Photovoltaic power station vector data set construction method based on deep learning semantic segmentation algorithm

Abstract

The invention discloses a photovoltaic power station vector data set construction method based on a deep learning semantic segmentation algorithm. Firstly, open source satellite remote sensing image data and multi-source heterogeneous data are obtained and subjected to geographical calibration processing, and the consistency of the image and the multi-source data is ensured. Secondly, segmenting a research area vector by adopting a fishing net division method, taking multi-source data as priori knowledge, and obtaining a sampling area through a spatial stratified sampling algorithm; a photovoltaic power station in a sampling area is manually marked, and a semantic segmentation training set and a verification set are constructed; and establishing a photovoltaic power station segmentation model based on the DSFA-SwinNet network and the training data. And finally, carrying out segmentation processing on the residual research area by using the model, optimizing a segmentation result through filtering and morphological processing, and converting raster data into a vector format to complete construction of a photovoltaic power station vector data set.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing image processing and geographic information technology, and specifically to a method for constructing a vector data set of a photovoltaic power station based on a deep learning semantic segmentation algorithm. The method comprehensively utilizes satellite remote sensing image data, spatial analysis and artificial intelligence technology to provide technical support for the accurate identification and efficient management of photovoltaic power stations. Background Art

[0002] As the global energy transition accelerates, photovoltaic power stations, as an important form of renewable energy, play an important role in optimizing the energy structure and achieving carbon neutrality goals. However, the traditional way of obtaining information on the distribution of photovoltaic power stations mainly relies on manual surveys or simple image interpretation, which has problems such as low efficiency and insufficient accuracy, and it is difficult to meet the needs of large-scale regional monitoring and dynamic management.

[0003] In recent years, with the rapid development of deep learning and remote sensing technology, the application of semantic segmentation algorithms in complex object recognition has gradually become a research hotspot. Through semantic segmentation technology, the spatial distribution information of photovoltaic power stations can be automatically extracted from satellite remote sensing images, significantly improving the efficiency and accuracy of information extraction. However, in large-scale regional applications, existing methods still face multiple technical challenges, including: the complexity of data preprocessing and fusion caused by the diversity of data sources; the insufficient recognition accuracy of segmentation models in small target areas or areas with complex boundaries; the lack of systematic methods for post-processing of large-scale segmentation results and conversion to vector data format, resulting in inconsistent data quality and difficulty in meeting standardization requirements.

[0004] To this end, the present invention proposes a method for constructing a photovoltaic power station vector data set based on a deep learning semantic segmentation algorithm. This method uses open source satellite remote sensing images as the main data source, combines historical photovoltaic power station data and population distribution density data, performs regional optimization sampling through a spatial stratified sampling algorithm, and uses semantic segmentation technology based on the DSFA-SwinNet network to accurately identify photovoltaic power stations. Through morphological processing, vectorization conversion, data organization and other technical steps, a high-precision photovoltaic power station vector data set is constructed. This invention helps to accurately locate photovoltaic power stations and obtain distribution information, and provides assistance for dynamic monitoring, scientific planning, accurate evaluation and intelligent management of large-scale photovoltaic power stations. Summary of the invention

[0005] The technical solution of the present invention is to propose a method for constructing a photovoltaic power station vector data set based on a deep learning semantic segmentation algorithm, efficiently process satellite remote sensing image data and accurately extract photovoltaic power stations, and realize automated, high-precision and consistent modeling of the distribution of photovoltaic power stations in large areas. At the same time, the present invention can significantly improve the efficiency of dynamic monitoring, scientific planning and intelligent management of photovoltaic power stations, and help the digital transformation of the renewable energy field.

[0006] The technical solution of the present invention is: a method for constructing a photovoltaic power station vector data set based on a deep learning semantic segmentation algorithm, firstly, obtaining open source satellite remote sensing image data, historical photovoltaic power station data, study area vector data and population distribution density data, performing geographic registration and fishing net division of vector data, combining spatial stratified sampling algorithm to generate sampling areas, and manually annotating photovoltaic power stations to form photovoltaic power station data set metadata; secondly, through pre-processing steps such as data enhancement and sample balancing, constructing training and validation sets, and using the DSFA-SwinNet network to train semantic segmentation models; then, based on the trained segmentation model, extracting photovoltaic power stations from the remaining study area remote sensing images; finally, through morphological processing, noise filtering, geographic reference information matching and vectorization conversion and other steps, the segmentation results are integrated into a high-quality vector data set, and organized and stored by administrative districts and counties. The specific steps are as follows:

[0007] (1) Collect multi-source heterogeneous data and construct metadata for PV power station datasets.

