Distributed photovoltaic site selection method

Abstract

The invention discloses a distributed photovoltaic site selection method, and the method comprises the steps: designing a clustering loss function with a consistent structure for a semantic analysis model, thereby enabling the features of each class to tend to be uniformly distributed, reducing the feature difference in different object classes, enlarging the distance between the classes, and improving the site selection accuracy. Therefore, more accurate edge perception is realized, structural adaptation of pixel-level categories is solved, and the problem of accurately extracting an adaptive photovoltaic laying roof from small and dense roof targets of a remote sensing image is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of semantic segmentation, and in particular relates to a distributed photovoltaic site selection method. Background Art

[0002] The distributed photovoltaic site selection method first needs to identify and extract building roof spots from remote sensing images, and make a preliminary judgment on whether to select photovoltaic sites based on the identified and extracted building roof spots. At present, the process of identifying and extracting building roof spots is mainly based on the semantic segmentation model of convolutional neural network. This type of model relies heavily on pixel-level semantic labels, and its annotation of source data requires intensive manual labor. In order to alleviate this problem, researchers have proposed the method of synthetic datasets / synthetic domains. However, since the above models trained on synthetic datasets / synthetic domains cannot be well generalized to real-world scenarios, because there is a clear domain gap between the two datasets / domains, the trained models are actually not effective for target extraction, especially for small and dense targets. For example, in the distributed rooftop photovoltaic survey, small and dense roof targets are difficult to extract accurately.

[0003] Some researchers have used transfer learning in the supervised neural network field of semantic segmentation to solve the extraction of small and dense targets in remote sensing patches, but there is still the problem of inaccurate extraction of building roof patches. This is because the existing transfer learning methods are often limited in generalization ability when faced with new data sets with different distributions, mainly due to the significant domain differences between the source domain and the target domain. At present, some researchers have adopted the supervised domain adaptation (UDA, Unsupervised Domain Adaptation) strategy for cross-domain semantic segmentation, aiming to effectively transfer the knowledge learned from the fully labeled source domain to the unlabeled target domain. The mainstream UDA method mostly uses adversarial training methods in order to learn domain-invariant feature representations and discriminative individual features. Specifically, when setting, there are two domains: source domain S = {X s , Y s} and target domain T = {X t}, one of the samples comes from the source domain X s ∈R H×W×3 and its corresponding ground truth label y s ∈(1, C) H×W , but one belongs to the target domain X t ∈R H×W×3 There are no corresponding labels. Well-annotated source datasets are mainly used to train segmentation models in a supervised manner. The unlabeled target dataset is combined with the source domain, and adversarial training is used to minimize the domain gap and achieve the migration between the source domain and the target domain. Therefore, the general objective function L(X s , Xt , Y s ) is defined as follows:

[0004]

[0005] Among them, L Seg and L adv are semantic loss and adversarial loss respectively, and η is their trade-off parameter. Seg is a standard cross entropy loss, where P s (h，w，c) Represents the model prediction probability of class c at pixel (h, w). C is the number of classes, H, W 120 is the height and width of the image. s (h、w、c) Represents pixel x s (h、w) The ground truth labels. is a segmentation model, f(·; θ) represents the feature embedding function, which maps the input sample xs to the feature space F. The following classifier Output prediction result P s (h，w，c) θ and are f and The model parameters.

[0006] However, most of these methods ignore the edge perception problem of category structure, which is particularly prominent when dealing with small-sized and densely distributed objects, and easily leads to category mismatch, so the extraction effect of building roof patches is still poor. This is because this adversarial alignment method is unaware of the edge because they do not adjust the intrinsic class-conditional structure. In this case, features with similar structures are difficult to distinguish because they are located in the boundary area. These fuzzy structures confuse the classifier and mistake some small objects as large-scale backgrounds, especially when extracting roofs, this phenomenon is more significant. Summary of the invention

[0007] In order to solve the problems existing in the prior art, the present invention proposes a distributed photovoltaic site selection method.

