A method and system for intelligent assessment of photovoltaic potential

Abstract

The present application relates to photovoltaic potential evaluation technical field, especially photovoltaic potential intelligent evaluation method and system, the method is based on original image data and carries out multi-scale feature extraction, identifies the area meeting photovoltaic installation condition and generates boundary vector polygon;Combined with obstacle information, the effective installation area is calculated;Identify the roof type and the inclination, and determine the orientation according to the geographic coordinates;Combined with solar radiation time series data, the annual power generation of unit area is predicted;And the effective photovoltaic installation area and the annual power generation of unit area are mapped to GIS to generate photovoltaic potential heat map;The present application can automatically identify the roof and accurately consider the influence of shielding, improve the reliability of power generation prediction accuracy, realize the whole process support from roof identification to investment evaluation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of photovoltaic potential estimation, and particularly to an intelligent photovoltaic potential estimation method and system. BACKGROUND

[0002] Photovoltaic potential estimation refers to estimating the photovoltaic power generation potential of a certain region or building surface by comprehensively considering factors such as geographical location, building characteristics, and climate data. In this process, the roof characteristics of the building and the solar radiation time series data are key factors. The photovoltaic installation area refers to the area suitable for installing photovoltaic panels, and the determination basis includes the effective area of the roof, unobstructed sunlight, and the influence of geographical location.

[0003] The judgment of the photovoltaic installation area mainly relies on manual or simple rule inference, and it is difficult to automatically identify specific areas that meet the installation conditions, and the accuracy is low. Secondly, the existing technology mostly ignores the obstacles that may exist around the building, such as high-rise buildings or other obstructions, which may cause the actual evaluated photovoltaic installation area to be lower than the theoretical value.

[0004] The information disclosed in this BACKGROUND section is only intended to enhance the understanding of the general background of the present application, and should not be construed as recognition or any form of suggestion that this information constitutes prior art that is known to those skilled in the art. SUMMARY

[0005] The present application provides an intelligent photovoltaic potential estimation method and system, thereby effectively solving the problems in the background art.

[0006] In order to achieve the above purpose, the technical solution adopted by the present application is as follows: an intelligent photovoltaic potential estimation method, comprising the following steps:

[0007] Based on the original image data, the areas that meet the photovoltaic installation conditions in the building surface are identified through multi-scale feature extraction and fusion, and the boundary vector polygon of each area is outputted;

[0008] According to the boundary vector polygon, the obstacle occupied area is deducted, and the effective photovoltaic installation area is calculated;

[0009] The roof type and inclination angle are identified, and the orientation is determined in combination with the geographical coordinates;

[0010] In combination with the solar radiation time series data, the roof orientation and inclination angle, the annual power generation per unit area is predicted;

[0011] The effective photovoltaic installation area and the annual power generation per unit area are associated with the geographic information system map to form a photovoltaic potential heat map.

[0012] Further, based on the original image data, by multi-scale feature extraction and fusion, the region meeting the photovoltaic installation condition in the building surface is identified, and the step of outputting the boundary vector polygon of each region includes:

[0013] Multi-scale feature map extraction is performed on the input high-resolution remote sensing image;

[0014] Based on the multi-scale feature map, the building bounding box is predicted;

[0015] Based on the building bounding box, the building roof category is classified;

[0016] From the multi-scale feature map and the building bounding box, sparse contour perception features and dense contour perception features are extracted, and the sparse contour perception features and the dense contour perception features are fused to obtain enhanced contour perception features;

[0017] Based on the enhanced contour perception features, the geometric primitives of the building contour are segmented, the connection order of the geometric primitives is determined, and the topological structure is optimized to generate a regular vector polygon of the building contour;

[0018] The roof category of the building and the regular vector polygon result of the building contour are output.

[0019] Further, the multi-scale feature map extraction includes:

[0020] The pre-trained backbone network is used to extract multi-scale feature maps;

[0021] A plurality of convolution kernels with different rotation angles are generated by a compass convolution module, multi-directional convolution operations are performed on the multi-scale feature maps, and multi-directional features are spliced and fused to obtain multi-directional contour features.

[0022] Further, the building bounding box prediction includes:

[0023] The multi-scale feature map is received, a single-stage detection structure based on full convolution is used to predict the coordinates and confidence of the corresponding bounding box at each spatial position of the multi-scale feature map, and the position information and confidence result of the building bounding box are output.

[0024] Further, based on the building bounding box, the building roof category is classified, and the step includes:

[0025] The image features and geometric features in the building bounding box region are extracted, wherein the geometric features include roof texture geometric features constructed by learned normals and calculated normals;

[0026] The geometric features are fused with the image features, and the fusion is based on a spatial gating mechanism to adjust the fusion proportion of the image features and the geometric features, and a geometric confidence evaluation module is combined for weighted processing;

[0027] A residual connection and a spatial-channel dual attention mechanism are applied to the fused features to enhance the expression ability of the roof region features;

[0028] The enhanced roof region features are input into a global average pooling layer, a fully connected layer and a Softmax activation function to output a probability distribution result of the roof type.

[0029] Further, sparse contour perception features and dense contour perception features are extracted from the multi-scale feature maps and the building bounding box, and the sparse contour perception features and the dense contour perception features are fused to obtain enhanced contour perception features, and the steps include:

[0030] Reference sampling points are generated in the building bounding box region, and offsets of the sampling points relative to the building contour edge are predicted by an offset prediction sub-network to extract sparse contour perception features;

[0031] The sparse contour perception features are input into a multi-head self-attention module as keys and values, and the multi-scale feature maps are input as queries to generate a building boundary segmentation map and output dense contour perception features;

[0032] Based on the building boundary segmentation map, a contour buffer mask region is constructed, and the sparse contour perception features and the dense contour perception features are fused according to the mask region to generate enhanced contour perception features.

[0033] Further, based on the enhanced contour perception features, geometric primitives of the building contour are segmented, the connection order of the geometric primitives is determined, and a geometric topological constraint is introduced to generate a topologically optimized building contour regular vector polygon, and the steps include:

[0034] The enhanced contour perception features are input into a primitive segmenter, and a learnable query vector is used to predict the position coordinates and class confidence of the geometric primitives of the building contour through cross-attention and self-attention mechanisms;

[0035] The geometric primitives are input into a sequential decoder, and a multi-layer self-attention structure is used to model the topological connection relationship between the geometric primitives to predict the connection order of the geometric primitives in the contour polygon;

[0036] The position coordinates and the connection order of the geometric primitives are input into a geometric constraint decoder, and a self-attention mechanism is used to model the geometric adjacency relationship and the angle constraint relationship to generate an optimized connection structure that satisfies geometric consistency.

