the i-th random number between [0, 1] rand i Let the total number of selection features for the i-th individual be .

[0025] Step 3.1.5: Starting from the i-th individual Find two meteorological cloud map features randomly selected from the chosen features, and denote the indices of the two selected meteorological cloud map features as c and d respectively;

[0026] make Indicates whether the c-th meteorological cloud image feature is selected for the i-th individual. Indicates whether the d-th meteorological cloud image feature is selected for the i-th individual.

[0027] If m k / c >m k / d Then let This indicates that the c-th meteorological cloud image feature will be selected into the meteorological cloud image feature selection scheme. If so, otherwise, order This indicates that the d-th meteorological cloud image feature will be selected into the feature selection scheme. Where, m k / cThis represents the importance of the c-th meteorological cloud map feature among the k-th feature importance indicators; m k / d This indicates the importance of the d-th meteorological cloud map feature among the k-th feature importance indicators;

[0028] Step 3.1.6: Repeat step 3.1.5 to execute num. i This process is repeated to obtain the updated i-th individual and assign it to...

[0029] Step 3.1.7: After assigning i+1 to i, determine if i > N. k Does this condition hold true? If it does, it indicates that the t-th generation of the population has completed the k-th task. Update the population and use it as the initial population; otherwise, return to step 3.1.4 to execute.

[0030] Step 3.2: Evaluate the initial population for the k-th task respectively. Each individual in the dataset is sorted using a non-dominated ordering method.

[0031] Step 3.2.1: Use equation (3) to calculate the i-th individual in the t-th generation of the population for the k-th task. Classification error rate metric

[0032]

[0033] In equation (3), This indicates the use of the i-th individual. The number of samples that were misclassified when selecting the meteorological cloud image features;

[0034] Step 3.2.2: Count the i-th individual. The number of meteorological cloud image features selected in the meteorological cloud image feature selection scheme represented by it

[0035] Step 3.2.3: Using the error rate of each individual in the t-th generation of the k-th task and the selected meteorological cloud image feature number, perform a non-dominated sort on all individuals to obtain the frontier number of the individual, denoted as . in, Represents the i-th individual The front surface number;

[0036] Step 3.2.3: After assigning i+1 to i, determine if i > N. k If the condition is met, it indicates that the evaluation of the t-th generation population for the k-th task has been completed; otherwise, return to step 3.2.1 for execution.

[0037] Step 3.2.4: After assigning k+1 to k, check if k > T is true. If true, it means that the initialization and evaluation of the population for T tasks have been completed; otherwise, return to step 3.1.

[0038] Step 4: Generate the meteorological cloud image feature selection scheme for the t-th generation of the k-th task. Let i represent the meteorological cloud image feature selection scheme in the t-th generation of the k-th task, initialized k=1;

[0039] Step 4.1: Initialize i = 1;

[0040] Step 4.2: Define the task number for knowledge transfer as R, and generate the i-th random number Rand between [0, 1]. i If Rand i If k > 0.5, then assign the value of k to R; if Rand i If the value is less than 0.25, then assign k-1 to R; otherwise, assign k+1 to R.

[0041] Step 4.3: Randomly select two individuals from the t-th generation population of the R-th task, and denote them as follows: and Compare the frontal surface numbers of the two individuals and like Then Assign it to Ind; otherwise, Assign the value to Ind;

[0042] Step 4.4: Find the i-th individual. Features that have been selected but not selected by individual Ind, or features selected by the i-th individual. Features that were not selected but were selected for individual Ind are combined into a set Diff;

[0043] Step 4.5: Individual Assign the value to the i-th individual in the t-th generation of the k-th task. Randomly select from set Diff One feature, and will The values ​​of each feature in the corresponding dimension are assigned to... The corresponding dimension in the set; where rand represents a random number between [0, 1]; |Diff| represents the size of the set Diff;

[0044] Step 4.6, for individuals Perform uniform mutation to obtain the mutated individuals and assign them to...

