Photovoltaic station fine identification method and device based on classifier and deep learning

Abstract

The invention belongs to the technical field of remote sensing detection, and particularly discloses a photovoltaic station fine identification method and device based on a classifier and deep learning, and the method comprises the steps: carrying out the preprocessing of a multi-source remote sensing image of a to-be-identified photovoltaic station in a research region range; inputting the preprocessed multi-source remote sensing image into a trained machine learning classifier to carry out preliminary coarse resolution recognition, and obtaining a high-resolution remote sensing image containing a coarse recognition contour of a potential photovoltaic region; and inputting the high-resolution remote sensing image containing the coarse identification contour of the potential photovoltaic region into a trained deep learning model for fine identification, and outputting a segmentation mask with panel type information, the segmentation mask being used for distinguishing a photovoltaic panel and a panel gap. According to the method, the potential area of the photovoltaic station can be quickly positioned in the specified area range, and fine identification is carried out on the basis, so that the accuracy and efficiency of identifying the photovoltaic station in a complex and diversified area are improved.

Description

Technical Field

[0001] The present application belongs to the field of remote sensing monitoring technology, and more specifically, to a method and device for refined identification of photovoltaic stations based on classifiers and deep learning. Background Art

[0002] With the increasing attention paid to renewable energy, photovoltaic power generation, as an important form of clean energy, has seen rapid growth in scale and speed. As the main carrier of photovoltaic power generation, the operating efficiency and safety of photovoltaic stations are directly related to the performance and reliability of the entire photovoltaic power generation system. With the increasing number of photovoltaic stations, the equipment is diverse, widely distributed and the environment is complex. How to achieve refined identification and management of photovoltaic stations over a large area is an urgent problem to be solved.

[0003] CN117671513A discloses an optical remote sensing image recognition method, device, equipment and storage medium, which combines the self-constructed photovoltaic panel index and recognition model to identify photovoltaic panels. However, this method is prone to lose the underlying feature information and cannot distinguish between photovoltaic panels and panel gaps.

[0004] CN116402109A discloses a method for obtaining a photovoltaic panel recognition model, a photovoltaic panel recognition method and a device, which obtains a deep learning model at a specific spatial resolution by creating a training set and its annotated labels. However, this method requires all remote sensing images of a specified area, has low computational efficiency, and is difficult to meet the needs of rapid recognition of large-scale field stations.

[0005] CN112348030A discloses a solar photovoltaic panel identification method, device, remote sensing big data processing platform and storage medium, which outputs the pixel points containing the photovoltaic panel in combination with the probability threshold. However, the probability threshold set by this method still has great limitations when dealing with complex environments, especially when confusing buildings and greenhouses, and cannot efficiently identify photovoltaic stations over a large range. Summary of the invention

[0006] In view of the above-mentioned defects in the prior art, the present application provides a method and device for fine-grained identification of photovoltaic stations based on classifiers and deep learning, aiming to solve the problem of fine-grained identification of photovoltaic stations over a large area.

[0007] In a first aspect, the present application provides a photovoltaic station refinement identification method based on a classifier and deep learning, comprising: Preprocess the multi-source remote sensing images of the photovoltaic stations to be identified within the research area; The pre-processed multi-source remote sensing images are input into the trained machine learning classifier for preliminary coarse resolution recognition, and a high-resolution remote sensing image containing the coarse recognition outline of the potential photovoltaic area is obtained; The high-resolution remote sensing image containing the rough identification outline of the potential photovoltaic area is input into the trained deep learning model for refined identification, and a segmentation mask with panel type information is output. The segmentation mask is used to distinguish photovoltaic panels and panel gaps.

[0008] In a second aspect, the present application also provides a photovoltaic station refined identification device based on a classifier and deep learning, comprising: A preprocessing module is used to preprocess the multi-source remote sensing images of the photovoltaic stations to be identified within the research area; The preliminary coarse resolution recognition module is used to input the pre-processed multi-source remote sensing images into the trained machine learning classifier for preliminary coarse resolution recognition, and obtain high-resolution remote sensing images containing the coarse recognition outline of the potential photovoltaic area; The refined recognition module is used to input the high-resolution remote sensing image containing the rough identification outline of the potential photovoltaic area into the trained deep learning model for refined recognition, and output the segmentation mask with panel type information. The segmentation mask is used to distinguish photovoltaic panels and panel gaps.

[0009] In the third aspect, the present application also provides a remote sensing big data processing platform, including: at least one memory for storing programs; at least one processor for executing the programs stored in the memory, when the program stored in the memory is executed, the processor is used to execute the method described in the first aspect or any possible implementation of the first aspect.

