Photovoltaic panel fine identification and segmentation method based on photovoltaic power station scene knowledge constraint

Abstract

The invention discloses a photovoltaic panel fine identification segmentation method based on photovoltaic power station scene knowledge constraint, and the method comprises the following steps: carrying out the thinning of a photovoltaic power station scene remote sensing image Iscene; the method comprises the following steps of: making a scene segmentation sample set Sscenelabel of the photovoltaic power station; performing training and knowledge reasoning on the photovoltaic power station scene model; the method comprises the following steps: manufacturing a photovoltaic panel instance segmentation sample set Spanellabel; a photovoltaic panel identification model Mpanel is trained; reasoning a photovoltaic panel result set of a single photovoltaic power station scene; and carrying out multi-scene photovoltaic panel achievement integration. The photovoltaic panel fine identification cascade model formed by the method does not depend on remote sensing image time phase change and territorial scope, the defect that a traditional method is low in false target discrimination capability is overcome, and the identification reliability of the photovoltaic panel in a complex scene is improved; the management efficiency is improved and the management cost is reduced for the management and maintenance of the photovoltaic array in the photovoltaic power industry, and a good support effect of a remote sensing technology is exerted.

Description

Technical Field

[0001] The present invention belongs to the field of photovoltaic panel target extraction from remote sensing images combined with scene knowledge, and specifically relates to a photovoltaic panel refined identification and segmentation method constrained by photovoltaic power station scene knowledge. Background Art

[0002] As a renewable and clean energy, solar energy has the advantages of low cost, wide application, and low emissions. It is an important part of the global energy revolution from traditional non-renewable energy to new renewable energy. By taking advantage of the large-scale, high-precision, and high-efficiency monitoring of remote sensing, high-precision and refined identification of photovoltaic panels based on remote sensing images can timely grasp the spatial scale and distribution of solar photovoltaic panels and reduce the monitoring and management costs of large-scale photovoltaic panels.

[0003] Photovoltaic panels, as the core components of photovoltaic power stations, show obvious dark tones and regular features in remote sensing images. Traditional machine learning methods based on spectral statistics (maximum likelihood method, support vector machine, etc.) can usually achieve certain accuracy recognition results, but there is still a big gap from engineering application. With the development of deep learning algorithms in the field of images, the recognition of remote sensing photovoltaic panels has begun to turn to deep intelligent network extraction algorithms based on semantic segmentation, mainly using encoder-decoder architecture. Usually, the continuous deepening of feature extraction networks can learn the optimal deep remote sensing image features of photovoltaic panels, which has great advantages over feature expressions based on traditional manual design. The extraction efficiency and accuracy of remote sensing photovoltaic panels have been greatly improved, especially when the remote sensing background is relatively simple (desert, grassland, plain, water body, etc.), basically reaching the level of engineering application.

[0004] However, when faced with complex scenes, especially in urban areas and densely built-up areas, the image features of photovoltaic panels are similar to those of certain buildings, roads or large vehicles. Even in drone images with high spatial resolution, the discrimination of target features will be affected to a certain extent. There is also a large redundancy in the feature information with pseudo-targets, which in turn affects the precision of photovoltaic panel extraction in remote sensing images over a large range. In addition, splicing training and inference are performed in a block-by-block manner over a large range, which further loses the contextual relationship features of photovoltaic panels over a large range, resulting in low accuracy and high false alarm rate in photovoltaic panel recognition in complex scenes.

[0005] Therefore, the phenomenon of different objects with the same spectrum is more obvious in complex scenes. It is difficult to separate photovoltaic panels from background features by using only a semantic segmentation framework. It is necessary to further combine the scene knowledge of the photovoltaic electric field itself at the global scale, and use spatial analysis and scene constraint algorithms to integrate them into the photovoltaic panel target recognition process, overcome the interference factors of complex background information, and further improve the refined recognition of photovoltaic panels. Summary of the invention