[0045] Step 4.7: After assigning i+1 to i, determine if i > N. k If the condition is met, it indicates that the feature selection scheme for the t-th generation of the sub-generation of the k-th task has been obtained. Otherwise, return to step 4.2 and execute;

[0046] Step 5: Merge the meteorological cloud image feature selection schemes of the offspring and parent generations of each task and perform environmental selection to select the meteorological cloud image feature selection schemes to enter the next generation.

[0047] Step 5.1: Transfer the t-th generation population of the k-th task. and the tth generation offspring population Merge the populations and perform a non-dominated sort on the merged populations; obtain the t-th generation merged population for the k-th task after sorting.

[0048] Step 5.2: Determine the front edge number L of the merged population of generation t for the k-th task. k and make the previous L k -1 The total number of individuals in a frontal plane is less than N k And the former L k The total number of individuals in each frontal plane is greater than or equal to N. k ; thereby making the first L k Individuals from the front surface are placed into the set Cand k ;

[0049] Step 5.3, if set Cand k The number of meteorological cloud image feature selection schemes exceeds N. k Then calculate Cand k Crowding distance in meteorological cloud image feature selection scheme, remove Cand k The selection scheme for meteorological cloud image features with the smallest crowding distance up to Cand k The number of meteorological cloud image feature schemes is equal to N. k So far, the updated meteorological cloud image feature selection scheme is obtained and used as the (t+1)th generation population for the k-th task.

[0050] Step 5.4: After assigning k+1 to k, determine whether k>T is true. If true, it means that the (t+1)th generation population of the T tasks has been obtained, and step 5.5 is executed. Otherwise, return to step 4.1 and execute sequentially.

[0051] Step 5.5: After assigning t+1 to t, determine whether t > maxT holds true. If true, obtain the maxT generation population of T tasks. From the union of the N feature selection schemes of the maxT generation population of T tasks, select the meteorological cloud image feature selection scheme P with the lowest classification error rate index error. minP was selected from the feature set of meteorological cloud images. min The selected meteorological cloud image features are used for cloud shape identification; otherwise, return to step four and execute sequentially.

[0052] The present invention provides an electronic device, including a memory and a processor, characterized in that the memory is used to store a program that supports the processor in executing the meteorological cloud image cloud pattern recognition method, and the processor is configured to execute the program stored in the memory.

[0053] The present invention provides a computer-readable storage medium on which a computer program is stored, characterized in that the computer program, when run by a processor, executes the steps of the meteorological cloud image cloud pattern recognition method.

[0054] Compared with existing technologies, the beneficial effects of this invention are reflected in:

[0055] 1. This invention proposes a set of efficient meteorological cloud image feature importance evaluation indicators that are independent of the specific characteristics of meteorological cloud image data. By combining distance-based ReliefF meteorological cloud image feature indicators and information entropy-based SU meteorological cloud image feature indicators with different weights, several sets of meteorological cloud image feature importance indicators with certain correlations are obtained. This diverse set of indicators can compensate for the shortcomings of indicators that only consider the correlation between meteorological cloud image features and cloud shape labels, thereby enabling a more accurate assessment of the importance of each meteorological cloud image feature. This, in turn, provides more effective guidance for the subsequent evolutionary process of using evolutionary algorithms to solve for high-quality subsets of meteorological cloud image features.

[0056] 2. Based on the meteorological cloud image feature importance evaluation index set proposed in the first point, this invention designs a fast and effective meteorological cloud image feature selection task generation strategy. By associating an index with each task and sorting the features in descending order based on the index, the top r features are collected to form the search dimension of the corresponding task. Thus, the feature selection optimization problem in the high-dimensional meteorological cloud image raw data is transformed into the optimization of multiple low-dimensional auxiliary tasks generated. Through knowledge transfer between tasks, the meteorological cloud image feature subset with the strongest cloud shape recognition capability is efficiently found.