[0010] In a fourth aspect, the present application also provides a computer-readable storage medium, which stores a computer program. When the computer program runs on a processor, the processor executes the method described in the first aspect or any possible implementation of the first aspect.

[0011] In a fifth aspect, the present application further provides a computer program product, which, when executed on a processor, enables the processor to execute the method described in the first aspect or any possible implementation manner of the first aspect.

[0012] The photovoltaic station refinement identification method and device based on classifier and deep learning provided by this application, combined with the multi-band spectral characteristics of remote sensing image data, performs multi-level identification and refinement processing on photovoltaic stations in a large range; by pre-processing image data, optimizing machine learning classifier input parameters, combining morphological operations and other related data sets to screen out potential photovoltaic areas at coarse resolution, and narrowing the scope of the area to be predicted; combined with SAM, an automated semantic segmentation tool, a web-based semi-automatic refinement annotation platform is established, and a deep learning model is used to refine the boundaries and types of photovoltaic panels; a fine-tuning model is established based on land use types, so that the final recognition result has the characteristics of clear boundaries and accurate morphology. This application can quickly locate the potential area of ​​photovoltaic stations within a specified area and perform refined identification based on this, minimizing manual intervention in each step, improving the accuracy and efficiency of identifying photovoltaic stations in complex and diverse areas, and laying the foundation for obtaining high-quality photovoltaic station information. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] In order to more clearly illustrate the technical solutions in the present application or related technologies, the following is a brief introduction to the drawings required for use in the embodiments or related technical descriptions. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0014] Figure 1 This is one of the flow charts of the photovoltaic station refinement identification method based on classifier and deep learning provided in the embodiment of the present application; Figure 2 This is the second flow chart of the photovoltaic station refinement identification method based on classifier and deep learning provided in the embodiment of the present application; Figure 3 It is a flowchart of the annotation process of the web version semi-automatic refined annotation platform provided in the embodiment of the present application; Figure 4 It is a schematic diagram of the architecture of the deep learning model provided in the embodiment of the present application; Figure 5 It is a schematic diagram of comparing the preliminary coarse-resolution recognition result and the refined recognition result provided in the embodiment of the present application; Figure 6 It is a structural schematic diagram of a photovoltaic station refined identification device based on a classifier and deep learning provided in an embodiment of the present application; Figure 7 It is a structural diagram of the remote sensing big data processing platform provided in the embodiment of the present application. DETAILED DESCRIPTION

[0015] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0016] Figure 1 This is one of the flow charts of the photovoltaic station refinement identification method based on classifier and deep learning provided in the embodiment of the present application, such as Figure 1 As shown, the method is applied to a remote sensing big data processing platform, such as a Google Earth Engine (GEE) platform, and the method includes at least the following steps: S101, preprocessing the multi-source remote sensing images of the photovoltaic stations to be identified within the research area; S102, inputting the preprocessed multi-source remote sensing image into a trained machine learning classifier for preliminary coarse resolution recognition, and obtaining a high-resolution remote sensing image containing a coarse recognition outline of a potential photovoltaic area; S103, inputting the high-resolution remote sensing image containing the rough identification outline of the potential photovoltaic area into the trained deep learning model for refined identification, and outputting a segmentation mask with panel type information, where the segmentation mask is used to distinguish photovoltaic panels and panel gaps.

[0017] The embodiment of the present application utilizes the GEE platform to quickly pre-process the multi-source remote sensing images of the photovoltaic stations to be identified within the study area; through the machine learning classifier, preliminary coarse-resolution identification is performed on the photovoltaic station area and background area to be identified within the study area, and the coarse identification outline of the potential photovoltaic area is identified, narrowing the identification range from a large area to a relatively small area; finally, the deep learning model is used to perform fine-grained identification of the relatively small identification range, and a segmentation mask with panel type information is output, which has a high spatial resolution and can distinguish photovoltaic panels and panel gaps, thereby obtaining fine-grained identification results of photovoltaic stations within the study area.

[0018] Figure 2 This is the second flow chart of the photovoltaic station refinement identification method based on classifier and deep learning provided in the embodiment of the present application, combined with Figure 1 and Figure 2 ,For S101, in the remote sensing big data processing platform, it is possible to quickly pre-process the multi-source remote sensing images of the photovoltaic stations to be identified within the research area.