[0008] The technical solution of the present invention is as follows:

[0009] A distributed photovoltaic site selection method, comprising:

[0010] Preprocessing the acquired remote sensing image with the building roof to obtain a source domain consisting of a source image and a label of the source image and a target domain consisting of a target image;

[0011] A semantic analysis model for building roof extraction including a generator, a classifier and a discriminator is constructed, wherein the objective function of the semantic analysis model includes a semantic loss for optimizing parameters of the generator and the classifier, a source domain prototype-based source unified clustering loss and a target domain empirical prototype-based target unified clustering loss for minimizing the intra-class difference of the generator and driving the uniform distribution of different classes relative to the classifier while forming edge-aware class-conditional clusters, and a discriminator adversarial loss for eliminating domain gaps by distinguishing domain labels of entropy mapping;

[0012] Based on the source domain and the target domain, the semantic analysis model is trained to transfer learning from the source image and the label of the source image to the target image to obtain the semantic label of the target image;

[0013] Based on the semantic label of the target image and the building POI information corresponding to the target image, the effective area of ​​the photovoltaic site selection and installation in the target image is obtained and compared with the preset minimum installation area to form a distributed photovoltaic site selection plan.

[0014] Furthermore, the specific process of training the transfer learning from the source image and the label of the source image to the target image includes:

[0015] Inputting the source image and the label domain of the source image into the generator, and generating a pseudo label with confidence semantic features;

[0016] The classifier is trained based on the labels of the original image and the pseudo labels, and generates entropy maps of the source image and the target image respectively to reflect the uncertainty of the pixels in the image;

[0017] The discriminator is trained based on the entropy map and the pseudo-labels to transfer the source domain knowledge to the target domain and distinguish between the source domain and the target domain.

[0018] Furthermore, the objective function of the semantic analysis model is:

[0019]

[0020] Where, X S , Y S and X t are the source image, the label of the source image and the target image respectively; and They are semantic loss, source unified clustering loss, target unified clustering loss and adversarial loss respectively.

[0021] Furthermore, the semantic loss The expression is:

[0022]

[0023] In the formula, represents the model predicted probability of category c at pixel (h,w); C is the number of categories, H and W are the height and width of the image respectively; Represents ground pixel The true label of; f(·; θ) represents the feature embedding function, which takes the input sample x s Mapped to feature space F, then the classifier Output prediction results θ and are the model parameters of functions f and φ respectively.

[0024] Furthermore, the source unified clustering loss The expression is:

[0025]

[0026] In the formula, is the activation function, if the pixel The ground truth label of If it belongs to category c, the function outputs 1, otherwise 0; is a prototype cluster set; represents the features of all pixels related to the same category c relative to f(·; θ); Indicates that the entropy of the classifier is maximized. The larger the entropy value, the better the class conditional distribution. The greater the discreteness.

[0027] Furthermore, the target unified clustering loss The expression is:

[0028]

[0029] In the formula, represents the pseudo-label of the target data and is expressed as is to use the predicted probability Features Empirical prototypes obtained by re-weighting; represents all samples from category c; Represents the empirical label distribution on the target domain.

[0030] Furthermore, the adversarial loss The expression is:

[0031]

[0032] In the formula, represents the pseudo label of the target domain; D represents and Entropy mapping of two domains.

[0033] Furthermore, the specific steps of obtaining the effective area of ​​photovoltaic site selection and installation in the target image based on the semantic label of the target image and the building POI information corresponding to the target image, and comparing it with the preset minimum installation area to form a distributed photovoltaic site selection plan include:

[0034] Based on the semantic labels of the target image, the building roof pattern is obtained, and combined with the building POI information, the effective area of ​​the photovoltaic site selection and installation corresponding to the building roof pattern in the target image is obtained. If the effective area is less than or equal to the preset minimum installation area, the area corresponding to the building roof pattern in the target image is not suitable for installing solar panels, and the area corresponding to the building roof pattern in the target image is not suitable as a distributed photovoltaic site selection plan; otherwise, the area corresponding to the building roof pattern in the target image is suitable for installing solar panels, and the area corresponding to the building roof pattern in the target image is suitable as a distributed photovoltaic site selection plan.