[0037] According to the optimized connection relationship of the geometric primitives, a building contour regular vector polygon with reasonable topological structure is constructed.

[0038] Further, according to the optimized connection relationship of the geometric primitives, a building contour regular vector polygon with reasonable topological structure is constructed, and the step comprises:

[0039] The predicted position coordinates and connection order of the geometric primitives are taken as inputs, and the geometric features of each geometric primitive are encoded into a vector form;

[0040] The vector is input into at least two layers of self-attention structure to learn the relative distance constraint relationship and directional constraint relationship between the geometric primitives, including parallelism and perpendicularity;

[0041] Based on the probability output of the connection order and the geometric relationship modeled by self-attention, the connection structure between the primitives is optimized to generate a connection path that meets the geometric consistency and topological rationality;

[0042] According to the connection path, the corner points of each geometric primitive are sequentially connected to construct a closed, connected and contour-regular building contour vector polygon.

[0043] Further, combined with solar radiation time series data, roof orientation and inclination, the unit area annual power generation is predicted, including:

[0044] According to the roof orientation and inclination, the solar radiation time series data is corrected by three-dimensional projection, and the actual radiation received by the roof plane is calculated;

[0045] Based on the obstacle position and height marked in the boundary vector polygon, combined with the solar orbit, dynamic shadow is simulated, and the shielding attenuation coefficient is calculated;

[0046] According to the periodic time power generation and time dimension accumulation, the unit area annual power generation is calculated, and the periodic time power generation is calculated as follows:

[0047] Periodic time power generation = periodic radiation × attenuation coefficient × photovoltaic component efficiency.

[0048] The present application also includes an intelligent evaluation system for photovoltaic potential, which uses the method as described above, comprising:

[0049] An image recognition and vectorization module is used to identify the areas meeting the photovoltaic installation conditions in the building surface based on the original image data through multi-scale feature extraction and fusion, and output the boundary vector polygon of each area;

[0050] An effective area calculation module is configured to calculate an effective photovoltaic installation area by deducting an obstacle occupied area from the boundary vector polygon;

[0051] A roof attribute identification module is configured to identify a roof type and an inclination angle, and determine an orientation in combination with geographic coordinates;

[0052] A power generation capacity prediction module is configured to predict an annual power generation capacity per unit area in combination with solar radiation time series data, a roof orientation and an inclination angle;

[0053] A visualization output module is configured to associate the effective photovoltaic installation area and the annual power generation capacity per unit area to a geographic information system map to form a photovoltaic potential heat map.

[0054] Further, the image recognition and vectorization module is configured to identify areas meeting photovoltaic installation conditions on a building surface based on original image data through multi-scale feature extraction and fusion, and output boundary vector polygons of the areas, and further includes:

[0055] A feature extractor is configured to extract multi-scale feature maps from an input high-resolution remote sensing image;

[0056] A detection head is configured to predict a building bounding box based on the multi-scale feature maps;

[0057] A roof classifier is configured to classify a building roof category based on the building bounding box;

[0058] A contour feature enhancer is configured to extract sparse contour perception features and dense contour perception features from the multi-scale feature maps and the building bounding box, and fuse the sparse contour perception features and the dense contour perception features to obtain enhanced contour perception features;

[0059] A building contour regular polygon extractor is configured to split a geometric primitive of a building contour based on the enhanced contour perception features, determine a connection order of the geometric primitive, and introduce a geometric topological constraint to generate a topologically optimized building contour regular vector polygon;

[0060] An output module is configured to output a roof category of a building and a regular vector polygon result of a building contour.

[0061] Further, the contour feature enhancer is configured to extract sparse contour perception features and dense contour perception features from the multi-scale feature maps and the building bounding box, and fuse the sparse contour perception features and the dense contour perception features to obtain enhanced contour perception features, and includes:

[0062] a sparse feature extractor configured to generate reference sampling points in the building bounding box region and predict offsets of each of the sampling points relative to the edge of the building contour by an offset predictor network to extract sparse contour-aware features;

[0063] a dense feature extractor configured to input the sparse contour-aware features as keys and values and the multi-scale feature maps as queries into a multi-head self-attention module to generate a building boundary segmentation map and output dense contour-aware features;

[0064] a feature fusion module configured to construct a contour buffer mask region based on the building boundary segmentation map and fuse the sparse contour-aware features and the dense contour-aware features according to the mask region to generate enhanced contour-aware features.

[0065] The present application has the following advantages:

[0066] By combining image recognition, geographic information processing and solar radiation simulation, the present application realizes full-process automatic evaluation of photovoltaic installation potential. Compared with the existing methods relying on manual or simple rule judgment, the present application can automatically identify suitable installation areas of building roofs under large-scale data and accurately calculate effective photovoltaic installation area in combination with obstacle factors, thereby avoiding evaluation deviation caused by neglecting shading. Meanwhile, the present application fully considers the dynamic changes of roof orientation, inclination and solar radiation when predicting power generation, thereby improving the accuracy and reliability of power generation prediction. Therefore, the present application not only improves the precision and efficiency of photovoltaic potential evaluation, but also provides a scientific and reliable basis for planning and layout of photovoltaic industry, and has significant practical value and popularization prospect. BRIEF DESCRIPTION OF DRAWINGS

[0067] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments described in the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0068] Figure 1 Flow chart of the intelligent evaluation method for photovoltaic potential;

[0069] Figure 2 Structure diagram of the image recognition and vectorization module in the intelligent evaluation system for photovoltaic potential;

[0070] Figure 3 Structure diagram of the contour feature enhancer (including sparse feature extractor, dense feature extractor and feature fusion module);

[0071] Figure 4 Fig. 1 is a schematic diagram of the working principle of a sparse feature extractor (uniform sampling points, adjusted sampling points, and bilinear interpolation). DETAILED DESCRIPTION

[0072] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments of the present application.

