A deep learning-based photovoltaic power station identification method

Abstract

The application relates to the technical field of classification recognition, in particular to a photovoltaic power station recognition method based on deep learning, which comprises the following steps: acquiring remote sensing recognition images of photovoltaic power stations in a target area; analyzing the gray difference degrees of each superpixel block to obtain photovoltaic ground object characteristic values of the superpixel blocks, constructing ground object similarity weights between each superpixel block and each adjacent superpixel block of the superpixel block, further constructing photovoltaic texture consistency degrees of the superpixel blocks, and constructing ground object texture similarities between each superpixel block and each adjacent superpixel block of the superpixel block based on the photovoltaic texture consistency degrees and the photovoltaic ground object characteristic values; based on the ground object texture similarities, the superpixel blocks are merged by using a selective search algorithm, remote sensing map tile images are acquired, and photovoltaic power stations in the remote sensing map tile images are recognized by using a neural network model. The application can improve the pixel classification accuracy, and further improve the recognition precision of photovoltaic power stations in remote sensing recognition images.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of classification recognition, in particular to a photovoltaic power station recognition method based on deep learning. BACKGROUND

[0002] At present, the application of photovoltaic power generation is more and more widely concerned, and the construction of photovoltaic power station distribution layer helps to scientifically analyze the power generation area and photovoltaic power generation potential of photovoltaic power stations in the region, and further provides a scientific basis for the promotion and utilization of sustainable energy. However, due to the complexity of the photovoltaic power station environment and the difficulty of obtaining photovoltaic data, how to accurately construct the photovoltaic power station distribution layer is one of the challenges currently faced.

[0003] The method based on deep learning is widely used in the field of image recognition and classification. At present, the Transformer deep learning model is used to establish a photovoltaic power station recognition model, and the remote sensing recognition image of the photovoltaic power station in the target area is recognized. By classifying the remote sensing image map tiles in the remote sensing recognition image, the photovoltaic tiles and other ground object tiles in the photovoltaic power station in the target area are identified, so as to quickly and efficiently draw the photovoltaic power station distribution layer of the target area based on the distribution position of the photovoltaic tiles and other ground object tiles on the photovoltaic power station. However, the current method of classifying remote sensing image map tiles considers the efficiency, and the tile edge ground object fragments are easily generated in the process of remote sensing map tiling, which may cause the phenomenon that the classification accuracy of the tile edge pixels is lower than that of the tile central area pixels, that is, the tile edge effect, resulting in poor classification accuracy of the remote sensing image map tiles, and the photovoltaic tiles in the photovoltaic power station recognition model cannot be accurately identified, thereby affecting the accuracy of the photovoltaic power station distribution layer drawn. SUMMARY

[0004] In order to solve the above technical problems, the present application provides a photovoltaic power station recognition method based on deep learning to solve the existing problems.

[0005] The photovoltaic power station recognition method based on deep learning provided by the present application adopts the following technical scheme:

[0006] One embodiment of the present application provides a photovoltaic power station recognition method based on deep learning, comprising the following steps:

[0007] Obtain the remote sensing recognition image of the photovoltaic power station in the target area;

[0008] Extract each superpixel block in the remote sensing recognition image, analyze the gray difference degree of each superpixel block to obtain the photovoltaic ground object feature value of each superpixel block, and construct the ground object similarity weight between each superpixel block and each adjacent superpixel block according to the gray deviation of each superpixel block and each adjacent superpixel block.

[0009] Constructing photovoltaic texture consistency of each superpixel block in combination with the photovoltaic ground object feature value of each superpixel block and photovoltaic ground object feature values of all adjacent superpixel blocks of each superpixel block, analyzing difference conditions of each superpixel block and each adjacent superpixel block about photovoltaic texture consistency and about photovoltaic ground object feature values, and constructing ground object texture similarity between each superpixel block and each adjacent superpixel block.