[0019] The GEE platform is a cloud geospatial computing platform that supports free and available Pb-level remote sensing data, multiple machine learning algorithms, and shared computing resources, and can quickly obtain more accurate ground object classification results. Using the GEE platform to process multi-source remote sensing data can combine the data advantages of different remote sensing image sources.

[0020] The specific process of preprocessing can be as follows: first, declouding is performed through the quality assurance (QA) band to screen out low-quality remote sensing images; then, the multi-temporal images are merged, and the spatial resolution of the final preprocessed image is the minimum spatial resolution among the multi-source remote sensing images.

[0021] Optionally, the preprocessing process also includes customized image spectral information for subsequent machine learning classification algorithm input, including red, green, blue visible light band and near infrared band information, as well as characteristic indexes such as vegetation index, building index, and customized index.

[0022] For S102, the machine learning classifier on the GEE platform was used to perform preliminary coarse-resolution identification of photovoltaic stations on the preprocessed multi-source remote sensing images.

[0023] Optionally, the machine learning classifier is a random forest classifier, a binary classifier, etc., which can classify potential photovoltaic areas and background areas in the input remote sensing image.

[0024] In some embodiments, S102 specifically includes: Based on the data set related to the site selection principles for photovoltaic station construction, the preliminary coarse-resolution recognition results output by the machine learning classifier are screened and eliminated and morphologically operated to obtain remote sensing images containing potential photovoltaic areas; Among them, data sets related to the site selection principles for photovoltaic station construction include one or more of terrain slope and mountain shadow data, urban night light data, road data, population density data, and solar radiation intensity data.

[0025] Specifically, after the machine learning classifier outputs the classification results of potential photovoltaic areas and background areas in the remote sensing image, it is still necessary to screen and eliminate them through the data set related to the photovoltaic construction site selection principles and optimize them through morphological operations to eliminate noise, so as to obtain the final remote sensing image containing the potential photovoltaic area.

[0026] Optionally, the data sets related to the photovoltaic construction site selection principle include one or more of terrain slope and mountain shadow data, urban night light data, road data, population density data, and solar radiation intensity data. Among them, terrain slope and mountain shadow data are used to calculate terrain slope and mountain shadow; solar radiation intensity data mainly include data such as solar azimuth and altitude angle; population density data and urban night light data are used to eliminate identification elements in towns and high-density population areas.

[0027] In some embodiments, the machine learning classifier is trained based on the following steps: Obtain remote sensing images containing vector polygon outlines of photovoltaic stations within the research area; Within the vector polygon, PV points are uniformly and randomly selected based on the size of the vector polygon; Outside the vector polygon, NPV points equivalent to the number of PV points were uniformly and randomly selected based on the land use type stratification; The training set and validation set are divided based on the selected PV points and NPV points, and the machine learning classifier is trained.

[0028] Specifically, the photovoltaic station vector polygon in the published literature is used as the basis of the training data set. Within the research area, points are randomly selected uniformly from the vector polygon according to the area size of the vector polygon, recorded as PV; points are randomly selected uniformly from outside the vector polygon according to the land use type, recorded as NPV, to ensure that the number of PV points and NPV points is equal, and the construction of the training data set is completed. Other parameters remain the default of the GEE platform, and the training set and validation set are divided to train the machine learning classifier.

[0029] The process of preliminary coarse resolution identification of a photovoltaic station in the embodiment of the present application is further explained below through a specific example.

[0030] Example 1: In this embodiment, based on the multi-source remote sensing images of the GEE platform, the vector polygon data of photovoltaic stations in 2020 in the public literature is used as training data, and the random forest classifier is selected as the machine learning classifier to obtain the preliminary coarse-resolution recognition results of photovoltaic stations in China in 2021, which specifically includes the following stages: Data preparation stage: Sentinel-2 Harmonized 2A and Landsat 8 SR satellite image products were selected as multi-source image sources, and cloud removal was performed using the QA60 band and the QA Pixel band, respectively. The bands selected from the two image sets as inputs for the subsequent machine learning classifier are shown in Table 1.

[0031] Table 1

[0032] As shown in Table 1, in the Sentinel-2 Harmonized 2A image set, B2, B3, B4, B8, B11, and B12 are blue, green, red, near infrared, short infrared 1 (1.614μm), and short infrared 2 (2.202μm) bands, respectively, and B8_stdDev is the standard deviation of the near infrared band value in a year; in the Landsat 8 SR image set, SR_B2, SR_B3, SR_B4, SR_B5, SR_B6, and SR_B7 are blue, green, red, near infrared, short infrared 1 (1.566-1.651μm), and short infrared 2 (2.107-2.294μm) bands, respectively, EVI is the enhanced vegetation index, NDBI is the normalized building index, mNDWI is the modified normalized difference water index, and NDVI is the normalized vegetation index.