[0073] Embodiment 1

[0074] As shown in the following: Figure 1 A method for intelligent evaluation of photovoltaic potential, comprising the following steps:

[0075] S10: Based on the original image data, through multi-scale feature extraction and fusion, the areas on the building surface that meet the photovoltaic installation conditions are identified, and the boundary vector polygons of each area are outputted;

[0076] S20: According to the boundary vector polygon, the obstacle occupied area is deducted, and the effective photovoltaic installation area is calculated; by obtaining the obstacle information around the building, the spatial analysis method is used to deduct the obstacle area; one optional specific implementation mode is that first, the positions and heights of obstacles such as high-rise buildings and trees around the building are obtained through laser radar scanning, laser radar data or data in a geographic information system database; then, the obstacle area is deducted from the boundary vector polygon of the building roof by using the polygon Boolean operation technology (such as difference set operation), and finally the effective area meeting the photovoltaic installation condition is calculated;

[0077] S30: Identify the roof type and inclination, and determine the orientation combined with the geographic coordinates; through image analysis combined with laser radar data, the type (such as flat roof, slope roof, curved roof, etc.) and inclination of the building roof are automatically identified; in specific implementation, first, a deep learning model (such as convolutional neural network CNN or U-Net) is used to segment the roof image to identify the shape of the roof; then, an image geometric analysis method is used to calculate the inclination of the roof, and combined with the geographic coordinates (such as latitude and longitude information) of the building, the orientation of the roof is determined; the geographic coordinates and inclination information can help to calculate the solar radiation intensity of the roof in different time periods, so as to provide accurate orientation data for power generation prediction;

[0078] S40: Combined with the solar radiation time series data, the roof orientation and inclination, the annual power generation per unit area is predicted;

[0079] S50: associate the effective photovoltaic installation area and the annual power generation per unit area to the geographic information system map to form a photovoltaic potential heat map; in this step, the effective photovoltaic installation area and the annual power generation per unit area calculated are associated to the geographic information system map for visual display; the specific implementation is that first, the photovoltaic installation area and the annual power generation per unit area of each building are taken as data points corresponding to the geographic position (latitude and longitude) thereof; then, the geographic information system platform is used to convert these data into a heat map to visually display the photovoltaic potential of different regions through color changes; this heat map can help users quickly identify which regions have high photovoltaic power generation potential, thereby providing data support for photovoltaic project decision-making;

[0080] Firstly, based on the original image data, a deep learning model is used for multi-scale feature extraction and fusion to identify the building roof area meeting the photovoltaic installation condition and output the boundary vector polygon of each area; then, combined with the surrounding obstacle information obtained by the laser radar data and other sensors, a Boolean operation is performed through a spatial analysis tool to deduct the obstacle occupied area and calculate the effective photovoltaic installation area; for the identification of roof type and inclination, image analysis and laser radar data are combined to further determine the orientation and geographic position of the roof, thereby providing data support for power generation prediction; next, combined with the solar radiation time series data, roof orientation and inclination, the solar radiation model is used to predict the annual power generation per unit area; finally, the effective photovoltaic installation area and the annual power generation per unit area are associated to the geographic information system to generate a photovoltaic potential heat map, thereby providing visual photovoltaic installation potential information for decision makers.

[0081] Through the organic combination of image recognition, geographic information processing and solar radiation simulation, the full-process automatic evaluation of photovoltaic installation potential is realized; compared with the existing method relying on manual or simple rule judgment, this method can automatically identify the suitable installation area of the building roof under a large range of data, and accurately calculate the effective photovoltaic installation area combined with the obstacle factor, thereby avoiding the evaluation deviation caused by ignoring the obstruction; at the same time, this method fully considers the dynamic changes of roof orientation, inclination and solar radiation when predicting the power generation, thereby improving the accuracy and reliability of power generation prediction; finally, through the visual display of the geographic information system platform, the user can intuitively obtain the photovoltaic potential heat map to realize the support from roof identification; therefore, the present application not only improves the precision and efficiency of photovoltaic potential evaluation, but also provides a scientific and reliable basis for the planning and layout of the photovoltaic industry, which has significant practical value and popularization prospect.

[0082] As a preferred embodiment of the above embodiment, in step S10, the step of identifying the area meeting the photovoltaic installation condition in the building surface based on the original image data through multi-scale feature extraction and fusion and outputting the boundary vector polygon of each area comprises:

[0083] S11: performing multi-scale feature map extraction on the input high-resolution remote sensing image;

[0084] S12: predicting a building bounding box based on the multi-scale feature map;

[0085] S13: classifying a building roof type based on the building bounding box;

[0086] S14: extracting sparse contour perception features and dense contour perception features from the multi-scale feature map and the building bounding box, and fusing the sparse contour perception features and the dense contour perception features to obtain enhanced contour perception features;

[0087] S15: segmenting a geometric primitive of a building contour based on the enhanced contour perception features, determining a connection order of the geometric primitive, and introducing a geometric topology constraint to generate a topologically optimized building contour regular vector polygon;

[0088] S16: outputting a roof type of the building and a regular vector polygon result of the building contour.

[0089] First, multi-scale feature map extraction is performed on the input high-resolution remote sensing image to capture the geometric and texture features of the building at different scales. Then, a target detection algorithm is used to predict the building bounding box based on the multi-scale feature map, and strategies such as non-maximum suppression are used to ensure the accuracy and uniqueness of the detection results. After obtaining the building bounding box, the roof area is cropped and input into a classification model to identify the roof type, such as flat roof, pitched roof, or curved roof, providing a basis for subsequent modeling. Further, sparse contour perception features and dense contour perception features are extracted from the multi-scale feature map and the bounding box, respectively, and enhanced contour perception features are obtained through a fusion strategy to balance the overall structure and local details of the roof boundary. On this basis, the building contour is segmented into geometric primitives, and the connection relationship is optimized with geometric topology constraints to avoid self-intersection or gap problems, ensuring that the generated roof boundary is regularized and regularized. The final output includes roof class labels and regular vector polygon data, which can be directly applied to geographic information system platforms in standard formats such as GeoJSON or Shapefile, enabling automatic and high-precision subsequent photovoltaic potential analysis.

[0090] Through key steps such as contour feature enhancement, geometric primitive segmentation, and topology optimization, the method overcomes the problems of inaccurate feature extraction, discontinuous contour expression, chaotic topology structure, and complex post-processing in existing methods. The method can achieve high-precision classification of building roof types and end-to-end output of regular vector contours, with the advantages of high precision, reasonable structure, high automation, and other beneficial effects, especially suitable for high-precision scenarios such as photovoltaic resource development potential assessment.