[0010] Merging superpixel blocks based on the ground object texture similarity by using a selective search algorithm, obtaining a remote sensing map tile image, and identifying a photovoltaic power station in the remote sensing map tile image by using a neural network model.

[0011] Preferably, the obtaining of the photovoltaic ground object feature value of each superpixel block further comprises: constructing a ground object feature vector of each superpixel block according to variation conditions of pixel gray values in each superpixel block; and determining the photovoltaic ground object feature value of each superpixel block in combination with differences between elements in the ground object feature vector of each superpixel block.

[0012] Preferably, the ground object feature vector is: constructing a gray feature sequence of each superpixel block according to gray values of all pixel points in the superpixel block, calculating an average value of absolute values of elements in a K-order difference sequence of the gray feature sequence, and taking the average value as a Kth ground object feature parameter of the superpixel block, and arranging all ground object feature parameters of the superpixel block in ascending order of order to form a vector as the ground object feature vector of the superpixel block, and K is a preset positive number not less than 1.

[0013] Preferably, the construction of the gray feature sequence of each superpixel block comprises: arranging all pixel gray values in the superpixel block in ascending order and in equal number of continuous arrangement in the superpixel block according to the gray values to form the gray feature sequence of the superpixel block.

[0014] Preferably, the expression of the photovoltaic ground object feature value is:

[0015] In the formula, F i is a photovoltaic ground object feature value of an i-th superpixel block in a remote sensing identification image, exp() is an exponential function with a natural constant as a base number, K is a number of elements in a ground object feature vector of the i-th superpixel block, S i,j and S i,j-1 are the jth and j-1th elements in the ground object feature vector of the i-th superpixel block, respectively.

[0016] Preferably, the expression of the ground object similarity weight is:

[0017] In the formula, D i,gwherein, W is the similarity weight between the i-th superpixel and the g-th neighboring superpixel in the remote sensing recognition image, norm() is an exponential normalization function, wherein, is the average of elements in the g-th neighboring superpixel's feature vector of the i-th superpixel, S i,k wherein, is the k-th element in the feature vector of the i-th superpixel, K is the number of elements in the feature vector of the i-th superpixel.

[0018] Preferably, the superpixels closest to each superpixel are taken as the neighboring superpixels of the superpixel.

[0019] Preferably, the expression of the photovoltaic texture consistency of each superpixel is as follows:

[0020] wherein, E i , F i are the photovoltaic texture consistency and the photovoltaic feature value of the i-th superpixel in the remote sensing recognition image, respectively, F i,g is the photovoltaic feature value of the g-th neighboring superpixel of the i-th superpixel, D i,g is the similarity weight between the i-th superpixel and the g-th neighboring superpixel in the remote sensing recognition image, G is the number of neighboring superpixels, and exp() is an exponential function with a natural constant as the base number.

[0021] Preferably, the expression of the feature texture similarity between each superpixel and its neighboring superpixel is as follows:

[0022] wherein, V i,g is the feature texture similarity between the i-th superpixel and the g-th neighboring superpixel in the remote sensing recognition image, E i,g is the photovoltaic texture consistency of the g-th neighboring superpixel of the i-th superpixel, and ∈ is a constant to avoid a denominator of 0.

[0023] Preferably, the method for obtaining the remote sensing map tile image comprises: inputting all the superpixels in the remote sensing recognition image of the photovoltaic power station in the target region into a selective search algorithm, wherein the feature texture similarity is taken as the similarity parameter between any two superpixels in the selective search algorithm, and the output of the selective search algorithm is the remote sensing image after the superpixels are merged, which is denoted as the remote sensing map tile image.

[0024] The present application has at least the following beneficial effects:

[0025] The application is based on the analysis of photovoltaic ground feature on the superpixel block segmented by the superpixel segmentation algorithm, and constructs a photovoltaic ground feature value through the distribution characteristics of the ground feature vector of the superpixel block. The photovoltaic ground feature value represents the size of the photovoltaic ground feature of the corresponding superpixel block in the remote sensing recognition image, and improves the accuracy of subsequent merging of the superpixel blocks in the same remote sensing influence map tile.