[0033] The two image sets are merged and median calculated to obtain the synthetic images of 2020 and 2021, which summarize all the bands and calculated indices in the two image sets, with 21 in each synthetic image. Among them, the two synthetic images of 2020 and 2021 are used for training and prediction of random forest classifiers, respectively.

[0034] Classifier training phase: The training data selected was the 2020 photovoltaic station vector polygon data with a spatial resolution of 30m from the public literature. Different numbers of points were randomly selected within the vector polygon according to the size of the polygon area, recorded as PV; outside the vector polygon, points were randomly selected on different types of land according to the ESA WorldCereal 10m land use type product, recorded as NPV. There are 40,000 PV points and 38,000 NPV points. 75% of the points were randomly divided into training sets and 25% into validation sets. The 21 band values ​​of the points contained in the training set were input into the random forest classifier for training. The number of trees was set to 500, and the other parameters were set by default on the GEE platform. The accuracy of the random forest classifier was verified using the validation set, and the accuracy was 0.881 and the Kappa coefficient was 0.723.

[0035] Optimization stage of coarse-resolution recognition results: After obtaining the preliminary coarse-resolution recognition results for 2021 through the random forest classifier, it is still necessary to further screen through the data sets related to the photovoltaic construction site selection principles. First, the solar azimuth angle is set to 180° and the solar altitude angle is set to 90°. The slope and mountain shadow are calculated according to the Shuttle Radar Topography Mission (30m spatial resolution) terrain data set, and the identification elements located in the area with a slope greater than 30 degrees and a mountain shadow less than 150 are removed. Secondly, since the research object is a large photovoltaic site, the identification elements in towns and high-density population areas are removed according to the VIIRS Nighttime urban night light data and the WorldPop Global Project Population data. Thirdly, pixels with less than 9 adjacent pixels are removed to eliminate noise. Finally, the circular kernel focalMax and focalMode filters are used to perform morphological operations on the recognition results, and the circular kernel radius is 1 pixel. Based on the above process, the preliminary coarse-resolution recognition results of photovoltaic sites in 2021 can be obtained, and Shapefile format files can be exported through the GEE platform. Similarly, the preliminary coarse-resolution identification results of photovoltaic stations in any year contained in the remote sensing image source used can also be obtained.

[0036] In the embodiment of the present application, a preliminary coarse-resolution identification process based on the GEE platform is performed to obtain potential photovoltaic areas at a coarse resolution by preprocessing image data, optimizing machine learning classifier input parameters, combining morphological operations and other related data sets to screen, and narrowing the area from a large area to the area to be predicted. Therefore, when using the deep learning model for subsequent prediction, there is no need to download all remote sensing images in the study area, saving memory space.

[0037] In order to further improve the recognition accuracy, a deep learning model architecture is adopted to optimize the refined photovoltaic site identification with high spatial resolution (below 10m).

[0038] For S103, a high-resolution remote sensing image containing the identification outline of the potential photovoltaic area is input into a trained deep learning model. The deep learning model can obtain refined identification results of photovoltaic sites within the research area by learning the boundaries and types of photovoltaic panels in the potential photovoltaic area.

[0039] In some embodiments, the deep learning model is trained based on the following steps: Acquire high-resolution remote sensing images containing the identified contours of potential photovoltaic areas; Target images are selected from high-resolution remote sensing images. The photovoltaic sites in the target images have diverse land use types and panel types and are evenly distributed within the research area. Input the target image into the web-based semi-automatic fine-grained annotation platform based on SAM for prediction, and output the segmentation mask as the image annotation result; Create a training set based on the image annotation results and panel type of the target image; Perform data augmentation on the training set and train the deep learning model.

[0040] Specifically, the Segment Anything Model (SAM) can achieve fast and accurate segmentation of any object in an image or video. During the regular training of the deep learning model, the sample images are annotated with the help of the web-based semi-automatic fine-grained annotation platform based on SAM.

[0041] Figure 2 : is a flow chart of the annotation process of the web version semi-automatic fine annotation platform provided in the embodiment of the present application, such as Figure 2 As shown in the figure, the platform can quickly and accurately annotate images of any spatial resolution, and is particularly advantageous for high-resolution images, because it can distinguish target objects from the background. High-resolution images include images taken by drones, commercial satellite images, etc. The platform uses the SAM tool with the help of user prompts and manual adjustments to achieve efficient semi-automatic fine annotation.