[0035] Furthermore, the preset minimum installation area is 25m 2 .

[0036] Furthermore, the building POI information includes building name, type, and house number information.

[0037] An electronic device comprises a memory and a processor, wherein the memory stores a computer program, and the processor is used to call and run the computer program stored in the memory to execute any of the methods described above.

[0038] A computer-readable storage medium stores a computer program, wherein the computer program implements the steps of any of the above methods when executed by a processor.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] The present invention proposes a distributed photovoltaic site selection method. The method designs a structural consistency clustering loss function for the semantic analysis model, thereby promoting the uniform distribution of features of each category, thereby reducing the feature differences within different object categories and simultaneously expanding the distance between categories, thereby achieving more accurate edge perception to solve the structural adaptation of pixel-level categories, thereby solving the problem of accurately extracting roofs suitable for photovoltaic laying from small and dense roof targets in remote sensing images. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 A flow chart of a distributed photovoltaic site selection method in an embodiment;

[0042] Figure 2A schematic diagram of the architecture of the semantic analysis model in the embodiment;

[0043] Figure 3 This is the distributed photovoltaic distribution diagram on the roof of an industrial park in Suzhou in the application example. DETAILED DESCRIPTION

[0044] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent modifications of the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0045] Embodiment 1:

[0046] A distributed photovoltaic site selection method of the present invention, such as Figure 1 As shown, including:

[0047] S1. Preprocessing the acquired remote sensing image with the building roof to obtain a source domain consisting of a source image and a label of the source image and a target domain consisting of a target image;

[0048] S2. Construct a semantic analysis model for building roof extraction that includes a generator, a classifier, and a discriminator. The objective function of the semantic analysis model includes a semantic loss for optimizing the parameters of the generator and the classifier, a source domain prototype-based source unified clustering loss and a target domain empirical prototype-based target unified clustering loss for minimizing the intra-class difference of the generator and driving the uniform distribution of different classes relative to the classifier while forming edge-aware class-conditional clusters, and a discriminator adversarial loss for eliminating domain gaps by distinguishing domain labels of entropy mapping;

[0049] S3. Based on the source domain and the target domain, the semantic analysis model is trained to transfer learning from the source image and the label of the source image to the target image to obtain the semantic label of the target image;

[0050] S4. Based on the semantic label of the target image and the building POI information corresponding to the target image, the effective area of ​​the photovoltaic site selection and installation in the target image is obtained, and compared with the preset minimum installation area to form a distributed photovoltaic site selection plan.

[0051] Embodiment 2:

[0052] This embodiment is further designed on the basis of the first embodiment in that: Figure 2 As shown in FIG. 1 , the specific process of training the source image and the label of the source image to transfer learning to the target image in this example includes:

[0053] Input the source image and the label domain of the source image into the generator, and generate a pseudo label with confidence semantic features;

[0054] The classifier is trained based on the labels of the original image and the pseudo labels, and generates entropy maps of the source image and the target image respectively to reflect the uncertainty of the pixels in the image;

[0055] The discriminator is trained based on the entropy map and pseudo labels to transfer the source domain knowledge to the target domain and distinguish between the source domain and the target domain.

[0056] Embodiment three:

[0057] This embodiment is further designed on the basis of the second embodiment in that the objective function of the semantic analysis model in this embodiment is:

[0058]

[0059] Where, X S , Y S and X t are the source image, the label of the source image, and the target image respectively; and L are semantic loss, source unified clustering loss, target unified clustering loss and adversarial loss respectively.

[0060] Embodiment 4:

[0061] This embodiment is further designed on the basis of the third embodiment in that the semantic loss in this embodiment The expression is:

[0062]

[0063] In the formula, represents the model predicted probability of category c at pixel (h,w); C is the number of categories, H and W are the height and width of the image respectively; Represents ground pixel The true label of; f(·; θ) represents the feature embedding function, which takes the input sample x s Mapped to feature space F, then the classifier Output prediction results θ and are the model parameters of functions f and φ respectively.