[0026] Further, based on the analysis of the ground feature similarity between adjacent superpixel blocks, a ground feature similarity weight is constructed, which represents the ground feature similarity between the superpixel block and its adjacent superpixel blocks in the remote sensing recognition image. The ground feature similarity weight is used to weight and sum the photovoltaic ground feature values of the adjacent superpixel blocks to measure the photovoltaic texture consistency degree of the superpixel block, representing the texture consistency feature of the photovoltaic tile in the photovoltaic power station in the remote sensing recognition image. At the same time, the ground feature texture similarity is constructed based on the difference of the photovoltaic texture consistency degree of the superpixel block and the difference of the photovoltaic ground feature value, so as to ensure the integrity of the photovoltaic ground tile after merging and avoid the phenomenon of tile edge effect of the photovoltaic ground tile.

[0027] Finally, the SS selective search algorithm is used to merge the superpixel blocks in the remote sensing image, avoiding the problem that the classification accuracy of the tile edge pixels is lower than that of the tile central area pixels, improving the classification accuracy of the remote sensing image map tile, and realizing more accurate drawing of the photovoltaic power station distribution layer of the target area. BRIEF DESCRIPTION OF DRAWINGS

[0028] In order to more clearly illustrate the technical solutions and advantages of the embodiments of the present application or the prior art, the drawings needed in the description of the embodiments or the prior art will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0029] Figure 1 A step flow chart of a photovoltaic power station recognition method based on deep learning provided by the present application;

[0030] Figure 2 An extraction flow chart of the photovoltaic ground feature value of each superpixel block provided by the present application. DETAILED DESCRIPTION

[0031] In order to further clarify the technical means and effects taken by the present application to achieve the predetermined object of the application, the specific implementation, structure, features and effects of the photovoltaic power station identification method based on deep learning according to the present application are described in detail as follows in combination with the drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. In addition, the specific features, structures or characteristics in one or more embodiments can be combined in any suitable form.

[0032] Unless otherwise defined, such as the term "comprise", "comprise" or any other variant thereof is intended to cover non-exclusive inclusion, so that the circuit structure, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or includes elements inherent to such article or device. Without more limitations, the element defined by the phrase "comprising one" does not exclude the presence of additional identical elements in the article or device comprising the element. In addition, the term "and / or" used herein includes any and all combinations of one or more related listed items. All technical and scientific terms used herein have the same meaning as understood by those skilled in the art to which the present application belongs.

[0033] The specific scheme of the photovoltaic power station identification method based on deep learning provided by the present application is described in detail below in combination with the drawings.

[0034] One embodiment of the photovoltaic power station identification method based on deep learning provided by the present application is described in detail below in combination with the drawings. Figure 1 , including the following steps:

[0035] Step 1: Obtain remote sensing identification image of photovoltaic power station in target area.

[0036] The purpose of the present embodiment is to consider the texture similarity features of ground features in the remote sensing image in the photovoltaic feature measurement remote sensing image map tile, and to use the region merging technique to merge the super pixel blocks in the remote sensing image, to avoid the problem that the classification accuracy of the tile edge pixels is lower than that of the tile central area pixels, to improve the classification accuracy of the remote sensing image map tile, and to draw the photovoltaic power station distribution layer of the target area more accurately.