[0042] In some embodiments, the process of semi-automatic fine annotation is as follows: Step a: Create a high-resolution image to be labeled.

[0043] Users can upload high-resolution images to the platform code folder. The image format supports TIFF format and must contain geospatial information. The image size is customizable and the common 256*256 size can be selected.

[0044] Step b: prompt for input.

[0045] The user logs in to the initial interface of the platform. The platform loads one image at a time and displays the original image on the left side of the interface. The user specifies the area of ​​interest (foreground) on the original image, such as photovoltaic panels, buildings, etc., by clicking, selecting boxes, etc. as needed. The platform background runs the SAM model and predicts according to the prompts, generates a segmentation mask and displays it on the right side of the page; if the foreground that the user wants to annotate does not exist in the displayed image, it can be skipped and directly enter the annotation of the next image.

[0046] Step c: Model feedback and optimization.

[0047] The user can enter hints about the foreground and background on the original image multiple times, and the mask will change and display synchronously on the right side of the page after each update. When the user is satisfied with the annotation result, he can click "Next" to continue the annotation of the next image; at the same time, the final annotation result is automatically saved in the folder of the platform code in the form of 0 / 1 TIFF raster image, named the same as the original image.

[0048] Optionally, the semi-automatic fine annotation platform supports multiple users to annotate the same batch of images at the same time, and sets three states of "annotating", "completed", and "to be annotated" in the user annotation status data table, and only reads the images corresponding to the "to be annotated" state to ensure that multiple users will not annotate the same image repeatedly. The semi-automatic fine annotation platform is suitable for the annotation tasks of fine objects such as photovoltaic stations and buildings, and is suitable for application scenarios that require fast and efficient processing of a large number of images.

[0049] Furthermore, the specific training process of the deep learning model is as follows: First, obtain high-resolution remote sensing images containing the identification contours of potential photovoltaic areas. Secondly, select a number of target images from them. The photovoltaic stations in the target images have a diversity of land use types and panel types, and are relatively evenly distributed within the research area. Furthermore, a web-based semi-automatic fine-grained annotation platform is pre-established based on the automated image segmentation tool SAM, and the target images are manually annotated in turn to create a training set for the deep learning model. SAM receives user-specified foreground or background points, rectangular boxes or free text prompts, and can generate and update the corresponding segmentation masks. Finally, perform data augmentation operations on the training set to expand the sample size and enhance the generalization ability of the model in different scenarios. At the same time, the time of image shooting and satellite information are input into the model, and multiple technical information outputs can be customized to train a deep learning model.

[0050] In some embodiments, the deep learning model is further fine-tuned based on the following steps: Classify the target images in the training set based on land use types and create training subsets corresponding to different land use types; Based on the training subsets corresponding to different land use types, the deep learning models are fine-tuned respectively to obtain the fine-tuned deep learning models corresponding to different land use types.

[0051] Specifically, the deep learning model trained using the aforementioned steps is a general deep learning basic model. However, the background morphology of different land use types on satellite images varies greatly. The accuracy of the segmentation mask output by inputting the preliminary coarse-resolution recognition results obtained by the machine learning classifier into the general deep learning basic model needs to be improved.

[0052] In order to adapt to the morphological characteristics of photovoltaic sites with different land use types and improve the recognition accuracy of the deep learning model, the general deep learning basic model is fine-tuned using the land use type to obtain the fine-tuned deep learning model corresponding to the different land use types. In practical applications, the land use type is judged in advance, and the preliminary coarse resolution recognition results obtained based on the machine learning classifier are input into the fine-tuned deep learning model of the corresponding land use type.

[0053] The fine-tuning process is as follows: first, the target images in the training set of the aforementioned training process are classified based on the land use type, and training subsets corresponding to different land use types are created; using the training subsets corresponding to different land use types, the general deep learning basic model is fine-tuned respectively to obtain the fine-tuned deep learning models corresponding to different land use types.

[0054] Optionally, the land use type of the potential photovoltaic area obtained by the preliminary coarse resolution identification process is pre-judged, and the fine-tuned deep learning model corresponding to different land use types is used to finely identify the potential photovoltaic areas of different land use types, and finally generate an accurate photovoltaic site distribution map. This distribution map has a high spatial resolution and can distinguish photovoltaic panels from panel gaps, laying the foundation for obtaining high-quality photovoltaic site information, and can support subsequent tasks such as power prediction and emission reduction benefit evaluation.