[0091] In the embodiment, in step S11, the multi-scale feature map extraction includes:

[0092] S111: extracting a multi-scale feature map by using a pre-trained backbone network;

[0093] S112: generating a plurality of groups of convolution kernels with different rotation angles by using a compass convolution module, performing multi-directional convolution operation on the multi-scale feature map, and fusing multi-directional features to obtain multi-directional contour features.

[0094] In a specific implementation, the feature extractor feature encoding first extracts a multi-scale feature map from an input high-resolution remote sensing image by using a pre-trained ResNet-50 backbone network; then the compass convolution module starts to work, and the specific operation is as follows: first, 8 groups of 3x3 convolution kernels with a rotation angle of 45 degrees (0 degrees, 45 degrees, 90 degrees, 135 degrees, 180 degrees, 225 degrees, 270 degrees, and 315 degrees) are generated; then, multi-directional convolution is performed on the input feature map in parallel to extract rotation features; finally, the multi-directional features are fused by channel splicing, the channel number is compressed by using 1x1 convolution, and a multi-scale feature map with enhanced multi-directional contour features of the building is output.

[0095] The input high-resolution remote sensing image is first subjected to multi-scale feature extraction by using a pre-trained backbone network, so as to capture low-level edge details and high-level semantic information of a roof at the same time, thereby obtaining a multi-scale feature map with local and global expressions; on this basis, a compass convolution module is introduced to generate a plurality of groups of convolution kernels with different rotation angles, perform multi-directional convolution operation on the multi-scale feature map, and enhance the perception ability of the model to multi-directional edges and contours; then, the features extracted in different directions are integrated by using channel splicing or adaptive weighting fusion, thereby forming multi-directional contour features, so that in the case that there are inclined lines, curves, or irregular boundaries on the roof of a complex building, complete and accurate contour extraction can still be realized.

[0096] In step S12, the building boundary box is predicted, including:

[0097] The multi-scale feature map is received, a single-stage detection structure based on full convolution is used, the coordinates and confidence of the corresponding boundary box are predicted at each spatial position of the multi-scale feature map, and the position information and confidence result of the building boundary box are output.

[0098] First, the detection head receives the multi-scale feature map output by the feature extractor; then, the multi-scale feature map is subjected to convolution processing to predict the probability distribution of the center point of the boundary box; finally, the four-dimensional coordinate offset of the boundary box is regressed based on the center point position, and the coordinates and confidence of the building boundary box are predicted.

[0099] Specifically, after obtaining the multi-scale feature map, the coordinates of the bounding box and the corresponding confidence score are directly predicted by using the full convolutional network at each spatial position of the feature map, so as to obtain a set of candidate bounding boxes. The predicted coordinates can be in the form of center point coordinates plus width and height, or in the form of absolute position representation of the upper left corner and the lower right corner. In order to improve the accuracy and robustness of detection, a non-maximum suppression strategy can be introduced in the output result to remove redundant boxes with too high overlap, and only the bounding boxes with high confidence are retained as the final result. The single-stage detection structure can use mature frameworks such as SSD, YOLO or RetinaNet, and combine the anchor mechanism in the output layer to adapt to buildings of different sizes and proportions. Not only does it avoid the complex process of candidate region generation and classification regression separation in two-stage detection, but it also greatly improves the detection speed and computational efficiency, while ensuring the accuracy of building bounding box prediction, providing reliable data input for subsequent roof classification and vectorization processing.

[0100] As a preferred embodiment of the above, in step S13, the building roof category is classified based on the building bounding box, and the step includes:

[0101] S131: Extracting image features and geometric features in the building bounding box region, wherein the geometric features include roof texture geometric features constructed by learning normal and calculated normal;

[0102] S132: Fusing the geometric features and the image features, wherein the fusion is based on a spatial gating mechanism to adjust the fusion proportion of the image features and the geometric features, and combined with a geometric confidence evaluation module for weighted processing;

[0103] S133: Applying residual connection and spatial-channel dual attention mechanism to the fused features to enhance the expression ability of the roof region features;

[0104] S134: Inputting the enhanced roof region features into a global average pooling layer, a fully connected layer and a Softmax activation function to output a probability distribution result of the roof type.

[0105] For the classification of building roof types, first, image features and geometric features are extracted within the building bounding box area, wherein the geometric features are composed of learned normals and computed normals to comprehensively reflect the spatial texture and geometric morphology of the roof. Then, the fusion ratio of the image features and the geometric features is adaptively adjusted through a spatial gating mechanism, and the feature reliability is weighted by a geometric confidence evaluation module to improve the accuracy of the fusion result. On this basis, the fused features are input into a residual connection structure, and a spatial and channel dual attention mechanism is introduced to strengthen the important features related to roof type discrimination and suppress irrelevant information. Finally, the enhanced features are output through a global average pooling layer, a fully connected layer and a Softmax activation function to obtain the probability distribution of the roof type, and thus the identification result of the roof type is obtained. Not only can it maintain high classification accuracy in complex texture or shadow interference conditions, but also significantly enhances the expression ability of the roof area features, providing a reliable data foundation for subsequent photovoltaic potential assessment.

[0106] In the embodiment, in step S14, sparse contour perception features and dense contour perception features are extracted from the multi-scale feature map and the building bounding box, and the sparse contour perception features and the dense contour perception features are fused to obtain enhanced contour perception features, and the step includes:

[0107] S141: generating reference sampling points in the building bounding box area, and predicting the offset of each sampling point relative to the building contour edge through an offset prediction subnetwork to extract sparse contour perception features;

[0108] The sparse feature extractor aggregates the feature sampling points to the real building contour edge, and the specific operation is as follows: first, initialize a uniformly distributed reference point grid on the feature map within the given building bounding box; then, input the linearly mapped feature map into a lightweight offset prediction subnetwork (including 5x5 depth convolution, GELU activation, 1x1 convolution, Tanh activation) to predict the offset of each reference point relative to the building contour edge (offset ); second, use a preset scaling factor (s) to scale the predicted offset, and add the scaled offset to the initial reference point coordinates to obtain the position of the final sampling point. Finally, the sampling points after scaling and offsetting tend to gather near the building contour, and the features at these sampling points are extracted by bilinear interpolation to obtain sparse contour perception building features (f ) and their position information;

[0109] S142: inputting the sparse contour perception features as keys (Key) and values (Value) and the multi-scale feature map as queries (Query) into a multi-head self-attention module to generate a building boundary segmentation map and output dense contour perception features;

[0110] Dense feature extraction can generate accurate building boundary segmentation maps, providing rich contour contextual semantic information and additional boundary supervision signals for the model. Specifically, the obtained sparse contour-aware features are processed... As keys and values, the multi-scale feature map of the building is used. As a query, the Key, Value, and Query are then fed into a multi-head self-attention module for information interaction, outputting an enhanced feature map. Then, it is fused with the multi-scale feature maps extracted from the backbone network, and after passing through convolutional layers and a sigmoid activation function, a building boundary segmentation probability map is generated. Finally, the segmentation probability map will be... Resize to match the original multi-scale feature map size to obtain dense contour-aware building features. );

[0111] S143: Construct a contour buffer mask region based on the building boundary segmentation map, and fuse sparse contour perception features and dense contour perception features according to the mask region to generate enhanced contour perception features.