[0057] 3. This invention proposes an efficient and sparse initialization strategy that can quickly remove irrelevant or redundant meteorological cloud image features from the meteorological cloud image feature selection scheme, thereby greatly reducing the time for subsequent meteorological cloud image feature selection and cloud shape recognition, and reducing the computational resources consumed. Attached Figure Description

[0058] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0059] In this embodiment, a meteorological cloud image cloud shape recognition method based on a large-scale evolutionary multi-task algorithm is described, such as... Figure 1 As shown, a high-quality meteorological cloud image feature selection scheme is obtained through the designed feature importance calculation strategy, task generation strategy, and sparse initialization strategy. This method can significantly reduce the time and space consumption in generating the meteorological cloud image feature selection scheme by optimizing multiple interconnected low-dimensional subtasks and transferring knowledge between subtasks, thus overcoming the "curse of dimensionality" problem in large-scale meteorological cloud image feature selection. Specifically, it proceeds according to the following steps:

[0060] Step 1: Obtain the raw observation data from the ground-based meteorological cloud map, and use the LBP method to extract features from the raw observation data to obtain an X×D dimensional meteorological cloud map feature set and an X-dimensional cloud label vector, where X represents the number of cloud map samples and D represents the number of meteorological cloud map features corresponding to each cloud map sample. As shown in Table 1, the meteorological cloud map dataset contains 10 meteorological cloud map features and 3 samples. The first 10 columns represent the relative numerical values ​​of the corresponding meteorological cloud map features in different samples, and the 11th column represents the label category of the cloud shape recognition corresponding to the sample, where 0 represents cirrus clouds, 1 represents cumulus clouds, and 2 represents stratiform clouds.

[0061] Table 1 Feature Set of Meteorological Cloud Images

[0062] Feature Index 1 2 3 4 5 6 7 8 9 10 11 Sample 1 0.32 0.61 0.84 0.42 0.98 0.56 0.48 0.12 0.43 0.23 1 Sample 2 0.71 0.89 0.31 0.29 0.72 0.67 0.94 0.31 0.83 0.56 0 Sample 3 0.52 0.67 0.12 0.59 0.12 0.78 0.22 0.92 0.71 0.76 2

[0063] Step 2: Generate multiple low-dimensional meteorological cloud image feature selection tasks;

[0064] Step 2.1: Calculate the k-th weight w using equation (1). k Thus, the weight set W = {w1, w2, ..., w...} is obtained. k , ..., w T};

[0065]

[0066] In equation (1), T is the weight number; in this embodiment, T is set to 3, so the three weights are calculated as {0.25, 0.5, 0.75} respectively;

[0067] Step 2.2: Calculate the importance index m of the kth feature using equation (2). k Thus, a list of feature importance indicators M = {m} is obtained. i m2, ..., m k , ..., m T};

[0068] m k =w k*ReliefF+(1-m k )*SU (2)

[0069] In equation (2), ReliefF represents the D-dimensional ReliefF index vector; SU represents the D-dimensional SU index vector; m k The vector is D-dimensional, and the value of each dimension represents the importance of the corresponding meteorological cloud map feature; Table 2 shows the ReliefF index and SU index of each feature and label, as well as the list of three feature importance indices in this embodiment.

[0070] Table 2 List of Feature Importance Indicators

[0071] Feature Index 1 2 3 4 5 6 7 8 9 10 ReliefF 0.2 0.7 0.6 0.8 0.4 0.2 0.1 0.8 0.4 0.2 SU 0.8 0.2 0.8 0.2 0.2 0.9 0.6 0.1 0.9 0.4 <![CDATA[m1]]> 0.65 0.325 0.75 0.35 0.25 0.725 0.475 0.275 0.775 0.35 <![CDATA[m2]]> 0.5 0.45 0.7 0.5 0.3 0.55 0.35 0.45 0.65 0.3 <![CDATA[m3]]> 0.35 0.575 0.65 0.65 0.35 0.375 0.225 0.625 0.525 0.25

[0072] Step 2.3: Generate T low-dimensional meteorological cloud image feature selection tasks based on the feature importance index set M, and construct a meteorological cloud image feature selection task set G = {g1, g2, ..., g...} k ,…,g T}, where g k This represents the subset of meteorological cloud image features searched by the k-th task; and each task is restricted to a low-dimensional feature space for searching.