[0055] The process of refined identification of photovoltaic stations in the embodiment of the present application is further explained below through a specific example.

[0056] Example 2: The deep learning model in this embodiment adopts the improved DeepLabV3+ architecture.

[0057] In this embodiment, the results obtained in Example 1 are refined and identified, and the panel type information is attempted to be obtained simultaneously. After downloading the 256*256 image of the screening area, a total of 6,000 manually annotated samples covering 122 photovoltaic stations are obtained through the web version of the semi-automatic refined annotation platform. In order to enable the model to learn the rules of panel types, the panel type information of these 122 stations is manually collected, including monocrystalline silicon, polycrystalline silicon and other types.

[0058] Once the label and panel type information of each image are prepared, the deep learning model can be used for training. In this embodiment, the DeepLabV3+ architecture and Xception41 encoder are selected to train the deep learning basic model. Figure 3 is a schematic diagram of the architecture of the deep learning model provided in the embodiment of the present application, such as Figure 3As shown in the figure, the RGB three channel values ​​of the image are used as the model input. After decoding, the shooting season and satellite type of the image are input in sequence, and the segmentation mask with panel type information is finally output. During the training process, the batch size is set to 36, the learning rate is set to 0.001, and the Adam optimizer is used for gradient optimization. The final validation set accuracy is 0.865.

[0059] Since different land use types present different background morphologies on satellite images, the models of photovoltaic stations with different land use types are fine-tuned and classified for training. In this embodiment, a fine-tuned deep learning model is established for 6 types of land use types (wasteland, farmland, forest, grassland, water body and others). For each fine-tuned deep learning model, the learning rate is set to 0.001, and the Adam optimizer is used to traverse the batch size in the interval [32, 48] to retrain the sub-samples, and the fine-tuned deep learning model with the highest accuracy corresponding to each land use type is selected as shown in Table 2. The land use type dataset comes from the publicly available 2021 CLCD China land cover type data, with a spatial resolution of 30m.

[0060] Table 2

[0061] It can be seen from Table 2 that after fine-tuning the model, the accuracy of the validation set in the training subsets of different land use types has been improved to varying degrees.

[0062] Figure 4 is a schematic diagram of the comparison between the preliminary coarse resolution recognition result and the refined recognition result provided in the embodiment of the present application, such as Figure 4 As shown, by comparing the difference between the refined recognition results in this embodiment and the 2020 30m coarse spatial resolution results used for training, it can be seen that the refined recognition results have a more complete recognition effect.

[0063] The photovoltaic station refinement identification method based on classifier and deep learning provided by the embodiment of the present application combines the multi-band spectral characteristics of remote sensing image data to perform multi-level identification and refinement processing on photovoltaic stations in a large range; by pre-processing image data, optimizing machine learning classifier input parameters, combining morphological operations and other related data sets to screen out potential photovoltaic areas at coarse resolution, and narrowing the scope of the area to be predicted; combining SAM, an automated semantic segmentation tool, to establish a web-based semi-automatic refinement annotation platform, using a deep learning model to refine the boundaries and types of photovoltaic panels; according to the land use type, a fine-tuning model is established, so that the final recognition result has the characteristics of clear boundaries and accurate morphology. The present application can quickly locate the potential area of ​​photovoltaic stations within the specified area and perform refined identification based on this, and minimize manual intervention in each step, thereby improving the accuracy and efficiency of identifying photovoltaic stations in complex and diverse areas, and laying the foundation for obtaining high-quality photovoltaic station information.

[0064] The photovoltaic station fine-grained identification device based on classifier and deep learning provided in the present application is described below. The photovoltaic station fine-grained identification device based on classifier and deep learning described below and the photovoltaic station fine-grained identification method based on classifier and deep learning described above can be referenced to each other.

[0065] Figure 6 is a schematic diagram of the structure of a photovoltaic station refinement identification device based on a classifier and deep learning provided in an embodiment of the present application, such as Figure 6 As shown, the device at least includes: A preprocessing module 601 is used to preprocess the multi-source remote sensing images of the photovoltaic stations to be identified within the research area; A preliminary coarse resolution recognition module 602 is used to input the pre-processed multi-source remote sensing image into a trained machine learning classifier for preliminary coarse resolution recognition, and obtain a high-resolution remote sensing image containing a coarse recognition outline of a potential photovoltaic area; The refined identification module 603 is used to input the high-resolution remote sensing image containing the rough identification outline of the potential photovoltaic area into the trained deep learning model for refined identification, and output a segmentation mask with panel type information. The segmentation mask is used to distinguish photovoltaic panels and panel gaps.