[0037] In the embodiment, the satellite remote sensing image data source is collected, and a photovoltaic power station recognition model is constructed by using a Transformer deep learning model to recognize the photovoltaic power station in the target area. In the embodiment, the Sentinel-2 satellite remote sensing image is taken as the remote sensing image data source, the Sentinel-2 satellite remote sensing image is preprocessed by radiation calibration, atmospheric correction and geometric correction, the geometric accuracy of the Sentinel-2 satellite remote sensing image is improved, and the image fusion and true color synthesis methods are used to finally obtain a true color fusion image with a spatial resolution of 1 m, which is recorded as the recognition training image. The recognition training image has high accuracy and ensures that the provided band can reflect the color characteristics of the ground objects, and can meet the accuracy requirements of photovoltaic power station recognition. The preprocessing of the satellite remote sensing image, the image fusion and the true color synthesis are all known technologies, and the specific process is not described in detail.

[0038] The recognition training image is convoluted by using an independent encoder to extract the photovoltaic power station features contained in the recognition training image. The positions with the photovoltaic power station features in the recognition training image are labeled by using a computer image labeling tool, and the labeled remote sensing image is recorded as a training label image. A deep learning model based on the Transformer target detection is used to input the recognition training image and the training label image into the deep learning model for training. The result of the training is recorded as a photovoltaic recognition model. The remote sensing image of the target area is determined by the latitude and longitude and the map level of the target area to be detected, and is input into the photovoltaic recognition model to obtain the remote sensing recognition image of the photovoltaic power station in the target area. The use of the image labeling tool, the image convolution, the training of the Transformer deep learning model and the target recognition are all known technologies in the field of deep learning, and the specific process is not described in detail.

[0039] Step two: extract each superpixel block in the remote sensing recognition image, analyze the gray difference degree of each superpixel block to obtain the photovoltaic ground feature value of each superpixel block, and construct the ground object similarity weight between each superpixel block and each adjacent superpixel block according to the gray deviation of each superpixel block and each adjacent superpixel block.

[0040] Generally, in order to consider the classification efficiency of the photovoltaic tiles, the remote sensing recognition image of the photovoltaic power station in the target area is subjected to remote sensing map tiling. Due to the complexity of the texture features and geometric features of the ground objects in the remote sensing map, tile edge ground object fragments are easily generated in the tiling process, which will affect the accuracy of subsequent classification and recognition of the photovoltaic tiles. Therefore, in order to improve the accuracy of subsequent classification and recognition of the photovoltaic tiles, it is necessary to solve the problem of tile edge ground object fragments generated in the tiling process.

[0041] The remote sensing recognition image is input into the SCLC superpixel segmentation algorithm (Simplified Color-Space K-Means Clustering). In this embodiment, the average superpixel size in the algorithm is 50, and the superpixel smoothness is 10. The SCLC superpixel segmentation algorithm outputs all superpixel blocks in the remote sensing recognition image. The SCLC superpixel segmentation algorithm is a known technology, and the specific process is not described again.

[0042] Generally, because the remote sensing recognition image of the photovoltaic power station has different features of different ground objects, different ground object features are formed, such as road tiles, building tiles, and photovoltaic tiles in the photovoltaic power station remote sensing image. After superpixel segmentation, each remote sensing map tile in the remote sensing recognition image is segmented into multiple superpixel blocks. If the superpixel blocks in a local area have higher similar features, the superpixel blocks in the local area are more likely to belong to the same remote sensing map tile.

[0043] In order to extract the connected domain of each remote sensing image map tile in the remote sensing recognition image, the gray feature sequence of the superpixel block is constructed according to the gray values of all pixel points in the superpixel block. Specifically, in this embodiment, all pixel gray values in each superpixel block in each remote sensing recognition image are arranged in order from small to large values and in a sequence in which the values are arranged in equal numbers and continuously in the superpixel block, such as (22, 22, 23, 23, 23, 23, 23, …, 25, 25, 26, 26, 26, 26), which is denoted as the gray feature sequence of each superpixel block. The gray feature sequence of each superpixel block reflects the gray change characteristics of various ground objects in the remote sensing recognition image of the photovoltaic power station.