[0006] In order to overcome the limitations of the single recognition network framework of remote sensing photovoltaic panels in complex scenes, the purpose of the present invention is to provide a photovoltaic panel refined recognition and segmentation method constrained by photovoltaic power station scene knowledge. The present invention is aimed at photovoltaic panels in complex urban backgrounds of remote sensing images. First, low-resolution large-scale remote sensing images are used to produce target samples and train network models for photovoltaic power station scenes to obtain the overall spatial range of photovoltaic power station scenes; secondly, within the scope of the photovoltaic power station scene, high-definition remote sensing images are used to train and reason about the deep semantic segmentation network of small-scale photovoltaic panels to obtain specific photovoltaic panel semantic information; finally, all photovoltaic panel data in the photovoltaic power station scene are vectorized and geographically mosaicked to form the target results of photovoltaic panels located in the photovoltaic power station scene in the whole scene image. The overall algorithm process further cascades the refined recognition of small-scale photovoltaic panels by integrating the large-scale spatial knowledge of photovoltaic power stations, forming a precision-refined target recognition algorithm nested with multi-scale scene knowledge.

[0007] The present invention provides a photovoltaic panel refined identification and segmentation method constrained by photovoltaic power station scene knowledge, comprising the following steps:

[0008] S1. Sparsely sampled photovoltaic power station scene remote sensing images I scene ;

[0009] S2. Create a photovoltaic power station scene segmentation sample set S scene_label ;

[0010] S3, training and knowledge reasoning of the photovoltaic power station scenario model;

[0011] S4. Create a photovoltaic panel instance segmentation sample set S panel_label ;

[0012] S5. Training photovoltaic panel recognition model M panel ;

[0013] S6, photovoltaic panel result set reasoning for a single photovoltaic power station scenario;

[0014] S7. Integration of photovoltaic panel results in multiple scenarios.

[0015] In some specific implementations, all remote sensing images are mainly true color bands.

[0016] In some specific implementations, in step S1, the photovoltaic power station scene remote sensing image I scene Relative to the original image I PV The downsampling ratio is usually 8 times;

[0017] That is, the following relationship is satisfied: PV =8×Resoscene , where Reso PV and Reso scene I PV and I scene The spatial resolution of remote sensing data is reduced by using bilinear interpolation or cubic convolution interpolation.

[0018] In some specific implementations, the method of step S2 is specifically as follows: the scene polygons in the photovoltaic power station scene sample set are usually the photovoltaic panel block cluster range, and the vectors are compared with the corresponding I scene The image is cropped to 1000*1000 pixels, where the scene label value is in the form of a PNG image with the inside of the vector marked as 1 and the outside marked as 0;

[0019] The random rotation geometric transformation in step S2 rotates the samples at random angles (usually multiples of 30 degrees) to retain the original sample image size. Taking the rotation of θ as an example, the scene image samples and label samples after rotation enhancement follow the following calculation formula:

[0020]

[0021] Among them, Im θ and La θ They are the rotated and enhanced matrices of the image matrix Im and label matrix La of a pair of photovoltaic scene training samples.

[0022] In some specific implementations, the specific method of step S3 is: the photovoltaic power station scene knowledge reasoning object is the upsampled remote sensing image I of the demonstration area scene ', the inference process uses a 1000*1000 pixel frame and 500 steps to traverse, and the inference vectorization result R scene Keep with I scene 'Same spatial coordinate system and geographic range.

[0023] In some specific implementations, the specific method of step S4 is: photovoltaic panel mask data Y panel It is usually stored in ESRShape format, where the attribute table contains the tag information of each photovoltaic panel, including the center point X coordinate, center point Y coordinate, rotation box width, rotation box height, and the rotation box angle α is defined as the acute angle between the center line of the rotation box and the x-axis with the horizontal axis of the center of the rotation box as the x-direction, with the counterclockwise rotation of the x-axis as positive and the clockwise rotation as negative. Then, the value range of α is [-90°, 90°];

[0024] In step S4, panel_labelThe sample label file is in json format. Since photovoltaic panels are directional and sensitive to directional information, random rotation enhancement in json format is required during sample enhancement. The basic principle is consistent with S2. The random change angle of the enhanced rotation angle θ is a multiple of 15 degrees. The training sample label file records the width, height, center point coordinates relative to the image sample, and rotation angle α of each photovoltaic panel target after enhancement.

[0025] In some specific implementations, the specific method of step S5 is: using the loss function in the instance segmentation in Mask_RCNN based on the rotation angle requires regressing the angle difference function of the rotation angle α, wherein the loss function of the angle loss part is shown as follows:

[0026] L α =α * Θα

[0027] Where α * is the model prediction angle, L α The difference between the angle of the true value α in the sample label.