[0073] Step 2.3.1: Based on the importance index m of the k-th feature k The D meteorological cloud image features of all cloud image samples are sorted in descending order to obtain the sorted feature vector, where k∈[1,T];

[0074] Step 2.3.2: Select the top r features from the sorted feature vectors to form the feature subset g of the meteorological cloud image searched for the k-th task. k Let D k Represents a subset g of meteorological cloud image features k Meteorological cloud image feature numbers;

[0075] In this embodiment, the result of sorting the features based on m1 is {f9,f3,f6,f1,f7,f4,f...} 10 The result of sorting the features based on m2 is {f3, f9, f6, f1, f4, f8, f2, f7, f6, f7, f8, f8, f9, f1, f2, f8, f9, f1, f1, f2, f9 ...1, f2, f1, f2, f3, f9, f1, f2, f3, f4, f8, f2, 10 The result of sorting the features based on m3 is {f4, f3, f8, f2, f9, f6, f1, f5, f5}; 10If r is set to 5, then g1 = {f9, f3, f6, f1, f7}, meaning task 1 only searches for features 9, 3, 6, 1, and 7; g2 = {f3, f9, f6, f1, f4}, meaning task 2 only searches for features 3, 9, 6, 1, and 4; g3 = {f4, f3, f8, f2, f9}, meaning task 3 only searches for features 4, 3, 8, 2, and 9.

[0076] Step 3: Set the total population size to N, then the population size for the k-th task is... Initialize k = 1;

[0077] Step 3.1: Define the population set of the k-th task as denoted as . p ki Let represent the i-th individual in the population for the k-th task, where each individual represents a subset g of features from the meteorological cloud image. k Choose a meteorological cloud image feature selection scheme composed of different meteorological cloud image features, and p ki ={p ki / 1 ,p ki / 2 ,…,p ki / j ,…,p ki / D}, p ki / j Represents the i-th individual p ki Whether the j-th meteorological cloud image feature should be included in the meteorological cloud image feature selection scheme, if p ki / j =1, which means that the j-th meteorological cloud image feature is selected into the i-th individual. If p ki / j =0, which means that the j-th meteorological cloud map feature was not selected for the i-th individual; where k∈[1,T];

[0078] Step 3.1.1: Define and initialize the number of iterations t = 1, and the maximum number of iterations is maxT;

[0079] Step 3.1.2: Define the t-th generation population for the k-th task. in, This represents the t-th generation population for the k-th task. The i-th individual, and initialized For a length of D k A vector consisting entirely of zeros;

[0080] Step 3.1.3: Initialize i = 1;

[0081] Step 3.1.4: Generate the i-th random number between [0, 1] rand i Let the total number of selection features for the i-th individual be .

[0082] Step 3.1.5: Starting from the i-th individual Two meteorological cloud map features are randomly selected from the unselected features, and the indices of the two selected meteorological cloud map features are denoted as c and d respectively;

[0083] make Indicates whether the c-th meteorological cloud image feature is selected for the i-th individual. Indicates whether the d-th meteorological cloud image feature is selected for the i-th individual.

[0084] If m k / c >m k / d Then let This indicates that the c-th meteorological cloud image feature will be selected into the meteorological cloud image feature selection scheme. If so, otherwise, order This indicates that the d-th meteorological cloud image feature will be selected into the feature selection scheme. Where mk / c represents the importance of the c-th meteorological cloud map feature among the k-th feature importance indicators; m k / d This indicates the importance of the d-th meteorological cloud map feature among the k-th feature importance indicators;

[0085] Step 3.1.6: Repeat step 3.1.5 to execute num. i This process is repeated to obtain the updated i-th individual and assign it to...