[0044] In order to more accurately extract the connected domain of the remote sensing image map tile, so as to accurately identify the photovoltaic tile in the photovoltaic power station identification model, in the embodiment, the feature vector of each superpixel block is constructed according to the change of the gray value of the pixel points in each superpixel block. Preferably, in the embodiment, the K-order difference sequence of the gray feature sequence of each superpixel block is calculated, wherein K is a positive number not less than 1, the value of K is 10, the mean value of the absolute value of the elements in the k-order difference sequence is recorded as the kth feature parameter of each superpixel block, and the vector composed of all the feature parameters of each superpixel block in the order from small to large is recorded as the feature vector of each superpixel block. The feature vector reflects the uniformity of the gray value on the photovoltaic tile in the photovoltaic power station. If all the gray values in the gray feature sequence of the superpixel block are almost the same, the feature parameters in the feature vector are almost close to 0, and the change of the feature parameters corresponding to each order is small. The road tile and the building tile in the photovoltaic power station are easily affected by obstacles, and in the remote sensing identification image, they will show the feature of non-uniform gray, which will cause a large change between the feature parameters in the feature vector.

[0045] Based on the above analysis, the photovoltaic feature value of each superpixel block in the remote sensing identification image is calculated based on the difference between the elements in the feature vector of each superpixel block. In the embodiment, the specific expression is:

[0046] In the formula, F i is the photovoltaic feature value of the ith superpixel block in the remote sensing identification image, exp() is the exponential function with the natural constant as the base, K is the number of elements in the feature vector of the ith superpixel block, S i,j and S i,j-1 are the jth and j-1th elements in the feature vector of the ith superpixel block. The photovoltaic feature value represents the size of the photovoltaic feature of the corresponding superpixel block in the remote sensing identification image. Since the gray value on the photovoltaic tile in the photovoltaic power station is relatively uniform, the smaller the change of the feature parameters corresponding to each order in the feature vector, the more obvious the photovoltaic feature on the superpixel block in the remote sensing identification image, and the larger the photovoltaic feature value, that is, the more likely the superpixel block in the remote sensing identification image belongs to the photovoltaic tile.

[0047] Specifically, the extraction flowchart of the photovoltaic feature value of each superpixel block is shown in Figure 2 .

[0048] Meanwhile, the photovoltaic tiles of the same material in the photovoltaic power station in the target area are usually unit modular photovoltaic arrays, and in order to ensure that the photovoltaic array receives sufficient sunlight, the photovoltaic array is usually not affected by the interference ground objects, so that the texture features on different photovoltaic tiles in the photovoltaic power station are consistent, and the texture features of other ground object tiles in the remote sensing recognition image are easily affected by the interference ground object features, and the texture features are generally more complex. Therefore, the smaller the difference between the ground object features of a superpixel block and the ground object features of the superpixel blocks around the superpixel block, the greater the possibility that the superpixel block is on the photovoltaic tile.

[0049] In order to identify the superpixel blocks belonging to the photovoltaic tiles in the remote sensing recognition image, the nearest neighbor superpixel blocks of each superpixel block are combined for joint analysis, and the multiple superpixel blocks closest to each superpixel block are used as the adjacent superpixel blocks of each superpixel block. Specifically, in the present embodiment, the G superpixel blocks closest to each superpixel block in terms of Euclidean distance are recorded as the G adjacent superpixel blocks of each superpixel block, and in the present embodiment, the value of G is 8.

[0050] Further, according to the difference between each element of the ground object feature vector of a superpixel block and the average of the elements of the ground object feature vector of its adjacent superpixel blocks, a ground object similarity weight between each superpixel block and its adjacent superpixel blocks is constructed, and in the present embodiment, the specific expression is:

[0051] In the formula, D i,g is the ground object similarity weight between the ith superpixel block and the gth adjacent superpixel block in the remote sensing recognition image, norm() is an exponential normalization function, is the average of the elements in the ground object feature vector of the gth adjacent superpixel block of the ith superpixel block, S i,k is the kth element in the ground object feature vector of the ith superpixel block. The ground object similarity weight represents the ground object similarity features between the superpixel block and its adjacent superpixel blocks in the remote sensing recognition image, and the smaller the difference between the ground object features, the smaller the , that is, the greater the ground object similarity, the higher the ground object similarity weight between the superpixel block and its adjacent superpixel blocks, that is, the more significant the texture consistency features between the superpixel block and its adjacent superpixel blocks.