[0028] In some specific implementations, the specific method of step S6 is: the cascade process of photovoltaic panel reasoning and photovoltaic power station scene reasoning, firstly, the coupled photovoltaic power station scene target result set R scene Mask rasterization is performed and compared with the original remote sensing image of the demonstration area. PV 'Perform logical AND operation, and then use 512*512 pixel frame and 256 steps to perform traversal reasoning, and finally obtain the semantic vectorization result R of the photovoltaic panel in the i-th photovoltaic power station scenario panel,i .

[0029] In some specific implementations, the specific method of step S7 is: when there is an area with overlapping photovoltaic electric field space in the entire demonstration area, it is necessary to use a non-extreme value suppression algorithm to eliminate repeated photovoltaic panel vector results, and the selected screening threshold is η≤0.4.

[0030] In some specific implementations, the final result RPV obtained in step S7 Panel It is saved in a vector file format. Its coordinate system information is consistent with the remote sensing image of the area and has a real geographical location. The file format is usually saved as ESRI Share File (.shp) or ESRI File GeoDatabase (.gdb).

[0031] The beneficial effects of the present invention are as follows: the present invention discloses a photovoltaic panel refined identification and segmentation method constrained by photovoltaic power station scene knowledge, and the beneficial effects of the method are mainly reflected in: ① The present invention adopts a double-layer serial cascade network of photovoltaic power station scene recognition network and photovoltaic panel recognition network to automatically identify photovoltaic panel information, which has higher photovoltaic panel recognition accuracy than the conventional recognition model relying on a single network model; ② The present invention integrates the concept of multi-scale remote sensing feature information, determines the scene geographic range of a large area of ​​a photovoltaic power station with low spatial resolution, and then performs targeted high spatial resolution photovoltaic panel refinement extraction based on the power station scene, simplifies the complex background of remote sensing images in the photovoltaic panel area, not only improves the photovoltaic panel recognition accuracy, but also improves the recognition speed of photovoltaic panels in areas with smaller photovoltaic ranges; ③ The present invention adopts a multi-type enhancement mode of scene recognition and rotation target recognition sample sets, expands the problem of few samples of photovoltaic power stations, and combines target information with geographic coding to achieve rapid positioning and rapid statistics of photovoltaic panel targets. In short, the present invention improves management efficiency and reduces management costs for the management and maintenance of photovoltaic panels in the photovoltaic power industry, and is an important reference model for the sustainable application of intelligent remote sensing monitoring technology in the photovoltaic power industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a schematic flow chart of a photovoltaic panel identification method under the constraints of photovoltaic power station scene knowledge in an embodiment of the present invention.

[0033] Figure 2 : Figure A is an example of a photovoltaic power station scene-related remote sensing image and a segmented sample original remote sensing image used in an embodiment of the present invention;

[0034] Figure 3 It is an example of a remote sensing image B of a photovoltaic power station scene-related remote sensing image and a remote sensing image of segmented samples after downsampling used in an embodiment of the present invention;

[0035] Figure 4 It is a photovoltaic power station scene-related remote sensing image and segmentation sample photovoltaic power station scene semantic segmentation initial sample set C figure used in the embodiment of the present invention;

[0036] Figure 5 is a photovoltaic power station scene semantic segmentation sample set D after the photovoltaic power station scene related remote sensing images and segmentation samples are enhanced used in the embodiment of the present invention;

[0037] Figure 6 It is an example of original remote sensing image A of the photovoltaic power station scene target result verification area obtained by automatic reasoning of the remote sensing image of the demonstration area in the embodiment of the present invention;

[0038] Figure 7It is an example of original remote sensing image B of the photovoltaic power station scene target result verification area obtained by automatic reasoning of the remote sensing image of the demonstration area in the embodiment of the present invention;

[0039] Figure 8 is a photovoltaic panel instance segmentation sample labeling related A graph recorded in an embodiment of the present invention;

[0040] Fig. 9 is a B-map related to the labeling of the photovoltaic panel instance segmentation samples recorded in the embodiment of the present invention;

[0041] Fig.10 It is a graph C of corresponding attribute table information related to the photovoltaic panel instance segmentation sample label recorded in an embodiment of the present invention;

[0042] Fig.11 It is a schematic diagram D of the calculation of the rotation angle related to the sample marking of the photovoltaic panel instance segmentation recorded in the embodiment of the present invention;