[0008] In step (1), a standard data structure is used to reduce the multi-source heterogeneous data such as the acquired open source satellite remote sensing image data, historical photovoltaic power station data, study area vector data, and population distribution density data.

[0009] Furthermore, multi-source heterogeneous data can be stored efficiently. Relational databases can be used in a structured manner, or NoSQL databases can be used in an unstructured manner to perform logical storage and management of data; centralized high-performance storage devices can be used, or distributed clusters can be used to perform physical storage and optimization of photovoltaic power station remote sensing image data.

[0010] Furthermore, to obtain the metadata of the PV power plant dataset, the specific steps include:

[0011] Data integration and georeferencing. Obtain the key for satellite image data from the Tiandi Map official website, and use the tianditu tool plug-in of QGis to import the latest online satellite images. Obtain and import historical photovoltaic power station vector data, study area vector data, and population distribution density raster data, and perform spatial projection conversion according to the unified coordinate system standard to ensure the spatial consistency of vector data and satellite remote sensing image data.

[0012] Fishnet division of the study area vector data. Using the fishnet division method, the study area vector in (a) is divided into a uniform grid of 10000*10000 meters.

[0013] Spatial stratified sampling. Using the historical PV power plant vector data and population distribution density raster data in (a) as prior knowledge, the average population density and historical PV power plant area of ​​each grid element in (b) are calculated. Using the spatial stratified sampling method, a specified number of grids are selected as sampling areas.

[0014] Manually label the photovoltaic power station area. Use various selection tools in the PS software to manually label the photovoltaic power stations in the satellite remote sensing image data of the sampling area in (c). Use Python scripts to unify the label colors, label photovoltaic power stations as white (255, 255, 255), and non-photovoltaic power stations as black (0, 0, 0), and save the labeling results in png format as sample label data. The label data is matched one by one with the satellite remote sensing image data, and the names are the same. The sample label data and satellite remote sensing image data initially constitute the metadata of the photovoltaic power station dataset.

[0015] (2) Preprocessing and enhancing the metadata of the photovoltaic power station dataset obtained in step (1) to construct a photovoltaic power station segmentation sample dataset. The specific steps include:

[0016] Channel adjustment. Remove the alpha channel of the satellite remote sensing image data, and copy the single channel of the label data to generate three-channel label data, so that both are in three-channel format.

[0017] Image data optimization: Histogram equalization and gamma correction are performed on the satellite remote sensing image data in (a) to enhance the image contrast.

[0018] Image segmentation: The label data and satellite remote sensing image data in (b) are segmented into non-overlapping tiles of 256*256 pixels.

[0019] Image cleaning. Filter the samples whose non-PV power station area accounts for less than 5% or more than 95% in the sample label data after cutting in (c) to exclude most images without PV areas, providing a balanced distribution of positive and negative samples for model training and accuracy evaluation. Since then, the original PV power station sample database has been constructed.

[0020] Data enhancement. The labeled data and satellite remote sensing image data in (d) are subjected to five data enhancements using a Python script, namely horizontal flipping, brightness reduction, brightness enhancement, Gaussian noise superposition, and random scaling. They are then concatenated with the sample database in (d) to obtain the data-enhanced photovoltaic power station sample database.

[0021] Training set and validation set division: The samples and sample labels in the photovoltaic power station sample database obtained in step (e) are divided into a training set train and a validation set val in a ratio of 8:2, one by one, to obtain a photovoltaic power station segmentation sample data set.