[0112] Feature fusion, by fusing sparse and dense features, generates more accurate contour-aware features; specific operation: first, based on the generated boundary segmentation map... A building outline edge buffer is constructed, centered on the predicted boundary line and extended inwards and outwards by 8 pixels each, forming a strip region with a certain width containing the most relevant outline information; then, in the dense outline-aware feature map... First, a buffer mask m is applied: feature values ​​within buffer m are retained, while feature values ​​outside the buffer are set to zero to suppress invalid features far from the contour, resulting in a dense feature map after masking. Second, based on the position information of the obtained sparse feature points, features at corresponding positions are extracted from the dense feature map after masking. Finally, these extracted features are compared with the original sparse contour-aware features. The features are then stitched together to obtain the final enhanced contour-aware features. The output is then fed to the subsequent primitive splitter.

[0113] The feature fusion performs the following operations: expand inward and outward based on the building boundary segmentation map to form an edge buffer region; retain features within the buffer region and set features outside the region to zero; extract reference grid point features from the retained features based on the location information of the sparse contour perception features; concatenate the reference grid point features with the sparse contour perception features to generate enhanced features; the edge buffer region expands inward and outward by 8 pixels from the building boundary as the center to form a 16-pixel wide feature retention region.

[0114] wherein the model of sparse contour-aware feature comprises:

[0115] ;

[0116] ;

[0117] ;

[0118] wherein, denotes the sparse contour-aware feature, denotes the input multi-scale feature map, denotes a linear transformation matrix, denotes mapping the feature map to a query vector, denotes an offset predictor network, denotes a preset scaling factor, denotes the offset of the sampling point relative to the edge of the building contour, denotes the initial uniform grid points of the sampling points, denotes a bilinear interpolation function, which samples the feature map at the deformed point to obtain the sparse contour-aware feature .

[0119] wherein the model of dense contour-aware feature comprises:

[0120] ;

[0121] ;

[0122] ;

[0123] wherein, denotes the dense contour-aware feature, denotes the sparse contour-aware feature, and denote linear transformation matrices, which map to keys ( ) and values ( ), denotes mapping the feature map to a query vector, is a multi-head self-attention mechanism, which realizes information interaction between and outputs an enhanced feature map , denotes that the multi-head self-attention mechanism outputs an enhanced feature map after interacting the query vector with the keys and values, denotes a boundary segmenter, which inputs Obtaining a building boundary segmentation probability map .

[0124] In step S15, based on the enhanced contour perception feature, the geometric primitives of the building contour are segmented, the connection order of the geometric primitives is determined, and the geometric topological constraint is introduced to generate the topological structure optimized building contour regular vector polygon, and the steps include:

[0125] S151: input the enhanced contour perception feature into the primitive segmenter, use the learnable query vector, and predict the position coordinates and category confidence of the geometric primitives of the building contour through the cross-attention mechanism and the self-attention mechanism;

[0126] S152: input the geometric primitives into the order decoder, model the topological connection relationship between the geometric primitives based on the multi-layer self-attention structure, and predict the connection order of each geometric primitive in the contour polygon;

[0127] S153: input the position coordinates and connection order of the geometric primitives into the geometric constraint decoder, model the geometric adjacency relationship and angle constraint relationship through the self-attention mechanism, and generate the optimized connection structure that satisfies the geometric consistency;

[0128] S154: according to the optimized connection relationship of the geometric primitives, construct a topological structure reasonable building contour regular vector polygon.

[0129] The primitive segmenter first inputs the contour enhancement feature output in step S40 Input the primitive segmenter, use the cross-attention and self-attention mechanism to process the input feature and a set of learnable queries (Query), update the position embedding of the query through the feedforward network (FFN) to predict the offset vector; Then, use the query and the updated position embedding to predict the accurate position coordinates and category confidence of the geometric primitives through the multi-layer perception (MLP).

[0130] The order decoder is responsible for predicting the topological connection order between the geometric primitives (such as vertices, lines, and angles) output by the primitive segmenter in step S51, and the specific operation is as follows: The order decoder explicitly models and strengthens the adjacency relationship between the geometric primitives through multi-layer self-attention, and the self-attention layer interacts with all primitive queries to learn the relative importance between the primitives and focus on the adjacent relationship of the primitives in the contour sequence; Finally, the decoder predicts the probability distribution of the primitives in the order category through the fully connected layer, and outputs the position and sequence relationship of the primitives in the building contour polygon according to the probability.

[0131] As a preferred embodiment of the above, in step S154, according to the optimized connection relationship of the geometric primitives, a topological structure reasonable building contour regular vector polygon is constructed, and the steps include:

[0132] S1541: encode the geometric features of each geometric primitive into vector form as input with the predicted position coordinates of the geometric primitives and the connection order;

[0133] S1542: input the vector into at least two layers of self-attention structure to learn the relative distance constraint relationship and directional constraint relationship between geometric primitives, including parallelism and perpendicularity;

[0134] S1543: based on the probability output of the connection order and the geometric relationship modeled by the self-attention, optimize the connection structure between primitives to generate a connection path that meets the geometric consistency and topological rationality;

[0135] S1544: connect the corner points of each geometric primitive in turn according to the connection path to construct a closed, connected and regular building contour vector polygon.

[0136] The geometric constraint decoder is responsible for optimizing the topological connection order based on geometric features. The specific operation is as follows: first, input the output of the primitive segmenter and the prediction result of the order decoder into the vector form of the geometric features of each primitive; then, the feature vector learns the relative distance constraint relationship and parallel and perpendicular position constraint relationship between primitives through two layers of self-attention; finally, the model connects the corner points according to the connection order probability and the topologically optimized primitive connection relationship, and directly outputs the topologically optimized vector polygon.