[0086] Step 3.1.7: After assigning i+1 to i, determine if i > N. k Does this condition hold true? If it does, it indicates that the t-th generation of the population has completed the k-th task. Update the population and use it as the initial population; otherwise, return to step 3.1.4 to execute.

[0087] As shown in Table 3, in this embodiment, the feature subset searched by the k-th task is set to g. k ={f1, f2, f5, f7, f 10 This means searching only the 1st, 2nd, 5th, 7th, and 10th features. The population size is set to 5, and the value of each dimension in an individual indicates whether g is included. k The corresponding meteorological cloud map features are selected into the meteorological cloud map feature selection scheme. For example, if individual 1 is represented by {1, 0, 0, 1, 0}, the meteorological cloud map feature selection scheme is {1, 7}, that is, the 1st and 7th meteorological cloud map features are selected.

[0088] Table 3. Population of meteorological cloud map feature selection schemes

[0089] Feature Index 1 2 3 4 5 Individual 1 1 0 0 1 0 Individual 2 1 1 0 1 0 Individual 3 0 1 1 1 0 Individual 4 0 0 1 0 0 Individual 5 1 0 1 0 0

[0090] Step 3.2: Evaluate the initial population for the k-th task respectively. And each individual, and perform non-dominated sorting;

[0091] Step 3.2.1: Use equation (3) to calculate the i-th individual in the t-th generation of the population for the k-th task. Classification error rate metric

[0092]

[0093] In equation (3), This indicates the use of the i-th individual. The number of samples that were misclassified when selecting the meteorological cloud image features;

[0094] Step 3.2.2: Count the i-th individual. The number of meteorological cloud image features selected in the meteorological cloud image feature selection scheme represented by it

[0095] Step 3.2.3: Using the error rate of each individual in the t-th generation of the k-th task and the selected meteorological cloud image feature number, perform a non-dominated sort on all individuals to obtain the frontier number of the individual, denoted as . in, Represents the i-th individual The front surface number;

[0096] Step 3.2.3: After assigning i+1 to i, determine if i > N. k If the condition is met, it indicates that the evaluation of the t-th generation population for the k-th task has been completed; otherwise, return to step 3.2.1 for execution.

[0097] Step 3.2.4: After assigning k+1 to k, determine whether k > T is true. If true, it means that the initialization and evaluation of the population of T tasks have been completed; otherwise, return to step 3.1 to execute.

[0098] Step 4: Generate the meteorological cloud image feature selection scheme for the t-th generation of the k-th task. Let i represent the meteorological cloud image feature selection scheme in the t-th generation of the k-th task, initialized k=1;

[0099] Step 4.1: Initialize i = 1;

[0100] Step 4.2: Define the task number for knowledge transfer as R, and generate the i-th random number Rand between [0, 1]. i If Rand i If k > 0.5, then assign the value of k to R; if Rand iIf the value is less than 0.25, then assign k-1 to R; otherwise, assign k+1 to R.

[0101] Step 4.3: Randomly select two individuals from the t-th generation population of the R-th task, and denote them as follows: and Compare the frontal surface numbers of the two individuals and like Then Assign it to Ind; otherwise, Assign the value to Ind;

[0102] Step 4.4: Find the i-th individual. Features that have been selected but not selected by individual Ind, or features selected by the i-th individual. Features that were not selected but were selected for individual Ind are combined into a set Diff;

[0103] Step 4.5: Individual Assign the value to the i-th individual in the t-th generation of the k-th task. Randomly select from set Diff One feature, and will The values ​​of each feature in the corresponding dimension are assigned to... The corresponding dimension in the set; where rand represents a random number between [0, 1]; |Diff| represents the size of the set Diff;