[0022] (3) Performing semantic segmentation of the photovoltaic power station sample database obtained in step (2) and constructing a photovoltaic power station segmentation model. The specific steps include:

[0023] (a) Constructing a photovoltaic power station segmentation model. The backbone of the DSFA-SwinNet model uses Swin-Transformer to extract photovoltaic power station features. By using its excellent window and layering ideas, it overcomes the limitations of traditional CNN segmentation methods in local receptive fields, thereby more effectively capturing long-range dependencies in satellite remote sensing images. In order to alleviate the limitations of the fixed window size of the Swin-Transformer backbone network, the present invention designs a dynamic window adjustment strategy based on the constraints of physical elements, striving to achieve excellent performance in extraction accuracy while improving the computational efficiency of the model and enhancing the adaptability and practicality of the model in photovoltaic target detection tasks. The model adopts a refined skip connection strategy, integrates the DSFA attention mechanism, and connects the output feature maps of the lower two layers of the encoder to the PAR bottleneck structure. By dynamically decoupling the multi-scale feature representation of the spatial domain and the frequency domain, it further enriches its understanding of the internal details of the features and improves the model's ability in fine-grained photovoltaic power station detection. Finally, a multi-level up-adoption head is constructed to classify the output feature maps of the bottleneck structure and the four-layer decoder into two categories, and upsamples them to the pixel-by-pixel results of the input data as the output sequence of the network.

[0024] (b) Model hyperparameter tuning. The Ray Tune automatic parameter tuning tool is used to perform hyperparameter tuning on the DSFA-SwinNet model in (a), combined with grid search and Bayesian optimization algorithms. The tuning experiment focuses on the loss function weight hyperparameter, learning rate hyperparameter, and Dropout rate. 1250 samples are randomly selected from the training set train of the photovoltaic power station segmentation sample data set for parameter search experiments. The samples and sample labels are divided into training sets and validation sets in a ratio of 3:1, with 1000 training sets and 250 validation sets. Each parameter experiment is conducted for 30 rounds. The learning rate hyperparameter, loss function weight hyperparameter, and Dropout rate are adjusted in turn. The optimal result of each hyperparameter is saved and the initial value is modified. A deep supervision strategy is adopted during the training process. The output feature maps from the bottleneck structure and the last three layers of the decoder in the network output sequence are used to calculate the auxiliary loss to supervise the model to learn more feature representations. The output feature map of the top layer of the decoder is used as the main loss, so that the model can more effectively capture the characteristic information of photovoltaic power stations of different scales.

[0025] (c) Model training and saving. Use the optimized hyperparameter combination in (b) to train the model and save the segmentation model containing the optimal weights.

[0026] (3) Based on the photovoltaic power station segmentation model and the optimal hyperparameter results obtained in (2), the remote sensing image dataset of the study area is segmented to construct a photovoltaic power station prediction result dataset of the study area. The specific steps include:

[0027] (a) Satellite remote sensing image data preprocessing: Channel processing, histogram equalization, gamma correction, filling and cropping are performed on the remote sensing images of the remaining study area.

[0028] (b) Photovoltaic power station segmentation. The tile data in (a) is put into the model test set, and the photovoltaic power station segmentation model is used to segment the test set into photovoltaic power station features, and the photovoltaic power station prediction results are stored in tif format.

[0029] (4) Optimize and vectorize all the regional segmentation prediction results obtained in step (3) to construct a vector dataset of photovoltaic power stations in the study area. The specific steps include:

[0030] (a) Raster stitching and clipping. Tiles are stitched together into a complete raster image according to the file name number. The data size is restored by clipping according to the filling record.

[0031] (b) Noise filtering and boundary optimization. The Canny edge detection algorithm is used to filter the noise of the spliced ​​raster data in (a) to remove isolated points or small area noise caused by misidentification of the segmentation model. Morphological processing techniques (such as expansion and corrosion) are combined to optimize the boundary of the photovoltaic power station, improve the smoothness and accuracy of the boundary, and ensure that the spatial boundary of the photovoltaic power station is complete and accurate.

[0032] (c) Geographic reference information matching. The geographic coordinate system information and projection parameters in the satellite remote sensing image data are extracted and copied to the corresponding raster data in (b) to ensure that the PV power station segmentation results are accurately matched with the real geographic space.

[0033] (d) Vector conversion and organization. According to the administrative district and county division, the raster data processed in (c) was converted into vector data by resolution using the professional geographic information processing software ArcMap. The generated PV power station vector data was integrated with geographic information, organized and stored in combination with the administrative district and county division of the study area, forming a PV power station vector data set that meets the standardization requirements.