[0043] Fig.12 is a diagram of an initial photovoltaic panel training set A for generating training samples for photovoltaic panels in a demonstration area according to an embodiment of the present invention;

[0044] Fig.13 is a training sample set B after rotation and enhancement for generating training samples for photovoltaic panels in a demonstration area according to an embodiment of the present invention;

[0045] Fig.14 is the original remote sensing image A of the final reasoning result of the photovoltaic panels in the demonstration area in the embodiment of the present invention;

[0046] Fig.15 FIG. B is a reasoning result diagram of the final reasoning result of the photovoltaic panels in the demonstration area in the embodiment of the present invention without any constraints;

[0047] Fig.16 Figure C is the reasoning result of the final reasoning result of the photovoltaic panels in the demonstration area in the embodiment of the present invention under the power station scenario knowledge and training constraints. DETAILED DESCRIPTION

[0048] In order to better illustrate the technical solutions and advantages of the present invention, the steps in the present invention are described in detail through experimental demonstration and related drawings. It should be understood that the specific implementation methods described here are only used to explain the present invention and are not used to limit the present invention.

[0049] Reference Figure 1-Figure 16 The present invention provides a photovoltaic panel refined identification and segmentation method constrained by photovoltaic power station scene knowledge, comprising the following steps:

[0050] S1. Sparsely sampled photovoltaic power station scene remote sensing images I sceneIn the demonstration area, the original remote sensing images used for photovoltaic panel identification are downsampled to generate low-resolution photovoltaic power station scene images with a large geographical receptive field scale. In the demonstration experimental area within the urban area, a true color remote sensing image with a size range of 9971*10913 is selected as the experimental object, such as Figure 2 As shown in Figure A, the original geometric resolution of the image is 0.01 meters and the pixel depth is 8 bits. When making the photovoltaic power station scene training sample, it is necessary to perform downsampling processing. Select 8 times downsampling to obtain the thinned photovoltaic power station scene remote sensing image, as shown in Figure 3 As shown in Figure B, the geometric resolution of the image is 0.08 meters, the length and width are 1246*1364, and the pixel depth and geographic coordinate range remain unchanged.

[0051] S2. Create a photovoltaic power station scene segmentation sample set S scene_label Photovoltaic power station scene remote sensing image I scene As the basis, we use the semantic segmentation sample making tool to scene The photovoltaic power station scene in the polygon vector is marked to obtain the photovoltaic power station scene mask data Y scene , using the block-by-block traversal method to scene and Y scene The training sample set is cut into 1000*1000 pixels, and enhanced by random rotation geometric transformation, Gaussian blur, color enhancement, etc., to finally form the photovoltaic power station scene segmentation training sample set S scene_label .

[0052] The photovoltaic power station scene is marked and obtained, such as Figure 3 The white polygonal box drawn in Figure B is used to make a semantic segmentation sample set of photovoltaic power station scenes through block-by-block traversal method. Figure 4 As shown in Figure C, there are 12 training sample pairs in this demonstration range. After rotation and other enhancements, the sample set is as follows Figure 5 As shown in Figure D, the number of training samples increased to 144 pairs.

[0053] S3, training the photovoltaic power station scene model and knowledge reasoning. The photovoltaic power station scene semantic detection model M is trained based on the TransUNet semantic segmentation network. scene The block-by-block traversal method is used to infer the knowledge of photovoltaic power stations on a large range of remote sensing images, and the quantity space is vectorized to obtain the photovoltaic power station scene target result set R scene The photovoltaic power station scene sample set S obtained by using S2 scene_label, TransUNet is used as the photovoltaic scene recognition network, the loss function uses the cross entropy function, the weight gradient is updated by back propagation, and a single-core 4-GPU (NVIDIA TESLA A100 40GB) hardware is used for training. The batch size is set to 24, and each GPU inputs 6 samples at a time. The initial learning rate is set to 0.01. When the training iterations reach 65 epochs, convergence is completed and training is stopped. Finally, the photovoltaic power station scene recognition training model file M is obtained. scene , the format is .pth.

[0054] S4. Create a photovoltaic panel instance segmentation sample set S panel_label Based on the original remote sensing image, scene Within the spatial range agreed by the data, the photovoltaic panel is framed using the instance segmentation sample tool to obtain the photovoltaic panel mask data Y panel The remote sensing image target detection sample generation method based on traversal source targets is used to produce the initial sample set of photovoltaic panel instance segmentation with a sample size of 512*512. Since photovoltaic panels have directionality, random rotation is needed to enhance them, and finally the photovoltaic panel instance segmentation sample training set S is formed. panel_label .