[0104] As shown in Table 4, in this embodiment, individuals The bits that differ from the individual Ind value are the 2nd, 3rd, and 4th bits, which are shown in gray in Table 4, i.e., Diff = {f2, f3, f4}. The generated random number rand is set to 0.6. Two features are randomly selected from Diff, let's say the 3rd and 4th features. The values ​​of the 3rd and 4th bits in Ind are then assigned to q. ki The 3rd and 4th positions (highlighted underlined in Table 4) generate offspring individuals q. ki As shown in the last row of Table 4.

[0105] Table 4 Knowledge Transfer Process

[0106]

[0107] Step 4.6, for individuals Perform uniform mutation to obtain the mutated individuals and assign them to...

[0108] Step 4.7: After assigning i+1 to i, determine if i > N. kIf the condition is met, it indicates that the feature selection scheme for the t-th generation of the sub-generation of the k-th task has been obtained. Otherwise, return to step 4.2 and execute.

[0109] Step 5: Merge the meteorological cloud image feature selection schemes of the offspring and parent generations of each task and perform environmental selection to select the meteorological cloud image feature selection schemes to enter the next generation.

[0110] Step 5.1: Transfer the t-th generation population of the k-th task. and the tth generation offspring population Merge the populations and perform a non-dominated sort on the merged populations; obtain the t-th generation merged population for the k-th task after sorting.

[0111] Step 5.2: Determine the front edge number L of the merged population of generation t for the k-th task. k and make the previous L k -1 The total number of individuals in a frontal plane is less than N k And the former L k The total number of individuals in each frontal plane is greater than or equal to N. k ; thereby making the first L k Individuals from the front surface are placed into the set Cand k ;

[0112] Step 5.3, if set Cand k The number of meteorological cloud image feature selection schemes exceeds N. k Then calculate Cand k Crowding distance in meteorological cloud image feature selection scheme, remove Cand k The selection scheme for meteorological cloud image features with the smallest crowding distance up to Cand k The number of meteorological cloud image feature schemes is equal to N. k So far, the updated meteorological cloud image feature selection scheme is obtained and used as the (t+1)th generation population for the k-th task.

[0113] Step 5.4: After assigning k+1 to k, determine whether k>T is true. If true, it means that the (t+1)th generation population of the T tasks has been obtained, and step 5.5 is executed. Otherwise, return to step 4.1 and execute sequentially.

[0114] Step 5.5: After assigning t+1 to t, determine whether t > maxT holds true. If true, obtain the maxT generation population of T tasks. From the union of the N feature selection schemes of the maxT generation population of T tasks, select the meteorological cloud image feature selection scheme P with the lowest classification error rate index error. min P was selected from the feature set of meteorological cloud images. minThe selected meteorological cloud image features are used for cloud shape identification; otherwise, return to step four and execute sequentially.

[0115] In this embodiment, an electronic device includes a memory and a processor. The memory is used to store a program that supports the processor in executing the above-described meteorological cloud image cloud pattern recognition method. The processor is configured to execute the program stored in the memory.

[0116] In this embodiment, a computer-readable storage medium stores a computer program, which, when run by a processor, executes the steps of the above-described meteorological cloud image cloud pattern recognition method.

Claims

1. A weather cloud map cloud shape recognition method based on a large-scale evolutionary multi-task algorithm, characterized by The following steps are performed: Step one, obtaining original observation data in a ground-based weather cloud image, and using a local binary pattern feature method to extract features of the original observation data to obtain a weather cloud image feature set of XxD dimensions and a cloud shape label vector of X dimensions, wherein X represents the number of cloud image samples, and D represents the number of weather cloud image features corresponding to each cloud image sample; Step two, generating a plurality of low-dimensional weather cloud image feature selection tasks; Step 2.1, calculate the kth weight w using formula (1) k Thus, the weight set W = {w1, w2,..., w k ,..., w T} is obtained; In formula (1), T is the number of weights; Step 2.2, calculate the kth feature importance indicator m using formula (2) k Thus, the feature importance indicator list M = {m1, m2,..., m k ,..., m T} is obtained; m k = w k * ReliefF + (1 - m k ) * SU (2) In the formula (2), ReliefF represents a ReliefF index vector of D dimensions; SU represents a SU index vector of D dimensions; m k is a D-dimensional vector, each dimension representing the importance of the corresponding meteorological cloud feature; Step 2.