[0137] In this embodiment, in the training process, the neural network structure for implementing the method is supervised learning, and the supervised learning is based on a weighted loss function model, which includes:

[0138] ;

[0139] In the formula, respectively represent the weight coefficients for balancing each loss; represents the detection loss, represents the primitive segmentation loss, represents the order regression loss, represents the boundary segmentation loss; represents the geometric topological constraint loss.

[0140] In this embodiment, in step S40, the unit area annual power generation is predicted in combination with the solar radiation time series data, the roof orientation and the inclination, including:

[0141] According to the roof orientation and the inclination, the solar radiation time series data is corrected by three-dimensional projection to calculate the actual radiation received by the roof plane;

[0142] Based on the obstacle position and height marked in the boundary vector polygon, and combined with the sun trajectory simulation dynamic shadow, the shielding attenuation coefficient is calculated.

[0143] The annual power generation per unit area is calculated according to the period power generation and time dimension accumulation, and the period power generation is calculated as follows:

[0144] Period power generation = Period radiation × Attenuation coefficient × Photovoltaic module efficiency.

[0145] The application also includes an intelligent evaluation system for photovoltaic potential, which uses the method as described above, including:

[0146] An image recognition and vectorization module is used to identify the area meeting the photovoltaic installation condition on the building surface based on the original image data through multi-scale feature extraction and fusion, and output the boundary vector polygon of each area.

[0147] An effective area calculation module is used to calculate the effective photovoltaic installation area according to the boundary vector polygon and deduct the obstacle occupied area.

[0148] A roof attribute recognition module is used to identify the roof type and inclination, and determine the orientation combined with the geographic coordinates.

[0149] A power generation prediction module is used to predict the annual power generation per unit area combined with the solar radiation time series data, roof orientation and inclination.

[0150] A visual output module is used to associate the effective photovoltaic installation area and the annual power generation per unit area to the geographic information system map to form a photovoltaic potential heat map.

[0151] As a preferred embodiment of the above, as shown in the image recognition and vectorization module, based on the original image data, through multi-scale feature extraction and fusion, the area meeting the photovoltaic installation condition on the building surface is identified, and the boundary vector polygon of each area is output, and further includes: Figures 2 to 4

[0152] A feature extractor is used to extract multi-scale feature maps from the input high-resolution remote sensing image.

[0153] A detection head is used to predict the building bounding box based on the multi-scale feature map.

[0154] A roof classifier is used to classify the building roof category based on the building bounding box.

[0155] A contour feature enhancer is used to extract sparse contour perception features and dense contour perception features from the multi-scale feature map and the building bounding box, and fuse the sparse contour perception features and the dense contour perception features to obtain enhanced contour perception features. ​

[0156] The building contour regular polygon extractor is used for segmenting geometric primitives of a building contour based on enhanced contour perception features, determining a connection order of the geometric primitives, and introducing geometric topology constraints to generate a topologically optimized building contour regular vector polygon;

[0157] The output module is configured to output the roof category of the building and the regular vector polygon result of the building contour.

[0158] The building contour regular polygon extractor comprises a primitive segmentation decoder, an order decoder and a geometric constraint decoder, wherein the primitive segmentation decoder predicts geometric primitive positions through cross-attention and self-attention mechanisms, the order decoder determines a primitive topology order through a self-attention layer, and the geometric constraint decoder disables image features, only uses geometric features of the geometric primitives, and predicts geometric topology relationships between the primitives through a self-attention layer.

[0159] To solve the problems of the existing geometric primitive-based extraction method, i.e., the accuracy of primitive segmentation is reduced due to invalid features introduced by ROI-Align sampling, and the topology structure is unreasonable due to the lack of explicit geometric constraints between primitives, an innovative double-module solution of a contour feature enhancer and a geometric topology constraint decoder is proposed to help accurately measure the development potential of photovoltaic resources. The contour feature enhancer dynamically adjusts the sampling points to the contour edge region through a deformable attention mechanism, generates a boundary segmentation map through multi-head self-attention to provide a supervision signal, and fuses sparse and dense features to enhance contour context perception. The geometric topology constraint decoder encodes the primitive coordinates, applies local neighborhood constraints and global geometric rule constraints through two self-attention layers, and outputs an optimized topology connection matrix. The enhanced features drive the primitive segmenter to accurately predict the corner positions, and the regular vector polygon is directly outputted after geometric topology optimization. The average precision (AP) is improved by 2.4% on the WHU-Mix dataset, which provides high-precision and regularized building roof data and vector contours for photovoltaic resource development potential measurement, and effectively improves the evaluation accuracy of photovoltaic resource development and layout planning.

[0160] In the embodiment, the contour feature enhancer is configured to extract sparse contour perception features and dense contour perception features from multi-scale feature maps and a building bounding box, and fuse the sparse contour perception features and the dense contour perception features to obtain enhanced contour perception features, including:

[0161] The sparse feature extractor is configured to generate reference sampling points in the building bounding box region, and predict offsets of the sampling points relative to a building contour edge through an offset prediction subnetwork to extract sparse contour perception features.

[0162] The dense feature extractor is used for inputting the multi-scale feature map as a query (Query) and the sparse contour perception feature as a key (Key) and a value (Value) into a multi-head self-attention module to generate a building boundary segmentation map and output the dense contour perception feature;

[0163] The feature fusion module is used for constructing a contour buffer mask region based on the building boundary segmentation map, fusing the sparse contour perception feature and the dense contour perception feature according to the mask region, and generating an enhanced contour perception feature.

[0164] The above adjustment system in the application can effectively realize the intelligent inspection method for road infrastructure, and the technical effects thereof are as described in the above embodiment, which will not be repeated here.

[0165] Similarly, the above optimization scheme of the system can also correspondingly realize the optimization effect of the method in Embodiment 1, which will not be repeated here.

[0166] Embodiment 2:

[0167] This embodiment shows the implementation process of identifying the region meeting the photovoltaic installation condition in the building surface based on the original image data by multi-scale feature extraction and fusion and outputting the boundary vector polygon of each region in step S10.