[0055] In the remote sensing images of the verification area (such as Figure 6 In Figure A), it is first downsampled by 8 times, and then the photovoltaic power station scene recognition training model file M obtained by S3 is used scene Perform block reasoning and vectorization of 1000*1000 pixels and 500 moving steps, merge repeated mask results, and finally obtain the photovoltaic power station scene target result set R scene (like Figure 7 The result file is stored in vector .shp format. Figure 6 The display is used to reason about the scene knowledge constraints of photovoltaic power plants. Fig.14 What is shown is the final result used to infer the photovoltaic panel.

[0056] like Figure 6 The white polygon area in Figure B is the surface vector range of the photovoltaic power station scene obtained after automatic reasoning. Figure 7 The white marks in Figure B are the vector marking results of the photovoltaic panel.

[0057] S5. Training photovoltaic panel recognition model M panel Based on S panel_label The photovoltaic panel instance detection model M is trained using the rotating box Mask_RCNN instance segmentation network with Swin-Transformer as the feature backbone network. panel .

[0058] The same remote sensing images as those used in photovoltaic scene training were selected in the experimental area. PV , by performing sample annotation on the photovoltaic panel instance segmentation, the photovoltaic panel mask data Y is obtained panel ,like Fig. 9 The white box in Figure B, where each photovoltaic panel instance attribute records the center point coordinates, length, width, and rotation angle α, as shown in Fig.10 As shown in Figure C, its definition principle is as follows Fig.11 As shown in Figure D.

[0059] S6, Reasoning of photovoltaic panel result set for single photovoltaic power station scenario. In step S3, within the demonstration area, for the spatial scope of a single photovoltaic power station, the target result set R of the photovoltaic power station scenario is used. scene As a scene knowledge mask, it is logically processed with the remote sensing image of the demonstration area, so that the photovoltaic panel recognition model M panel We only focus on the photovoltaic panels inside the photovoltaic power station and obtain the detailed photovoltaic panel set R in a single scene through block-by-block traversal reasoning and spatial vectorization processing. panel,i , where i∈(0,N rscene ), N rscene is the number of photovoltaic electric field scenes in the demonstration area.

[0060] Since the monitoring of photovoltaic panels needs to be based on independent instances, PV and Y panel The remote sensing image target detection sample generation method that traverses the source target is adopted to ensure that each photovoltaic panel object can generate a corresponding training sample unit. The initial training sample set is generated with a sample frame size of 512*512, with a total of 1446 training sample pairs, such as Fig.12 As shown in the example of Figure A, after rotation enhancement, the photovoltaic panel instance segmentation sample training set S is finally formed. panel_label , the number is expanded to 5784 pairs, such as Fig.13 As shown in the example of Figure B, it is finally converted into the standard COCO format and the label file is stored in json format.

[0061] S7, Multi-scenario photovoltaic panel results integration. panel,i , using the geographic space location for spatial mosaicking, for the overlapping areas, using the non-maximum suppression algorithm to filter the photovoltaic panel vector results in all power station scenarios, and obtain the photovoltaic panel rotation frame recognition result RPV in the final demonstration area PanelAfter the above steps, in the process of photovoltaic panel extraction from high spatial resolution remote sensing images, non-photovoltaic electric field interference factors such as roads, buildings, and vehicles can be effectively excluded, and the complex background of the photovoltaic panel refinement extraction process can be simplified, thereby improving the refinement of photovoltaic panels located inside photovoltaic power stations.