[0052] The above process of the present embodiment is repeated to obtain the ground object similarity weight between each superpixel block and its adjacent superpixel blocks.

[0053] Step three: combine the ground object similar weight and the photovoltaic ground object feature value of all adjacent super pixel blocks of each super pixel block to construct the photovoltaic texture consistency of each super pixel block, analyze the difference of the photovoltaic texture consistency and the difference of the photovoltaic ground object feature value between each super pixel block and its adjacent super pixel blocks, and construct the ground object texture similarity between each super pixel block and its adjacent super pixel blocks.

[0054] Further, combine the ground object similar weight and the photovoltaic ground object feature value of all adjacent super pixel blocks of each super pixel block to construct the photovoltaic texture consistency of each super pixel block, and the specific calculation formula in the embodiment is:

[0055] In the formula, E i is the photovoltaic texture consistency of the i-th super pixel block in the remote sensing recognition image, F i,g is the photovoltaic ground object feature value of the g-th adjacent super pixel block of the i-th super pixel block, and G is the number of adjacent super pixel blocks. In the embodiment, the number of adjacent super pixel blocks G of each super pixel block is 8. The photovoltaic texture consistency represents the texture consistency feature of the photovoltaic tile in the photovoltaic power station in the remote sensing recognition image. The ground object similar weight is used to weight and sum the photovoltaic ground object features of the adjacent super pixel blocks. The result represents the comprehensive level of the ground object features of the photovoltaic tile around the super pixel block. The smaller the difference between the photovoltaic ground object feature value of the super pixel block and the comprehensive level of the ground object features of the photovoltaic tile around the super pixel block, the more it can reflect the consistent feature of the texture feature of the photovoltaic tile in the photovoltaic power station. Therefore, the photovoltaic texture consistency is greater, that is, the super pixel block has a greater possibility of being on the photovoltaic tile.

[0056] Further, in the embodiment, according to the difference of the photovoltaic texture consistency between each super pixel block and its adjacent super pixel blocks and the difference degree of the photovoltaic ground object feature value between each super pixel block and its adjacent super pixel blocks, the ground object texture similarity between each super pixel block and its adjacent super pixel blocks is calculated, and the specific expression is:

[0057] In the formula, V i,g is the ground object texture similarity between the i-th super pixel block and the g-th adjacent super pixel block in the remote sensing recognition image, E i,gis the photovoltaic texture consistency of the gth adjacent superpixel block of the ith superpixel block, and is a constant to avoid a denominator of 0, and the value range is 0 to 1, and in the embodiment, the value is 0.1. The ground feature texture similarity represents the ground feature texture similarity between two adjacent superpixel blocks in the remote sensing recognition image, and the ground feature texture similarity is constructed by using the difference between the photovoltaic ground feature value and the difference between the photovoltaic texture consistency, so as to ensure the integrity of the photovoltaic ground feature tile after merging, and avoid the phenomenon of tile edge effect of the photovoltaic ground feature tile. The smaller the difference is, the greater the ground feature texture similarity between the adjacent superpixel blocks is, and at this time, the smaller area is more likely to be in the same remote sensing image tile.

[0058] Step four: based on the ground feature texture similarity, the superpixel blocks are merged by using the selective search algorithm, the remote sensing map tile image is obtained, and the photovoltaic power station in the remote sensing map tile image is identified by using the neural network model.