[0034] The advantages of the present invention compared with the prior art are:

[0035] 1. A photovoltaic power station extraction method based on deep learning semantic segmentation algorithm is provided. Compared with the existing technology, the DSFA-SwinNet model adopts the Swin-Transformer backbone network with integrated dynamic window adjustment strategy, and proposes the DSFA attention mechanism and PAR bottleneck structure to extract features from the local and global image levels, respectively, so that DSFA-SwinNet can better integrate contextual information and show higher accuracy and efficiency in feature extraction of photovoltaic power stations. Compared with traditional methods, the model segmentation takes less time and can effectively reduce the consumption of computing resources. At the same time, the model ensures more accurate segmentation results through optimized deep supervision strategy, which can meet the needs of large-scale regional photovoltaic power station automation detection and provide strong support for intelligent management.

[0036] 2. This method builds a high-quality PV power station segmentation sample database by combining preprocessing methods such as spatial stratified sampling and data enhancement. The DSFA-SwinNet model is used to efficiently extract PV power stations from satellite remote sensing images with high accuracy and significantly improved computational efficiency. Compared with traditional image processing methods, the present invention not only greatly reduces the need for manual intervention, but also improves the automation level of large-scale PV power station detection. This method can meet the needs of fast and accurate detection and is suitable for monitoring and analysis of large-scale PV power station areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is the overall flow chart of the present invention. DETAILED DESCRIPTION

[0038] In order to enable persons skilled in the art to better understand the solutions of the embodiments of the present invention, the embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings and implementation modes.

[0039] like Figure 1 As shown, the present invention comprises the following steps:

[0040] Data acquisition and sampling: To store multi-source heterogeneous data, centralized high-performance computing facilities can be used to physically store high-resolution remote sensing data using high-performance servers, high-speed networks, and high-speed storage device clusters. Distributed clusters can also be used to physically store high-resolution remote sensing data of photovoltaic power stations using distributed computing environments and storage structures. The steps for constructing the sampling area include: geographic alignment, fishing net division, spatial stratified sampling, and data labeling;

[0041] Geo-reference: Use QGis software to geo-calibrate the acquired multi-source heterogeneous data. Obtain the key from the Tiandi Map official website and import online satellite remote sensing image data through the tianditu tool plug-in. Import historical photovoltaic power station vector data, study area vector data, and population distribution density raster data. Use the study area vector to clip the online satellite remote sensing image data. Use the layer reprojection and raster reprojection tools to ensure that all data are superimposed and displayed in the unified coordinate system WGS 84.

[0042] Fishnet division: The fishnet division method is used to divide the vector data of the study area into uniform grids. The grid resolution is 10000*10000 meters. The divided grid vector data will be used for subsequent sampling analysis.

[0043] Spatial stratified sampling: Using historical photovoltaic power station data and population distribution density data as prior knowledge, the photovoltaic power station area and average population density in each grid element are counted. According to the above indicators, the grid units are divided into several categories. The photovoltaic power station area stratification includes: low photovoltaic area, medium photovoltaic area and high photovoltaic area; the population density stratification includes: low population density, medium population density and high population density. Combining the two indicators of photovoltaic power station area and population density, a total of 9 stratified categories are obtained. According to the number of grids in each category, the sampling quantity is allocated proportionally to ensure that at least 1 grid is extracted from each category. The specific rules are as follows: Assuming that it is planned to extract M grids, the number of grids in category i is N i , the total number of grids is N, the number of samples for each category is S i The calculation formula is:

[0044]

[0045] Data labeling: Use PS software to open the satellite remote sensing image data of the sampling area, and use a variety of selection tools to manually label the photovoltaic power stations. Use Python scripts to unify the label colors, label the photovoltaic power stations as white (255, 255, 255), and label non-photovoltaic power stations as black (0, 0, 0), and save the labeling results in png format. Ensure that the label file name corresponds to the satellite remote sensing image data file name and is named the same. Archive the labeled label data and the corresponding remote sensing image data as the metadata of the photovoltaic power station dataset.

[0046] The steps of data preprocessing and enhancement include: channel adjustment, image data optimization, image cutting, image cleaning, data enhancement, and division of training set and validation set;

[0047] Channel adjustment: Perform channel format processing on the label data and satellite remote sensing image data of the PV power plant dataset metadata. Satellite remote sensing image metadata contains four channels: R, G, B, and Alpha. Remove the Alpha channel and keep only the RGB channels. The label metadata is a grayscale image. Copy the single-channel label data and generate a three-channel format to make it consistent with the satellite remote sensing image data format.