[0062] The photovoltaic panel example analysis sample training set S obtained by S6 panel_label , the photovoltaic panel instance training is realized by using the rotating box Mask_RCNN instance segmentation network based on Swin-Transformer as the feature backbone network. The loss function is composed of the center point coordinate loss, length and width loss, rotation angle loss and the cross entropy loss of the instance segmentation surface. The weight gradient is updated by back propagation. The single-core 4 GPU (NVIDIA TESLA A100 40GB) hardware is used for training. The batch size is set to 16, and each GPU inputs 4 samples at a time. The initial learning rate is set to 0.001. When the training iterations reach 95 epochs, convergence is completed and training is stopped. Finally, the photovoltaic panel instance detection model file M is obtained. panel , the format is .pth;

[0063] Select the same remote sensing image in step S4 in the verification area, and infer the target result R in the photovoltaic power station scene. scene Based on the vectorization, the original remote sensing image is used as a knowledge mask to perform spatial logic and operation on the original remote sensing image to obtain effective photovoltaic power station scene image data, and the photovoltaic panel result inference is performed with a 512*512 sliding block and a moving step of 256. The photovoltaic panel result corresponding to the photovoltaic electric field scene of the image is obtained as follows Fig.16 Figure C, after statistics without integrating the knowledge of photovoltaic power plants, inferences were obtained to obtain 3726 photovoltaic panels, such as Fig.15 In Figure B, the extraction accuracy is only 55.3%. After further reasoning under the knowledge constraints of photovoltaic power stations, 2078 photovoltaic panels are obtained. Since there are multiple photovoltaic power station scenarios, 2061 photovoltaic panel results are obtained after NMS, and the accuracy is improved to 99.6%. The specific experimental results are shown in Table 1.

[0064] Table 1 Statistical table of inference results of photovoltaic panels in the verification area

[0065]

[0066] In some specific implementations, all remote sensing images are mainly true color bands.

[0067] In some specific implementations, in step S1, the photovoltaic power station scene remote sensing image I scene Relative to the original image I PV The downsampling ratio is usually 8 times;

[0068] That is, the following relationship is satisfied: PV =8×Reso scene , where Reso PV and Reso scene I PV and I scene The spatial resolution of remote sensing data is reduced by using bilinear interpolation or cubic convolution interpolation.

[0069] In some specific implementations, the method of step S2 is specifically as follows: the scene polygons in the photovoltaic power station scene sample set are usually the photovoltaic panel block cluster range, and the vectors are compared with the corresponding I scene The image is cropped to 1000*1000 pixels, where the scene label value is in the form of a PNG image with the inside of the vector marked as 1 and the outside marked as 0;

[0070] The random rotation geometric transformation in step S2 rotates the samples at random angles (usually multiples of 30 degrees) to retain the original sample image size. Taking the rotation of θ as an example, the scene image samples and label samples after rotation enhancement follow the following calculation formula:

[0071]

[0072] Among them, Im θ and La θ They are the rotated and enhanced matrices of the image matrix Im and label matrix La of a pair of photovoltaic scene training samples.

[0073] In some specific implementations, the specific method of step S3 is: the photovoltaic power station scene knowledge reasoning object is the upsampled remote sensing image I of the demonstration area scene ', the inference process uses a 1000*1000 pixel frame and 500 steps to traverse, and the inference vectorization result R scene Keep with I scene 'Same spatial coordinate system and geographic range.

[0074] In some specific implementations, the specific method of step S4 is: photovoltaic panel mask data Y panel It is usually stored in ESRShape format, where the attribute table contains the tag information of each photovoltaic panel, including the center point X coordinate, center point Y coordinate, rotation box width, rotation box height, and the rotation box angle α is defined as the acute angle between the center line of the rotation box and the x-axis with the horizontal axis of the center of the rotation box as the x-direction, with the counterclockwise rotation of the x-axis as positive and the clockwise rotation as negative. Then, the value range of α is [-90°, 90°];

[0075] In step S4,panel_label The sample label file is in json format. Since photovoltaic panels are directional and sensitive to directional information, random rotation enhancement in json format is required during sample enhancement. The basic principle is consistent with S2. The random change angle of the enhanced rotation angle θ is a multiple of 15 degrees. The training sample label file records the width, height, center point coordinates relative to the image sample, and rotation angle α of each photovoltaic panel target after enhancement.

[0076] In some specific implementations, the specific method of step S5 is: using the loss function in the instance segmentation in Mask_RCNN based on the rotation angle requires regressing the angle difference function of the rotation angle α, wherein the loss function of the angle loss part is shown as follows:

[0077] L α =α * Θα

[0078] Where α * is the model prediction angle, L α The difference between the angle of the true value α in the sample label.