[0066] In some embodiments, the deep learning model is trained based on the following steps: Acquire high-resolution remote sensing images containing the identified contours of potential photovoltaic areas; Target images are selected from high-resolution remote sensing images. The photovoltaic sites in the target images have diverse land use types and panel types and are evenly distributed within the research area. Input the target image into the web-based semi-automatic fine-grained annotation platform based on SAM for prediction, and output the segmentation mask as the image annotation result; Create a training set based on the image annotation results and panel type of the target image; Perform data augmentation on the training set and train the deep learning model.

[0067] In some embodiments, the deep learning model is further fine-tuned based on the following steps: Classify the target images in the training set based on land use types and create training subsets corresponding to different land use types; Based on the training subsets corresponding to different land use types, the deep learning models are fine-tuned respectively to obtain the fine-tuned deep learning models corresponding to different land use types.

[0068] In some embodiments, acquiring a remote sensing image containing a potential photovoltaic area includes: Based on the data set related to the site selection principles for photovoltaic station construction, the preliminary coarse-resolution recognition results output by the machine learning classifier are screened and eliminated and morphologically operated to obtain remote sensing images containing potential photovoltaic areas; Among them, data sets related to the site selection principles for photovoltaic station construction include one or more of terrain slope and mountain shadow data, urban night light data, road data, population density data, and solar radiation intensity data.

[0069] In some embodiments, the machine learning classifier is trained based on the following steps: Obtain remote sensing images containing vector polygon outlines of photovoltaic stations within the research area; Within the vector polygon, PV points are uniformly and randomly selected based on the size of the vector polygon; Outside the vector polygon, NPV points equivalent to the number of PV points were uniformly and randomly selected based on the land use type stratification; The training set and validation set are divided based on the selected PV points and NPV points, and the machine learning classifier is trained.

[0070] In some embodiments, the machine learning classifier is a random forest classifier.

[0071] In some embodiments, the deep learning model adopts an improved DeepLabV3+ network, which includes a cascaded encoder, a decoder and a one-hot encoding module. The encoder adopts the Xception41 encoder, and the one-hot encoding module is used to perform one-hot encoding on the shooting season and satellite type of the input remote sensing image.

[0072] It can be understood that the detailed functional implementation of each of the above-mentioned units / modules can be found in the introduction of the aforementioned method embodiment, and will not be repeated here.

[0073] It should be understood that the above-mentioned device is used to execute the method in the above-mentioned embodiment. The implementation principle and technical effect of the corresponding program module in the device are similar to those described in the above-mentioned method. The working process of the device can refer to the corresponding process in the above-mentioned method, which will not be repeated here.

[0074] Based on the method in the above embodiment, the embodiment of the present application provides a remote sensing big data processing platform. The device may include: at least one memory for storing programs and at least one processor for executing the programs stored in the memory. When the program stored in the memory is executed, the processor is used to execute the method described in the above embodiment.

[0075] Figure 7 is a schematic diagram of the structure of the remote sensing big data processing platform provided in the embodiment of the present application, such as Figure 7 As shown, the remote sensing big data processing platform may include: a processor (Processor) 701, a communication interface (CommunicationsInterface) 702, a memory (Memory) 703 and a communication bus 704, wherein the processor 701, the communication interface 702, and the memory 703 communicate with each other through the communication bus 704. The processor 701 may call the software instructions in the memory 703 to execute the method described in the above embodiment.

[0076] In addition, the logic instructions in the above-mentioned memory 703 can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application.

[0077] Based on the method in the above embodiment, an embodiment of the present application provides a computer-readable storage medium, which stores a computer program. When the computer program runs on a processor, the processor executes the method in the above embodiment.

[0078] Based on the method in the above embodiment, an embodiment of the present application provides a computer program product. When the computer program product runs on a processor, the processor executes the method in the above embodiment.

[0079] It is understandable that the processor in the embodiment of the present application may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, transistor logic devices, hardware components or any combination thereof. The general-purpose processor may be a microprocessor or any conventional processor.

[0080] The method steps in the embodiments of the present application can be implemented by hardware or by a processor executing software instructions. The software instructions can be composed of corresponding software modules, and the software modules can be stored in random access memory (RAM), flash memory, read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, mobile hard disks, CD-ROMs, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor so that the processor can read information from the storage medium and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can be located in an ASIC.