[0064] Embodiment five:

[0065] This embodiment is further designed on the basis of the third embodiment in that the source unified clustering loss in this embodiment The expression is:

[0066]

[0067] In the formula, is the activation function, if the pixel The ground truth label of If it belongs to category c, the function outputs 1, otherwise 0; is a prototype cluster set; represents the features of all pixels related to the same category c relative to f(·; θ); Indicates that the entropy of the classifier is maximized. The larger the entropy value, the better the class conditional distribution. The greater the discreteness.

[0068] Embodiment six:

[0069] This embodiment is further designed on the basis of the third embodiment in that the target unified clustering loss in this embodiment The expression is:

[0070]

[0071] In the formula, represents the pseudo-label of the target data and is expressed as is to use the predicted probability Features Empirical prototypes obtained by re-weighting; represents all samples from category c; Represents the empirical label distribution on the target domain.

[0072] Embodiment seven:

[0073] This embodiment is further designed on the basis of the third embodiment in that the adversarial loss in this embodiment The expression is:

[0074]

[0075] In the formula, represents the pseudo label of the target domain; D is the discriminator; is the entropy graph of the source domain; is the entropy graph of the target domain.

[0076] Embodiment eight:

[0077] This embodiment is further designed on the basis of the first embodiment in that, in this embodiment, based on the semantic label of the target image and the building POI information corresponding to the target image, the effective area of ​​the photovoltaic site selection and installation in the target image is obtained and compared with the preset minimum installation area. The specific steps of forming a distributed photovoltaic site selection plan include:

[0078] The building roof pattern is obtained based on the semantic label of the target image, and combined with the building POI information, the effective area of ​​the photovoltaic site selection and installation corresponding to the building roof pattern in the target image is obtained. If the effective area is less than or equal to the preset minimum installation area, the area corresponding to the building roof pattern in the target image is not suitable for installing solar panels, and the area corresponding to the building roof pattern in the target image is not suitable as a distributed photovoltaic site selection plan; otherwise, the area corresponding to the building roof pattern in the target image is suitable for installing solar panels, and the area corresponding to the building roof pattern in the target image is suitable as a distributed photovoltaic site selection plan.

[0079] Embodiment nine:

[0080] This embodiment is further designed on the basis of the eighth embodiment in that the preset minimum installation area in this embodiment is 25m 2 .

[0081] Embodiment ten:

[0082] This embodiment is further designed on the basis of the eighth embodiment in that the building POI information in this example includes information such as the building name, type, house number, etc.

[0083] Embodiment eleven:

[0084] An electronic device comprises a memory and a processor, wherein the memory stores a computer program, and the processor is used to call and run the computer program stored in the memory to execute any of the methods described above.

[0085] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of any of the above methods are implemented.

[0086] Application examples:

[0087] In this example, the distributed photovoltaic site selection method of the present invention is tested on the remote sensing image of an industrial park in Jiangsu. The total area of ​​the industrial park is 278.28 km 2 The rooftop distributed photovoltaic distribution extracted by the method of the present invention is as follows: Figure 3 As shown, the total extracted building roof area is about 32613746.38m2, the area that can be installed with photovoltaic panels is about 18732038.60m2, and the installed capacity is about 3336.52MW; among them, the total developed area is 3415917.04m2, the area that can be installed with photovoltaic panels is 2253999.06m2, and the installed capacity is 402.83MW; the total area suitable for development is about 25112550.44m2, the area that can be installed with photovoltaic panels is about 16478039.54m2, and the installed capacity is about 2933.69MW; the total area not suitable for development is about 4085278.90m2.

[0088] The above description is only a specific implementation of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by a technician familiar with the technical field within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention.