[0048] Image data optimization: Perform histogram equalization and gamma correction on satellite remote sensing images. Enhance the contrast and brightness of the image, improve the clarity of the edge features of the photovoltaic power station, make it conform to the visual perception characteristics, and thus enhance the characteristic performance of the target area.

[0049] Image cutting: Use an automated program to cut the processed label data and satellite remote sensing image data into non-overlapping tiles of 256*256 pixels.

[0050] Image cleaning: Screen the cut tiles and filter out samples where the proportion of photovoltaic power stations in the tile label data is less than 5% or greater than 95%, so as to exclude most irrelevant areas or single category areas.

[0051] Data enhancement: Five data enhancement operations are performed on the cleaned data and the corresponding label data, namely horizontal flipping, brightness reduction, brightness enhancement, superimposition of Gaussian noise and random scaling. The enhanced data are merged with the original data to construct a data-enhanced photovoltaic power station sample database.

[0052] Training set and validation set division: The data-enhanced photovoltaic power station sample database is divided into a training set train and a validation set val in a ratio of 8:2.

[0053] Perform semantic segmentation of photovoltaic power stations on the photovoltaic power station segmentation sample data set and build a photovoltaic power station segmentation model. The specific steps include: building a photovoltaic power station feature learning model, model hyperparameter tuning, model training, and storing optimal model weights;

[0054] Constructing a feature learning model for photovoltaic power stations: Swin-Transformer is used as the encoder backbone network, and its hierarchical window attention mechanism is used to capture long-range dependencies in high-resolution remote sensing images, while alleviating the problem of insufficient feature extraction caused by local receptive field limitations of traditional convolutional networks. Combined with the constraints of physical elements, a dynamic window size adjustment mechanism is designed to dynamically optimize the spatial size of the window to more accurately extract local and global features of photovoltaic power stations, thereby improving extraction efficiency and accuracy. The DSFA attention mechanism is incorporated into the jump connection strategy, which is divided into a channel-space path and a channel-frequency path. In the channel-space path, the spatial features are decoupled from the H and W dimensions of the input features to obtain the spatial attention weights. In the channel-frequency path, based on the DCT method, according to multiple frequency indexes freq x and freq y Construct DCT filters in different frequency spaces, map the input features from the spatial domain to the frequency domain, and obtain the frequency attention weight. The mapping formula is as follows:

[0055]

[0056] Split the input tensor by channel dimension, for each channel slice X t (i=0,1,2,…,15), F i [u i ,v i ] represents the two-dimensional frequency space coefficient, T i [x, y] represents the tensor value of the (x, y) coordinate in the spatial domain. H and W represent the resolution of the input tensor, and u i and v i is the frequency coordinate, which is defined as follows:

[0057]

[0058] λ(u i ) and λ(v i ) is the regularization coefficient, which is defined as follows:

[0059]

[0060] The spatial attention weight and the frequency attention weight are respectively multiplied with the input features, and the obtained feature tensor is sigmoid processed to obtain dynamic weights that adapt to the input tensor features. The dynamic weight is disassembled, and the disassembled items are multiplied with the spatial attention tensor and the frequency attention tensor to obtain the result feature map of the attention mechanism. The PAR bottleneck structure is introduced between the encoder and the decoder. This structure uses 1*1 traditional convolution and depth-separable convolution with different void rates {2, 4, 8} to construct a four-layer pyramid structure. The fusion features of the lower two layers of the encoder are input, and after multi-scale processing of the pyramid structure, they are input into the DSFA attention mechanism to decouple the spatial and frequency features, and refine the information expression of the photovoltaic power station in the high-level features.

[0061] Model hyperparameter tuning: 1250 samples were randomly selected from the training set train of the photovoltaic power station segmentation sample data set for parameter search experiments. The samples and sample labels were divided into training set and validation set in a ratio of 3:1, with 1000 training sets and 250 validation sets. The learning rate, loss function weight and Dropout rate were adjusted in turn using the model convergence speed and IoU as indicators. Grid search was used to tune the Dropout rate hyperparameter, and the search space was defined as [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8]. Bayesian optimization was used to tune the learning rate and loss function weight hyperparameters. 20 attempts were made in the [1e-4, 1e-1] and [0, 1] search spaces to obtain the optimal solution during the trial process. Each parameter experiment was conducted for 30 rounds. ASHA Scheduler was used during the experiment to terminate the test with poor results as soon as possible.