[0079] In some specific implementations, the specific method of step S6 is: the cascade process of photovoltaic panel reasoning and photovoltaic power station scene reasoning, firstly, the coupled photovoltaic power station scene target result set R scene Mask rasterization is performed and compared with the original remote sensing image of the demonstration area. PV 'Perform logical AND operation, and then use 512*512 pixel frame and 256 steps to perform traversal reasoning, and finally obtain the semantic vectorization result R of the photovoltaic panel in the i-th photovoltaic power station scenario panel,i .

[0080] In some specific implementations, the specific method of step S7 is: when there is an area with overlapping photovoltaic electric field space in the entire demonstration area, it is necessary to use a non-extreme value suppression algorithm to eliminate repeated photovoltaic panel vector results, and the selected screening threshold is η≤0.4.

[0081] In some specific implementations, the final result RPV obtained in step S7 Panel It is saved in a vector file format. Its coordinate system information is consistent with the remote sensing image of the area and has a real geographical location. The file format is usually saved as ESRI Share File (.shp) or ESRI File GeoDatabase (.gdb).

[0082] By adopting the above technical solution disclosed in the present invention, the following beneficial effects are obtained:

[0083] The beneficial effects of the method are mainly reflected in: ① The present invention adopts a double-layer serial cascade network of a photovoltaic power station scene recognition network and a photovoltaic panel recognition network to automatically recognize photovoltaic panel information, which has a higher photovoltaic panel recognition accuracy than the conventional recognition model relying on a single network model; ② The present invention integrates the concept of multi-scale remote sensing feature information, determines the scene geographic range of a large area of ​​a photovoltaic power station with low spatial resolution, and then performs targeted high spatial resolution photovoltaic panel refinement extraction based on the power station scene, simplifies the complex background of remote sensing images in the photovoltaic panel area, not only improves the photovoltaic panel recognition accuracy, but also improves the recognition speed of photovoltaic panels in areas with smaller photovoltaic ranges; ③ The present invention adopts a multi-type enhancement mode of scene recognition and rotating target recognition sample sets, expands the problem of few samples of photovoltaic power stations, and combines target information with geographic coding to achieve rapid positioning and rapid statistics of photovoltaic panel targets. In short, the present invention improves management efficiency and reduces management costs for the management and maintenance of photovoltaic panels in the photovoltaic power industry, and is an important reference model for the sustainable application of intelligent remote sensing monitoring technology in the photovoltaic power industry.

[0084] The present invention provides a photovoltaic panel refined identification and segmentation method constrained by photovoltaic power station scene knowledge. When the photovoltaic panels are located in a relatively complex urban remote sensing background, the method will be superior to the traditional deep learning extraction algorithm. First, by downsampling the original image, the main information within the photovoltaic power station is retained, and the interference ability of non-photovoltaic panel elements is reduced. Through geographic mapping of the same area, the double-layer serial cascade reasoning of the photovoltaic power station scene recognition network and the photovoltaic panel recognition network is realized. Secondly, the multi-scale monitoring concept is integrated, and the basic range of the photovoltaic power station where the photovoltaic panel is located is locked by large-scale monitoring. Then, the detailed information of the photovoltaic panel is obtained by further using small-scale precise reasoning. Although the execution steps are slightly more complicated than the original one-step method of direct photovoltaic panel extraction, the purpose of obtaining higher recognition accuracy is achieved without losing much time.

[0085] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be considered as the scope of protection of the present invention.

Claims

1. A photovoltaic panel refined identification and segmentation method constrained by photovoltaic power station scene knowledge, characterized in that: The following steps are involved: S1. Sparsely sampled photovoltaic power station scene remote sensing images I scene ; S2. Create a photovoltaic power station scene segmentation sample set S scene_label ; S3, training and knowledge reasoning of the photovoltaic power station scenario model; S4. Create a photovoltaic panel instance segmentation sample set S panel_label ; S5. Training photovoltaic panel recognition model M panel ; S6, photovoltaic panel result set reasoning for a single photovoltaic power station scenario; S7. Integration of photovoltaic panel results in multiple scenarios.

2. The photovoltaic panel refined identification and segmentation method constrained by photovoltaic power station scene knowledge according to claim 1 is characterized in that: All remote sensing images are mainly in true color bands.

3. The photovoltaic panel refined identification and segmentation method constrained by photovoltaic power station scene knowledge according to claim 1 is characterized in that: In step S1, the photovoltaic power station scene remote sensing image I scene Relative to the original image I PV The downsampling ratio is usually 8 times; That is, the following relationship is satisfied: PV =8×Reso scene , where Reso PV and Reso scene I PV and I scene The spatial resolution of remote sensing data is reduced by using bilinear interpolation or cubic convolution interpolation.