Claims

1. A distributed photovoltaic site selection method, characterized in that: include: Preprocessing the acquired remote sensing image with the building roof to obtain a source domain consisting of a source image and a label of the source image and a target domain consisting of a target image; A semantic analysis model for building roof extraction including a generator, a classifier and a discriminator is constructed, wherein the objective function of the semantic analysis model includes a semantic loss for optimizing parameters of the generator and the classifier, a source domain prototype-based source unified clustering loss and a target domain empirical prototype-based target unified clustering loss for minimizing the intra-class difference of the generator and driving the uniform distribution of different classes relative to the classifier while forming edge-aware class-conditional clusters, and a discriminator adversarial loss for eliminating domain gaps by distinguishing domain labels of entropy mapping; Based on the source domain and the target domain, the semantic analysis model is trained to transfer learning from the source image and the label of the source image to the target image to obtain the semantic label of the target image; Based on the semantic label of the target image and the building POI information corresponding to the target image, the effective area of ​​the photovoltaic site selection and installation in the target image is obtained and compared with the preset minimum installation area to form a distributed photovoltaic site selection plan.

2. The distributed photovoltaic site selection method according to claim 1, characterized in that: The specific process of training the transfer learning from the source image and the label of the source image to the target image includes: Inputting the source image and the label domain of the source image into the generator, and generating a pseudo label with confidence semantic features; The classifier is trained based on the labels of the original image and the pseudo labels, and generates entropy maps of the source image and the target image respectively to reflect the uncertainty of the pixels in the image; The discriminator is trained based on the entropy map and the pseudo-labels to transfer the source domain knowledge to the target domain and distinguish between the source domain and the target domain.

3. The distributed photovoltaic site selection method according to claim 2, characterized in that: The objective function of the semantic analysis model is: Where, X S , Y S and X t are the source image, the label of the source image, and the target image respectively; and They are semantic loss, source unified clustering loss, target unified clustering loss and adversarial loss respectively.

4. The distributed photovoltaic site selection method according to claim 3, characterized in that: The semantic loss The expression is: In the formula, represents the model predicted probability of category c at pixel (h,w); C is the number of categories, H and W are the height and width of the image respectively; Represents ground pixel The true label of; f(·; θ) represents the feature embedding function, which takes the input sample x s Mapped to feature space F, then the classifier Output prediction results θ and are the model parameters of functions f and φ respectively.

5. The distributed photovoltaic site selection method according to claim 3, characterized in that: The source unified clustering loss The expression is: In the formula, is the activation function, if the pixel The ground truth label of If it belongs to category c, the function outputs 1, otherwise 0; is a prototype cluster set; represents the features of all pixels related to the same category c relative to f(·; θ); Indicates that the entropy of the classifier is maximized; the larger the entropy value, the higher the class conditional distribution The greater the discreteness.

6. The distributed photovoltaic site selection method according to claim 3, characterized in that: The objective is to unify the clustering loss The expression is: In the formula, represents the pseudo-label of the target data and is expressed as is to use the predicted probability Features Empirical prototypes obtained by re-weighting; represents all samples from category c; Represents the empirical label distribution on the target domain.

7. The distributed photovoltaic site selection method according to claim 3, characterized in that: The adversarial loss The expression is: In the formula, represents the pseudo label of the target domain; D represents and Entropy mapping of two domains.

8. The distributed photovoltaic site selection method according to claim 1, characterized in that: The specific steps of obtaining the effective area of ​​photovoltaic site selection and installation in the target image based on the semantic label of the target image and the building POI information corresponding to the target image, and comparing it with the preset minimum installation area to form a distributed photovoltaic site selection plan include: Based on the semantic labels of the target image, the building roof pattern is obtained, and combined with the building POI information, the effective area of ​​the photovoltaic site selection and installation corresponding to the building roof pattern in the target image is obtained. If the effective area is less than or equal to the preset minimum installation area, the area corresponding to the building roof pattern in the target image is not suitable for installing solar panels, and the area corresponding to the building roof pattern in the target image is not suitable as a distributed photovoltaic site selection plan; otherwise, the area corresponding to the building roof pattern in the target image is suitable for installing solar panels, and the area corresponding to the building roof pattern in the target image is suitable as a distributed photovoltaic site selection plan.

9. The distributed photovoltaic site selection method according to claim 8, characterized in that: The preset minimum installation area is 25m 2 .

10. The distributed photovoltaic site selection method according to claim 8, characterized in that: The building POI information includes building name, type, and house number information.

Images