3. Generating T low-dimensional meteorological cloud map feature selection tasks based on the feature importance indicator set M, and constituting a meteorological cloud map feature selection task set G = {g1, g2,..., gT}, wherein gk represents the meteorological cloud map feature subset searched by the kth task; and each task is limited to search in a low-dimensional feature space; k ,..., g T}, wherein g k represents the meteorological cloud map feature subset searched by the kth task; and each task is limited to search in a low-dimensional feature space; Step 2.3.1, based on the kth feature importance indicator m k The D meteorological cloud features of all cloud samples are sorted in descending order to obtain a sorted feature vector, where k ∈ [1, T]. Step 2.3.2, select the top r features from the sorted feature vectors to form the subset of weather map features g searched by the kth task k ; let D k denote the number of weather map features in the subset of weather map features g k ; Step three, set the total population size as N, then the population size of the kth task is Initialize k = 1; Step 3.1, define the population set of the kth task as p ki represents the ith individual in the population of the kth task, and each individual represents a weather map feature selection scheme composed of different weather map features selected from the weather map feature subset g k ki ki / 1 ki / 2 ki / j ki / D ki / j ki ki / j ki / j wherein k ∈ [1, T]​​​​​​​​​ Step 3.1.1, define and initialize the iteration number t=1, and the maximum iteration number is maxT; Step 3.1.2, define the tth generation population for the kth task where, denotes the ith individual of the tth generation population for the kth task and initializes to a D k length all-zero vector; Step 3.1.3, initialize i=1; Step 3.1.

4. Generating a random number rand between 0 and 1 for the i-th individual i Setting the number of total selection features for the i-th individual to Step 3.1.5, selecting two weather cloud features from the i-th individual Two weather cloud features are randomly selected from the unselected features, and the indices of the selected two weather cloud features are denoted as c and d, respectively. Let represents whether the cth weather map feature is selected into the ith individual represents whether the dth weather map feature is selected into the ith individual If m k / c > m k / d , then let denote that the cth weather map feature is selected into the weather map feature selection scheme , otherwise, let denote that the dth weather map feature is selected into the feature selection scheme wherein m k / c denotes the importance of the cth weather map feature in the kth feature importance indicator; and m k / d denotes the importance of the dth weather map feature in the kth feature importance indicator. Step 3.1.

6. Repeat the procedure of Step 3.1.5 for num i times, thus obtaining the updated ith individual and assigning it to Step 3.1.7, after i+1 is assigned to i, judge whether i>N k is true, if true, it means the tth generation population of the kth task is completed is updated and is used as the initial population; otherwise, return to step 3.1.4 for execution; Step 3.2, evaluate the initial population of the kth task, respectively each individual and perform non-dominated sorting; Step 3.2.1, calculating the classification error rate indicator for the i-th individual in the t-th generation of the k-th task using formula (3) ​ In formula (3), represents the weather map feature selection scheme using the i-th individual represents the number of samples classified incorrectly when the weather map feature selection scheme represented by the i-th individual Step 3.2.2, Statistics of the i-th individual The number of selected weather chart features in the selected weather chart feature selection scheme Step 3.2.