[0062] Model training: The training set in the photovoltaic power station segmentation sample data set is used as input, and the validation set is used for model evaluation. According to the optimal hyperparameter results searched in the experiment, the training set is trained for 100 rounds using cosine annealing and label fuzzification methods. The loss function calculation process uses a deep supervision mechanism to perform weighted summation on the model output sequence and apply it to the loss function calculation. Through layer-by-layer optimization, the model's generalization and extraction capabilities for multi-scale photovoltaic features are improved. The evaluation indicators of the model training results use ACC, precision, recall, and F1.

[0063] Store optimal model weights: Compare the best performance on the validation set and use centralized high-performance computing facilities or distributed computing environments and storage structures to store the optimal model weight files.

[0064] Based on the photovoltaic power station segmentation model, the satellite remote sensing image data of the remaining research area is segmented to construct the segmentation prediction results of the entire area. The specific steps include: test set preparation, feature segmentation of all photovoltaic power stations, and storage of segmentation prediction results;

[0065] Test set preparation: Check whether the size of the label data and remote sensing image data is divisible by 256. If not, fill the image with the upper left corner as the origin: the fill color is (0, 0, 0), the width expands from the right, and the height expands from the bottom to ensure data integrity. The filling operation will generate a txt document containing the image filling situation, which is used to record the parameters of the image filling operation so that the image size can be restored in subsequent processing. The filled data is cropped into non-overlapping tiles of 256*256 size. For data cut tiles from the same image, they are stored in the same folder. The folder is named after the image data for easy subsequent management and call.

[0066] Segmentation of all PV power station features: Put all folders containing tiles into the test set, use the optimal PV power station segmentation model to perform PV power station feature segmentation, and obtain the segmentation prediction results of all regions.

[0067] Storing segmentation prediction results: Using centralized high-performance computing facilities or utilizing distributed computing environments and storage structures, the prediction results of PV power station segmentation are physically stored according to administrative districts and counties.

[0068] The obtained regional segmentation prediction results are optimized and vectorized to construct a vector dataset of photovoltaic power stations in the study area. The specific steps include: raster stitching and clipping, noise filtering and boundary optimization, geographic reference information matching, vector conversion and organization.

[0069] Raster stitching and clipping: According to the file name number of the tile data, the tile data in the segmentation prediction result is stitched into a complete raster image. Read the filling record txt file, and clip and restore the stitched raster image to restore the original data size.

[0070] Noise filtering and boundary optimization: Using the Canny edge detection algorithm, the minimum bounding rectangle of each result object is constructed with the upper left corner as the origin, and isolated points, small area noise and other pseudo information generated by the segmentation model misidentification in the spliced ​​image are detected and removed. Based on the neat strip texture of the photovoltaic array, combined with the expansion and corrosion operations of morphological processing, the boundary of the photovoltaic power station is optimized to improve the smoothness and accuracy of the boundary, reduce the occlusion of background information such as vegetation or buildings, and ensure that the segmentation result can accurately reflect the actual outline and boundary characteristics of the photovoltaic power station.

[0071] Geographic reference information matching: Use the GDAL library and Python script to batch extract geographic coordinate information, projection parameters and resolution information from satellite remote sensing image metadata. Copy and match the extracted spatial reference information to the optimized raster segmentation results to ensure that the photovoltaic power station segmentation results can be accurately connected with the real geographic space, and also lay the foundation for converting raster data into vector format.

[0072] Vector conversion and organization: Use arcpy library and python script to batch vectorize the segmentation prediction results to generate high-precision PV power station vector data. Combined with the division of administrative districts and counties in the study area, organize and classify the vector data, and intersect the PV power station vector data with the regional boundary information. Store the integrated vector data in a standardized data format Shapefile, and classify and store them according to the administrative districts and counties to form a PV power station vector data set that meets the needs of regional research.