[0059] Further, in order to merge the superpixel blocks in the remote sensing recognition image and avoid the phenomenon that the classification accuracy of the tile edge pixels is lower than that of the tile central area pixels, and improve the classification accuracy of the remote sensing image map tile, the embodiment adopts the SS selective search algorithm (selective search) to extract the remote sensing map tile image by iteratively merging similar superpixel blocks.

[0060] All superpixel blocks in the remote sensing recognition image of the photovoltaic power station in the target area are input into the SS selective search algorithm, and the ground feature texture similarity is used as the similarity parameter between any two superpixel blocks in the SS selective search algorithm. The output of the SS selective search algorithm is the remote sensing image after the superpixel blocks are merged, which is recorded as the remote sensing map tile image. The SS selective search algorithm is a known technology, and the specific process is not described again.

[0061] Different connected domains in the remote sensing map tile image represent different remote sensing image map tiles. Further, in order to accurately identify different types of remote sensing image map tiles in the remote sensing map tile image, in the embodiment, 1000 remote sensing image road tile images, remote sensing image building tile images, and remote sensing image photovoltaic tile images are collected, and the CNN convolutional neural network is trained by using the collected remote sensing image ground feature tile images of each type. The pooling layer of the CNN convolutional neural network adopts maximum pooling, and the Sigmoid function is used as the activation function, and a trained neural network model is obtained.

[0062] The neural network model is trained and classified, and specific processes are not described again.

[0063] At this point, different types of remote sensing image ground object tiles in the remote sensing recognition image of the photovoltaic power station in the target area are obtained. In this embodiment, according to the different types of remote sensing image ground object tiles in the remote sensing recognition image of the photovoltaic power station in the target area, a target area photovoltaic power station distribution layer is drawn by using CAD software (Autodesk Computer Aided Design), and the distribution of the photovoltaic tiles can be observed more directly through the target area photovoltaic power station distribution layer, which is helpful for scientific analysis of the power generation area and photovoltaic power generation potential of the photovoltaic power station in the region.

[0064] It can be understood that the reference "one embodiment" or "some embodiments" and the like described in the specification means that the specific features, structures or characteristics described in connection with the embodiment are included in one or more embodiments of the application. Therefore, if "in one embodiment", "in some embodiments", "in other some embodiments", "in additional some embodiments" and the like appear in the specification, it does not necessarily refer to the same embodiment, but means "one or more but not all embodiments", unless otherwise specifically emphasized. The terms "include", "contain", "have" and their variants mean "include but not limited to", unless otherwise specifically emphasized.

[0065] It should be noted that the above sequence of the embodiments of the application is only for description, and does not represent the advantages and disadvantages of the embodiments. The above description of the specific embodiments of the specification is described. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multi-task processing and parallel processing are possible or may be advantageous. At the same time, the size of the serial number of each step in the embodiment does not mean the execution order, and the execution order of each process should be determined by its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments in the specification.