3. Non-dominated sorting is performed on all individuals using the error rate of each individual in the tth generation of the kth task and the selected number of weather map features, and the individual is numbered in the front, denoted as wherein, denotes the front number of the ith individual . Step 3.2.3, after i+1 is assigned to i, judging whether i>N k is established, if yes, it indicates the evaluation of the tth generation population of the kth task; otherwise, returning to step 3.2.1 to execute; Step 3.2.4, after k+1 is assigned to k, it is judged whether k>T is true, if true, it indicates that the initialization and evaluation of the population of T tasks are completed; otherwise, return to step 3.1 for execution; Step four, generating the weather map feature selection scheme of the tth generation offspring of the kth task representing the i th weather map feature selection scheme in the t th generation offspring of the k th task, initializing k = 1; Step 4.1, initialize i=1; Step 4.2, define the task number R for which knowledge transfer is performed, generate a random number Rand between [0, 1] for the i-th i If Rand i > 0.5, assign k to R; if Rand i < 0.25, assign k-1 to R; otherwise, assign k+1 to R; Step 4.

3. Randomly select two individuals from the tth generation of the Rth task, denoted as and Compare the front face numbers of the two individuals and If assign to Ind, otherwise, assign to Ind; Step 4.4, find the ith individual selected features but individual Ind has not selected features, or the ith individual has not selected features but individual Ind has selected features, and form the set Diff; Step 4.5, assign the individual to the i-th individual in the t-th generation offspring of the k-th task randomly selected from the set Diff characteristics, and assign the value of the i-th characteristic in Ind to the corresponding dimension in characteristics, and assign the value of the i-th characteristic in Ind to the corresponding dimension in ; wherein rand represents a random number between 0 and 1; |Diff| represents the size of the set Diff. Step 4.6, select individual Perform uniform variation, get the varied individual and assign it to Step 4.7, after i+1 is assigned to i, judge whether i>N k is true, if true, it means that the kth task tth generation of offspring weather cloud feature selection scheme is obtained Otherwise, return to step 4.2 for execution; Step five, merging the weather cloud image feature selection scheme of each task of the offspring and the parent and performing environmental selection to screen out the weather cloud image feature selection scheme entering the next generation; Step 5.1, merging the tth generation of the kth task and the tth generation of the offspring population and non-dominantly sorting the merged population; obtaining the sorted tth generation of the kth task merged population; Step 5.2, determine the number of front faces as L for the tth generation of the merged population of the kth task k , and make the total number of individuals in the first L k -1 front faces less than N k , and make the total number of individuals in the first L k front faces greater than or equal to N k ; thus put the individuals in the first L k front faces into the set Cand k ; Step 5.3, if the number of weather cloud map feature selection schemes in Cand k exceeds N k , calculate the crowding distance of the weather cloud map feature selection schemes in Cand k , and delete the weather cloud map feature selection scheme with the smallest crowding distance in Cand k until the number of weather cloud map feature selection schemes in Cand k equals N k , thereby obtaining the updated weather cloud map feature selection schemes and taking them as the (t+1)th generation population of the kth task Step 5.4, after k+1 is assigned to k, it is judged whether k>T is true, if true, it indicates that the t+1 generation population of T tasks is obtained, and step 5.5 is executed, otherwise, return to step 4.1 for sequential execution; Step 5.5: After assigning t+1 to t, determine whether t > maxT holds true. If true, obtain the maxT generation population of T tasks. From the union of the N feature selection schemes of the maxT generation population of T tasks, select the meteorological cloud image feature selection scheme P with the lowest classification error rate index error. min P was selected from the feature set of meteorological cloud images. min The selected meteorological cloud image features are used for cloud shape identification; otherwise, return to step four and execute sequentially.

2. An electronic device comprising a memory and a processor, characterized in that The memory is used to store a program supporting the processor to execute the weather cloud image cloud shape recognition method of claim 1, and the processor is configured to execute the program stored in the memory.

3. A computer-readable storage medium having stored thereon a computer program, characterized in that The computer program is executed by the processor to perform the steps of the weather cloud image cloud shape recognition method of claim 1.

Images