[0168] Firstly, the pre-trained ResNet-50 backbone network is used to extract the multi-scale feature map of the input high-resolution remote sensing image. The compass convolution module reduces the input multi-scale feature map to 64 channels by 1*1 convolution and divides it into 8 8-channel branches, which are processed by 3*3 convolution kernels respectively. After feature splicing, it is fused to 256 channels by 1*1 convolution, and the multi-scale feature map (P2, P3, P4, P5) is output. Then, the feature map is input into the FCOS detection head to predict the two-dimensional bounding box of the building instance. The roof classifier receives the building boundary box output by the detection head, inputs the classifier composed of 3*3 deep convolution layer, ReLU activation function, full connection layer and Softmax activation function, and divides the building roof into flat roof, slope roof, curved roof and multi-wave folded plate roof.

[0169] For each building instance output by the detection head, the contour feature enhancer starts to work. This module first initializes an 8*8 uniform reference point grid on the multi-scale feature map in the bounding box area by the sparse feature extractor, calculates the offset of each point towards the contour edge by using the offset prediction sub-network composed of 5*5 deep convolution, GELU activation, 1*1 convolution and Tanh activation, and extracts the sparse contour perception feature of the offset point set by bilinear interpolation , so that the sampling points are gathered to the effective contour area; then, the dense feature extractor inputs the contour feature map and the sparse contour perception feature into the multi-head self-attention module to generate the dense contour perception feature The original features, serving as both Key and Value, are input into an 8-head self-attention module for interaction. After interaction, the features are processed through multi-scale ROI-Align, convolution, and Sigmoid activation to generate a 32×32 boundary probability map. The image was upsampled back to its original scale as dense contour features. It also provides supervision using binary cross-entropy loss; finally, the feature fusion module is based on... A contour edge buffer mask is constructed, and the mask is applied to the original features to suppress invalid regions. The masked features are then extracted based on the positions of sparse feature points. splicing to form enhanced features .

[0170] Enhanced features The input primitive segmenter module adopts a 6-layer Transformer decoder structure. It processes learnable queries through cross-attention and self-attention mechanisms, and finally outputs corner point category scores, position offsets and confidence scores by the feedforward network.

[0171] The effective corner points predicted by the primitive segmenter are input into a sequential decoder consisting of three self-attention layers and a geometric constraint decoder consisting of two self-attention layers. The sequential decoder predicts the corner connection order probability; subsequently, the geometric constraint decoder begins its work. First, the coordinates of each corner point are positionally encoded; then, the geometric constraint relationships between primitives are learned through two self-attention layers. The first self-attention layer calculates the distance matrix between corner points, focusing on enhancing the geometric positional relationships between neighboring corner points; the second self-attention layer calculates the directional similarity matrix, enhancing the parallel and perpendicular constraints between building outline edges. Finally, the model connects the corner points according to the corner connection order probability and the topology-optimized primitive connection relationships, directly outputting a regular vector polygon. The geometric constraint decoder disables image features, allowing the model to complete the task using only geometric information, thereby improving the model's geometric reasoning and regular building extraction capabilities.

[0172] During training, specific loss functions are used to supervise the output of each part of the convolutional neural network. Focal loss is used to supervise the corner category score; L1 loss is used to supervise the position offset; and cross-entropy loss is used to supervise the predicted corner connection order probability and geometric constraint score.

[0173] To verify the effectiveness and advancement of the proposed method, we compared it with other advanced deep learning building contour extraction methods and validated it on the WHU-Mix aerial dataset, selecting the average accuracy ( ) as the core evaluation indicator, supplemented by and Multi-dimensional performance analysis is performed. The proposed method is compared with representative instance segmentation models including Mask R-CNN, SOLO, and advanced vector polygon extraction methods such as PolarMask, BuildMapper, and Line2Poly. Table 1 shows the experimental results, which indicate that the present application reaches 63.0% on the WHU-Mix dataset with a ResNet-50 backbone network, which is 2.4 percentage points higher than the model P2PFormer and 4.4 percentage points higher than the current optimal vector method Line2Poly. The three indicators (mAP, F1, and F2) all set new performance records. 、 、

[0174] Table 1 Quantitative comparison of WHU-Mix dataset

[0175]

[0176] Although the present application is described in connection with specific features and embodiments thereof, it is evident that various modifications and combinations can be made thereto without departing from the spirit and scope of the application. Accordingly, the present description and drawings are to be regarded simply as illustrative of the present application and are to be construed in such a manner that they cover any and all modifications, variations, combinations and equivalents that fall within the scope of the present application. Obviously, various modifications and changes can be made to the present application by those skilled in the art without departing from the scope of the present application. Thus, it is intended that the present application encompass all such modifications and changes and, accordingly, the application is not to be construed as limited to the examples described herein.​

Claims

1. A method for intelligent assessment of photovoltaic potential, characterized in that, The method comprises the following steps: Based on the original image data, the areas meeting the photovoltaic installation conditions in the building surface are identified through multi-scale feature extraction and fusion, and the boundary vector polygons of each area are output; Multi-scale feature maps are extracted from the input high-resolution remote sensing image; Based on the multi-scale feature maps, the building boundary box is predicted; Based on the building boundary box, the building roof type is classified; From the multi-scale feature maps and the building boundary box, sparse contour perception features and dense contour perception features are extracted, and the sparse contour perception features and the dense contour perception features are fused to obtain enhanced contour perception features; Based on the enhanced contour perception features, the geometric primitives of the building contour are segmented, the connection order of the geometric primitives is determined, and the topological structure is optimized to generate a regular vector polygon of the building contour, which comprises the following steps: The enhanced contour perception features are input into the primitive segmenter, the position coordinates and category confidence of the geometric primitives of the building contour are predicted through cross-attention mechanism and self-attention mechanism by using a learnable query vector; The geometric primitives are input into the order decoder, and the topological connection relationship between the geometric primitives is modeled based on the multi-layer self-attention structure to predict the connection order of each geometric primitive in the contour polygon; The position coordinates and connection order of the geometric primitives are input into the geometric constraint decoder, and the geometric adjacency relationship and angle constraint relationship are modeled through the self-attention mechanism to generate an optimized connection structure that meets the geometric consistency; According to the optimized connection relationship of the geometric primitives, a regular vector polygon of the building contour with reasonable topological structure is constructed; The roof type of the building and the regular vector polygon of the building contour are output; According to the boundary vector polygon, the obstacle occupied area is deducted, and the effective photovoltaic installation area is calculated; The roof type and inclination are identified, and the orientation is determined in combination with the geographic coordinates; In combination with the solar radiation time series data, the roof orientation and inclination, the annual power generation per unit area is predicted; wherein, based on the obstacle position and height marked in the boundary vector polygon, the dynamic shadow is simulated by combining the sun's orbit, and the shading attenuation coefficient is calculated; The effective photovoltaic installation area and the annual power generation per unit area are associated with the geographic information system map to form a photovoltaic potential heat map.