[0066] The above examples are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that the technical solutions recorded in the foregoing examples can still be modified, or some technical features therein can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1.A deep learning-based photovoltaic power station identification method, characterized in that, The method comprises the following steps: acquiring remote sensing identification images of photovoltaic power stations in a target area; extracting each superpixel block in the remote sensing identification images, analyzing the gray difference degree of each superpixel block to obtain a photovoltaic ground object feature value of each superpixel block, and constructing a ground object similarity weight between each superpixel block and each adjacent superpixel block according to the gray deviation of each superpixel block and each adjacent superpixel block; combining the ground object similarity weight and the photovoltaic ground object feature values of all adjacent superpixel blocks of each superpixel block to construct a photovoltaic texture consistency degree of each superpixel block, analyzing the difference in the photovoltaic texture consistency degree and the difference in the photovoltaic ground object feature value between each superpixel block and each adjacent superpixel block, and constructing a ground object texture similarity between each superpixel block and each adjacent superpixel block; based on the ground object texture similarity, using a selective search algorithm to merge the superpixel blocks, acquiring a remote sensing map tile image, and using a neural network model to identify the photovoltaic power stations in the remote sensing map tile image; an expression of the photovoltaic texture consistency degree of each superpixel block is: , wherein, respectively represent photovoltaic texture consistency of the i-th superpixel block in the remote sensing identification image, and photovoltaic ground feature values, respectively represent photovoltaic ground feature values of the g-th adjacent superpixel block of the i-th superpixel block, respectively represent ground feature similarity weights between the i-th superpixel block and the g-th adjacent superpixel block thereof in the remote sensing identification image, and G represents the number of adjacent superpixel blocks, is an exponential function with a natural constant as the base. 2.The photovoltaic power station identification method based on deep learning of claim 1, wherein, the acquisition of the photovoltaic ground object feature value of each superpixel block further comprises: constructing a ground object feature vector of each superpixel block according to the change of the gray value of each pixel in each superpixel block; and determining the photovoltaic ground object feature value of each superpixel block in combination with the difference between the elements in the ground object feature vector of each superpixel block. 3.The photovoltaic power station identification method based on deep learning of claim 2, wherein, the ground object feature vector is: constructing a gray feature sequence of each superpixel block according to the gray value of all pixels in the superpixel block, calculating the average value of the absolute value of the elements in the K-order difference sequence of the gray feature sequence, and taking the average value as the Kth ground object feature parameter of the superpixel block, arranging all the ground object feature parameters of the superpixel block in ascending order of order to form a vector, and taking the vector as the ground object feature vector of the superpixel block, K being a positive number greater than or equal to 1. 4.The photovoltaic power station identification method based on deep learning of claim 3, wherein, the construction of the gray feature sequence of each superpixel block comprises: arranging all the pixel gray values in the superpixel block in ascending order and in equal number of continuous arrangement in the superpixel block according to the gray value to form the gray feature sequence of the superpixel block. 5.The photovoltaic power station identification method based on deep learning according to claim 3, wherein, an expression of the photovoltaic ground object feature value is: wherein, is a photovoltaic ground feature value of the i-th superpixel block in the remote sensing recognition image, is an exponential function with a natural constant as the base number, is the number of elements in the ground feature vector of the i-th superpixel block, and are the j-th and j-1-th elements in the ground feature vector of the i-th superpixel block, respectively. 6.The photovoltaic power station identification method based on deep learning according to claim 1, wherein, an expression of the ground object similarity weight is: wherein, is a ground object similarity weight between the i-th superpixel patch and its g-th neighboring superpixel patch in the remote sensing recognition image, is an exponential normalization function, is the mean of elements within the ground object feature vector of the g-th neighboring superpixel patch of the i-th superpixel patch, is the k-th element within the ground object feature vector of the i-th superpixel patch, is the number of elements within the ground object feature vector of the i-th superpixel patch. 7.The photovoltaic power station identification method based on deep learning of claim 1, wherein, the plurality of superpixel blocks closest to each superpixel block are taken as the adjacent superpixel blocks of each superpixel block. 8.The photovoltaic power station identification method based on deep learning of claim 1, wherein, an expression of the ground object texture similarity between each superpixel block and each adjacent superpixel block is: wherein, is the texture similarity of the ground object between the i-th superpixel block and its g-th neighboring superpixel block in the remote sensing recognition image, is the photovoltaic texture consistency of the g-th neighboring superpixel block of the i-th superpixel block, is a constant to avoid the denominator being 0. 9.The photovoltaic power station identification method based on deep learning of claim 1, wherein, the acquisition of the remote sensing map tile image comprises: inputting all the superpixel blocks in the remote sensing identification images of the photovoltaic power stations in the target area into the selective search algorithm, wherein the ground object texture similarity is taken as the similarity parameter between any two superpixel blocks in the selective search algorithm, and the output of the selective search algorithm is the remote sensing image after the superpixel blocks are merged, which is recorded as the remote sensing map tile image.

Images