4. The photovoltaic panel refined identification and segmentation method constrained by photovoltaic power station scene knowledge according to claim 1 is characterized in that: The method of step S2 is specifically as follows: the scene polygons in the photovoltaic power station scene sample set are usually the photovoltaic panel block cluster range, and the vectors are compared with the corresponding I scene The image is cropped to 1000*1000 pixels, where the scene label value is in the form of a PNG image with the inside of the vector marked as 1 and the outside marked as 0; The random rotation geometric transformation in step S2 rotates the samples at random angles (usually multiples of 30 degrees) to retain the original sample image size. Taking the rotation of θ as an example, the scene image samples and label samples after rotation enhancement follow the following calculation formula: Among them, Im θ and La θ They are the rotated and enhanced matrices of the image matrix Im and label matrix La of a pair of photovoltaic scene training samples.

5. The photovoltaic panel refined identification and segmentation method constrained by photovoltaic power station scene knowledge according to claim 1 is characterized in that: The specific method of step S3 is: the photovoltaic power station scene knowledge reasoning object is the upsampled remote sensing image I of the demonstration area scene ', the inference process uses a 1000*1000 pixel frame and 500 steps to traverse, and the inference vectorization result R scene Keep with I scene 'Same spatial coordinate system and geographic range.

6. The photovoltaic panel refined identification and segmentation method constrained by photovoltaic power station scene knowledge according to claim 1 is characterized in that: The specific method of step S4 is: photovoltaic panel mask data Y panel It is usually stored in ESRI Shape format, in which the attribute table contains the tag information of each photovoltaic panel, including the center point X coordinate, center point Y coordinate, rotation box width, rotation box height, and the rotation box angle α is defined as the acute angle between the center line of the rotation box and the x-axis with the horizontal axis of the center of the rotation box as the x-direction, with the counterclockwise rotation of the x-axis as positive and the clockwise rotation as negative. Then, the value range of α is [-90°, 90°]; In step S4, panel_label The sample label file is in json format. Since photovoltaic panels are directional and sensitive to directional information, random rotation enhancement in json format is required during sample enhancement. The basic principle is consistent with S2. The random change angle of the enhanced rotation angle θ is a multiple of 15 degrees. The training sample label file records the width, height, center point coordinates relative to the image sample, and rotation angle α of each photovoltaic panel target after enhancement.

7. The photovoltaic panel refined identification and segmentation method constrained by photovoltaic power station scene knowledge according to claim 1 is characterized in that: The specific method of step S5 is: using the loss function in the instance segmentation in Mask_RCNN based on the rotation angle requires regressing the angle difference function of the rotation angle α, wherein the loss function of the angle loss part is shown as follows: L α =a * Will Where α * is the model prediction angle, L α The difference between the angle of the true value α in the sample label.

8. The photovoltaic panel refined identification and segmentation method constrained by photovoltaic power station scene knowledge according to claim 1 is characterized in that: The specific method of step S6 is: the cascade process of photovoltaic panel reasoning and photovoltaic power station scene reasoning, firstly, the target result set R coupled to the photovoltaic power station scene is scene Mask rasterization is performed and compared with the original remote sensing image of the demonstration area. PV 'Perform logical AND operation, and then use 512*512 pixel frame and 256 steps to perform traversal reasoning, and finally obtain the semantic vectorization result R of the photovoltaic panel in the i-th photovoltaic power station scenario panel,i .

9. The photovoltaic panel refined identification and segmentation method constrained by photovoltaic power station scene knowledge according to claim 1 is characterized in that: The specific method of step S7 is: when there is an area with overlapping photovoltaic electric field space in the entire demonstration area, it is necessary to use a non-extreme value suppression algorithm to eliminate repeated photovoltaic panel vector results, and the selected screening threshold is η≤0.

4.

10. The photovoltaic panel refined identification and segmentation method constrained by photovoltaic power station scene knowledge according to claim 9 is characterized in that: The final result RPV obtained in step S7 Panel It is saved in a vector file format, and its coordinate system information is consistent with the remote sensing image of the area. It has a real geographical location. The file format is usually saved in ESRI ShapeFile (.shp) or ESRIFile GeoDatabase (.gdb).

Images