2. The method for intelligent assessment of photovoltaic potential according to claim 1, characterized in that, The multi-scale feature map extraction comprises: Multi-scale feature maps are extracted by using a pre-trained backbone network; A plurality of convolution kernels with different rotation angles are generated by a compass convolution module, multi-directional convolution operations are performed on the multi-scale feature maps, and multi-directional contour features are obtained by splicing and fusing the multi-directional features.

3. The method for intelligent assessment of photovoltaic potential according to claim 1, characterized in that, The method for predicting the building boundary box comprises: The multi-scale feature maps are received, a single-stage detection structure based on full convolution is adopted to predict the coordinates and confidence of the corresponding boundary box at each spatial position of the multi-scale feature maps, and the position information and confidence result of the building boundary box are output.

4. The method for intelligent assessment of photovoltaic potential according to claim 1, characterized in that, The building roof type is classified based on the building boundary box, which comprises the following steps: extracting image features and geometric features in the building bounding box region, wherein the geometric features include roof texture geometric features constructed by learned normals and computed normals; fusing the geometric features and the image features, the fusion being based on a spatial gating mechanism to adjust the fusion proportion of the image features and the geometric features, and being combined with a geometric confidence evaluation module for weighted processing; applying residual connection and spatial-channel dual attention mechanisms to the fused features to enhance the expression ability of the roof region features; inputting the enhanced roof region features into a global average pooling layer, a fully connected layer and a Softmax activation function to output a probability distribution result of the roof type.

5. The method for intelligent assessment of photovoltaic potential according to claim 1, characterized in that, from the multi-scale feature maps and the building bounding box, extracting sparse contour perception features and dense contour perception features, and fusing the sparse contour perception features and the dense contour perception features to obtain enhanced contour perception features, the steps comprising: generating reference sampling points in the building bounding box region, and predicting the offset of each sampling point relative to the building contour edge through an offset prediction sub-network to extract sparse contour perception features; inputting the sparse contour perception features as keys and values and the multi-scale feature maps as queries into a multi-head self-attention module to generate a building boundary segmentation map and output dense contour perception features; based on the building boundary segmentation map, constructing a contour buffer mask region, and fusing the sparse contour perception features and the dense contour perception features according to the mask region to generate enhanced contour perception features.

6. The method for intelligent assessment of photovoltaic potential according to claim 1, characterized in that, based on the optimized connection relationship of the geometric primitives, constructing a building contour regular vector polygon with a reasonable topological structure, the steps comprising: inputting the predicted position coordinates and connection order of the geometric primitives as input, and encoding the geometric features of each geometric primitive into vector form; inputting the vectors into at least two layers of self-attention structure to learn the relative distance constraint relationship and directional constraint relationship between geometric primitives, including parallelism and perpendicularity; based on the probability output of the connection order and the geometric relationship modeled by self-attention, optimizing the connection structure between primitives to generate a connection path that conforms to geometric consistency and topological rationality; connecting the corner points of each geometric primitive in sequence according to the connection path to construct a closed, connected and contour regular building contour vector polygon.

7. The method for intelligent assessment of photovoltaic potential according to claim 1, characterized in that, in combination with solar radiation time series data, roof orientation and inclination, predicting the annual power generation per unit area, including: based on the roof orientation and inclination, performing three-dimensional projection correction on the solar radiation time series data to calculate the actual radiation received by the roof plane; calculating the annual power generation per unit area according to the period time power generation and time dimension accumulation, the period time power generation being calculated as follows: period time power generation = period radiation × attenuation coefficient × photovoltaic module efficiency.

8. An intelligent assessment system for photovoltaic potential, characterized in that, using the method of any one of claims 1 to 7, comprising: an image recognition and vectorization module for identifying regions in the building surface that meet the photovoltaic installation conditions based on original image data through multi-scale feature extraction and fusion, and outputting the boundary vector polygons of each region; An effective area calculation module is configured to calculate an effective photovoltaic installation area by deducting an obstacle occupied area from the boundary vector polygon; A roof attribute identification module is configured to identify a roof type and an inclination angle, and determine an orientation in combination with geographic coordinates; A power generation prediction module is configured to predict an annual power generation per unit area in combination with solar radiation time series data, a roof orientation and an inclination angle; A visualization output module is configured to associate the effective photovoltaic installation area and the annual power generation per unit area to a geographic information system map to form a photovoltaic potential heat map.

9. The intelligent assessment system of photovoltaic potential according to claim 8, characterized in that, The image recognition and vectorization module is configured to recognize areas meeting photovoltaic installation conditions in a building surface based on original image data through multi-scale feature extraction and fusion, and output boundary vector polygons of the areas, and further includes: A feature extractor is configured to extract multi-scale feature maps from input high-resolution remote sensing images; A detection head is configured to predict a building bounding box based on the multi-scale feature maps; A roof classifier is configured to classify a building roof category based on the building bounding box; A contour feature enhancer is configured to extract sparse contour perception features and dense contour perception features from the multi-scale feature maps and the building bounding box, and fuse the sparse contour perception features and the dense contour perception features to obtain enhanced contour perception features; A building contour regular polygon extractor is configured to segment geometric primitives of a building contour based on the enhanced contour perception features, determine a connection order of the geometric primitives, and introduce a geometric topology constraint to generate a topologically optimized building contour regular vector polygon; An output module is configured to output a roof category of a building and a regular vector polygon result of a building contour.

10. The intelligent assessment system of photovoltaic potential according to claim 8, characterized in that, The contour feature enhancer is configured to extract sparse contour perception features and dense contour perception features from the multi-scale feature maps and the building bounding box, and fuse the sparse contour perception features and the dense contour perception features to obtain enhanced contour perception features, and includes: A sparse feature extractor is configured to generate reference sampling points in a building bounding box area, and predict offset amounts of the sampling points relative to a building contour edge through an offset prediction sub-network to extract sparse contour perception features; A dense feature extractor is configured to input the sparse contour perception features as keys and values, and the multi-scale feature maps as queries into a multi-head self-attention module to generate a building boundary segmentation map, and output dense contour perception features; A feature fusion module is configured to construct a contour buffer mask area based on the building boundary segmentation map, and fuse the sparse contour perception features and the dense contour perception features according to the mask area to generate enhanced contour perception features.

Images