[0073] The embodiments of the present invention are described in detail above. The present invention is explained in this article using specific implementation methods. The description of the above embodiments is only used to help understand the method of the present invention. At the same time, for those skilled in the art, according to the ideas of the present invention, there will be changes in the specific implementation methods and application scopes. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A method for constructing a photovoltaic power station vector dataset based on a deep learning semantic segmentation algorithm, characterized by the following: Steps: Obtain open-source satellite remote sensing image data, study area vector data, historical photovoltaic power station data and population distribution density data, and perform georeferencing on the data; divide the study area vector data through fishing nets, use historical photovoltaic power station data and population distribution density data as prior knowledge, adopt spatial stratified sampling algorithm to obtain sampling areas, manually label photovoltaic power stations in the sampling areas, and export and store photovoltaic power station dataset metadata containing satellite remote sensing image data and labels; preprocess and enhance the metadata of photovoltaic power station datasets, divide them into training sets and validation sets, and construct photovoltaic power station segmentation sample datasets; Based on the DSFA-SwinNet network, the PV power station segmentation sample data set is semantically segmented and a PV power station segmentation model is constructed. Based on the PV power station segmentation model, the satellite remote sensing image data of the remaining study area is segmented and the segmentation prediction results of all areas are stored. The segmentation prediction results of the entire area are filtered and morphologically processed, the raster data are converted into vector format, and a vector dataset of photovoltaic power stations in the study area is constructed.

2. The method for constructing a photovoltaic power station vector data set based on a deep learning semantic segmentation algorithm according to claim 1, characterized in that: In the step (1), the vector data of the study area is divided into fishing nets to limit the analysis scope, and the historical photovoltaic power station data and population distribution density data are combined as prior knowledge, and the spatial stratified sampling algorithm is used to optimize the sampling scheme to ensure that the sampling area is representative; for the storage of the metadata of the photovoltaic power station data set, a relational database or a NoSQL database is used for logical storage and management of data; a centralized high-performance storage device is used, or a distributed cluster method is used for physical storage and optimization of data.

3. The method for constructing a photovoltaic power station vector data set based on a deep learning semantic segmentation algorithm according to claim 1, characterized in that: In the step (2), the satellite remote sensing image data and labels are channel adjusted, and the image contrast is enhanced by using histogram equalization and gamma correction; the two are cut into non-overlapping tiles according to the specified tile size, and sample balance filtering is performed based on the positive and negative sample ratios; and the data are divided into a training set and a validation set in a ratio of 8:

2.

4. The method for constructing a photovoltaic power station vector data set based on a deep learning semantic segmentation algorithm according to claim 1, characterized in that: In the step (3), semantic segmentation of the photovoltaic power station segmentation sample data set includes: Randomly extract a small training set from the training set and use the automated parameter tuning tool Ray Tune to tune the model hyperparameters; Based on the hyperparameter combination of the optimal experimental indicators, the DSFA-SwinNet network is used to train the characteristics of the photovoltaic power station; and the segmentation model containing the optimal weight parameters is stored.

5. The method for constructing a photovoltaic power station vector data set based on a deep learning semantic segmentation algorithm according to claim 1, characterized in that: In the step (4), segmenting the satellite remote sensing image data of the remaining study area includes: performing channel adjustment, histogram equalization and gamma correction, as well as image filling and cutting on the remote sensing images of the remaining study area; extracting photovoltaic power station features based on a segmentation model with optimal weight parameters; and storing the model segmentation prediction results according to the administrative district and county scope.

6. The method for constructing a photovoltaic power station vector data set based on a deep learning semantic segmentation algorithm according to claim 1, characterized in that: In the step (5), the construction of the vector data set includes: according to the administrative district and county scope, the tiles of the segmentation prediction results are spliced ​​into a complete raster image, and the filling area and the boundary are cropped and optimized; the spliced ​​raster data is subjected to noise filtering using the Canny edge detection algorithm, and the integrity and spatial accuracy of the photovoltaic power station boundary are further optimized in combination with morphological processing; the processed raster data is matched with the geographic reference information of the satellite remote sensing image data to ensure the spatial consistency of the data; the optimized raster data is converted into a vector format by resolution using ArcMap or other professional geographic information processing software; the photovoltaic power station vector data is organized and stored according to the administrative district and county scope to form a complete photovoltaic power station vector data set in the study area, and provide standardized data support for subsequent applications;.

Images