[0081] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented by software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted through the computer-readable storage medium. The computer instructions may be transmitted from a website site, computer, server or data center to another website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more available media integrated. The available medium may be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive (SSD)), etc.

[0082] It should be understood that the various numerical numbers involved in the embodiments of the present application are only used for the convenience of description and are not used to limit the scope of the embodiments of the present application.

[0083] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

Claims

1. A photovoltaic station refinement identification method based on classifier and deep learning, characterized in that: Applied to a remote sensing big data processing platform, the method includes: Preprocess the multi-source remote sensing images of the photovoltaic stations to be identified within the research area; The pre-processed multi-source remote sensing images are input into the trained machine learning classifier for preliminary coarse resolution recognition, and a high-resolution remote sensing image containing the coarse recognition outline of the potential photovoltaic area is obtained; The high-resolution remote sensing image containing the rough identification outline of the potential photovoltaic area is input into the trained deep learning model for refined identification, and a segmentation mask with panel type information is output, wherein the segmentation mask is used to distinguish photovoltaic panels and panel gaps.

2. The photovoltaic station refined identification method according to claim 1 is characterized in that: The deep learning model is trained based on the following steps: Acquire high-resolution remote sensing images containing the identified outlines of potential photovoltaic areas; Selecting target images from the high-resolution remote sensing images, wherein the photovoltaic sites in the target images have a diversity of land use types and panel types and are evenly distributed within the research area; Input the target image into a web-based semi-automatic fine-grained annotation platform based on SAM for prediction, and output a segmentation mask as an image annotation result; Create a training set based on the image annotation results and panel type of the target image; Performing data augmentation operations on the training set to train the deep learning model.

3. The photovoltaic station refined identification method according to claim 2 is characterized in that: The deep learning model is also fine-tuned based on the following steps: Classifying the target images in the training set based on land use types, and creating training subsets corresponding to different land use types; Based on the training subsets corresponding to different land use types, the deep learning model is fine-tuned respectively to obtain the fine-tuned deep learning models corresponding to different land use types.

4. The photovoltaic station refined identification method according to claim 1 is characterized in that: The step of obtaining a remote sensing image containing a potential photovoltaic area includes: Based on a data set related to the photovoltaic site selection principle, the preliminary coarse-resolution recognition results output by the machine learning classifier are screened and eliminated and morphologically operated, and the remote sensing image containing the potential photovoltaic area is obtained; The data sets related to the site selection principles for photovoltaic station construction include one or more of terrain slope and mountain shadow data, urban night light data, road data, population density data and solar radiation intensity data.

5. The photovoltaic station refined identification method according to claim 1 or 4, characterized in that: The machine learning classifier is trained based on the following steps: Acquire a remote sensing image containing vector polygon outlines of photovoltaic stations within the research area; Within the vector polygon, PV points are uniformly and randomly selected based on the size of the vector polygon; Outside the vector polygon, NPV points equivalent to the number of PV points are uniformly and randomly selected based on the land use type stratification; The training set and the validation set are divided based on the selected PV points and NPV points to train the machine learning classifier.

6. The photovoltaic station refined identification method according to claim 1 is characterized in that: The machine learning classifier is a random forest classifier.

7. The photovoltaic station refined identification method according to claim 1 is characterized in that: The deep learning model adopts an improved DeepLabV3+ network, which includes a cascaded encoder, a decoder and a one-hot encoding module. The encoder adopts an Xception41 encoder, and the one-hot encoding module is used to perform one-hot encoding on the shooting season and satellite type of the input remote sensing image.

8. A photovoltaic station refined identification device based on classifier and deep learning, characterized in that: include: A preprocessing module is used to preprocess the multi-source remote sensing images of the photovoltaic stations to be identified within the research area; The preliminary coarse resolution recognition module is used to input the pre-processed multi-source remote sensing images into the trained machine learning classifier for preliminary coarse resolution recognition, and obtain high-resolution remote sensing images containing the coarse recognition outline of the potential photovoltaic area; The refined recognition module is used to input the high-resolution remote sensing image containing the rough recognition outline of the potential photovoltaic area into the trained deep learning model for refined recognition, and output a segmentation mask with panel type information, wherein the segmentation mask is used to distinguish photovoltaic panels and panel gaps.

9. A remote sensing big data processing platform, characterized in that: include: at least one memory for storing a computer program; At least one processor is used to execute the program stored in the memory. When the program stored in the memory is executed, the processor is used to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program runs on a processor, the processor is caused to execute the method according to any one of claims 1 to 7.

Images