Method for constructing spatio-temporal database by fusing supervised segmentation and digital twinning and related equipment

Abstract

The application discloses a kind of spatiotemporal database construction method and related equipment of fusion supervision segmentation and digital twin, belong to geographic information system field, method includes obtaining remote sensing image data and digital twin base space;Loss function constraint term of weak supervision segmentation network is constructed based on prior constraint information;Remote sensing image data is trained based on weak supervision segmentation network to obtain initial spatial distribution mask;Based on digital twin base space, time series signal in remote sensing image data is decoupled to obtain attribute inversion information and generate feedback check signal;Predicted results are calculated based on initial spatial distribution mask and attribute inversion information;The above data is stored as spatiotemporal database, and data meeting preset quality requirement is fed back to digital twin base space to update prior constraint information, with the advantages that misextraction and missing extraction caused by spectral confusion are reduced, and spatial distribution and attribute information are cooperatively checked by feedback mechanism.

Description

Technical Field

[0001] This invention belongs to the field of geographic information system technology, and specifically relates to a method and related equipment for constructing a spatiotemporal database that integrates supervised segmentation and digital twins. Background Technology

[0002] With the accelerated global energy transition, the installed capacity of photovoltaic (PV) systems continues to grow rapidly. Accurately grasping the spatial distribution, installation sequence, type classification, and deployment density of PV systems plays an important foundational supporting role in power grid planning, carbon emission reduction accounting, PV industry policy formulation, and electricity market forecasting.

[0003] Currently, the mainstream technical approach for constructing a spatiotemporal database of photovoltaic facilities relies on high-resolution remote sensing imagery and deep learning models. It employs a sequential processing method of first spatial segmentation and extraction, followed by temporal attribute inversion to build the photovoltaic spatiotemporal database. However, the above approach has the following drawbacks: The photovoltaic panel spectrum is confused with that of buildings and water bodies, and the model lacks prior constraints such as geographical location, geometric arrangement, and orientation, resulting in frequent false positives and false negatives. Remote sensing signals are affected by atmospheric and sensor interference, and the inversion does not utilize physical attributes such as tilt angle, orientation, and conversion efficiency for decoupling, resulting in large deviations in power generation and attenuation rate results and low reliability. Segmentation and inversion are sequential, and attributes cannot be used to correct segmentation errors.

[0004] In summary, existing technologies suffer from problems such as spectral confusion leading to erroneous or missed extractions, low accuracy of attribute inversion, and isolation between segmentation and inversion. Therefore, there is an urgent need in this field for a spatiotemporal database construction method that can integrate spatial-attribute co-verification. Summary of the Invention

[0005] This application provides a spatiotemporal database construction method and related equipment that integrates supervised segmentation and digital twins, which can reduce false extraction and missed extraction caused by spectral confusion, and realize the collaborative verification of spatial distribution and attribute information through a feedback mechanism.

[0006] On the one hand, embodiments of this application provide a method for constructing a spatiotemporal database that integrates supervised segmentation and digital twins, including: Acquire remote sensing image data and digital twin base space of the target area, wherein the digital twin base space includes prior constraint information characterizing the spatial distribution and physical attributes of geographic entities, and the geographic entities include building vector outlines and permanent feature distributions; Based on the prior constraint information, a loss function constraint term for a weakly supervised segmentation network is constructed. The weakly supervised segmentation network is trained based on the remote sensing image data, and the initial spatial distribution mask of the photovoltaic system output by the trained weakly supervised segmentation network is obtained. Based on the digital twin base space, the temporal signals in the remote sensing image data are decoupled by aliasing to obtain the attribute inversion information of the photovoltaic system in the temporal dimension, and a feedback verification signal for the initial spatial distribution mask is generated based on the attribute inversion information. The initial spatial distribution mask and the attribute inversion information are input into a preset scene semantic understanding model to obtain the predicted results of the category label and spatial deployment density of the photovoltaic system. The initial spatial distribution mask, the attribute inversion information, the category label, and the prediction result are fused and stored to obtain the spatiotemporal database. Data in the spatiotemporal database that meets the preset quality requirements are fed back to the digital twin base space to update the prior constraint information.

[0007] Furthermore, the step of decoupling aliased signals from the temporal signals in the remote sensing image data based on the digital twin base space includes: Simulate the time-series characteristic curves of the surface response of the photovoltaic system under different physical states in the digital twin substrate space; The time-series feature curves are transformed into the frequency domain space to generate a twin frequency domain fingerprint database; The pixel-level signals of the low spatial resolution temporal images in the remote sensing image data are sparsely decomposed in the frequency domain space with the twin frequency domain fingerprint database to obtain sparse decomposition coefficients. The installation time of the photovoltaic system is determined based on the step abrupt change time node of the sparse decomposition coefficient, and the sub-pixel area ratio information of the photovoltaic system in the low spatial resolution time series image is inferred based on the amplitude of the sparse decomposition coefficient.

[0008] Furthermore, the feedback verification signal includes an area consistency loss term, and the generation of the feedback verification signal for the initial spatial distribution mask based on the attribute inversion information includes: The sub-pixel area ratio information is mapped to the output space of the weakly supervised segmentation network to construct an area consistency loss function; The area consistency loss function is used as the feedback verification signal and backpropagated to the weakly supervised segmentation network to constrain the global total area of ​​the initial spatial distribution mask and the sub-pixel area ratio information to maintain consistency.

[0009] Furthermore, the prior constraint information includes a distribution map of negative samples of permanent features, and the loss function constraint term for constructing the weakly supervised segmentation network based on the prior constraint information includes: Construct a twin collision loss function, wherein the twin collision loss function is used to calculate the spatial overlap penalty value between the initial spatial distribution mask and the permanent feature negative sample distribution map; If the spatial overlap penalty value is greater than a preset spatial overlap penalty threshold, the weakly supervised segmentation network is forced to set the activation value of the region corresponding to the region with a spatial overlap penalty value greater than the preset spatial overlap penalty threshold to zero.

[0010] Furthermore, the step of inputting the initial spatial distribution mask and the attribute inversion information into the preset scene semantic understanding model includes: Based on the initial spatial distribution mask and the spatial attribution relationship of building vector contours in the digital twin base space, an attribute topology graph is constructed with individual buildings as nodes and spatial adjacency relationships between buildings as edges. The attribute topology graph is processed by a graph neural network to perform message passing and aggregation, thereby obtaining the residential category label or non-residential category label corresponding to the building. The node characteristics of the attribute topology graph include photovoltaic coverage, photovoltaic panel main axis dispersion, and building height.

[0011] Furthermore, obtaining the predicted results of the category label and spatial deployment density of the photovoltaic system includes: In the digital twin base space, a first simulation case and a second simulation case are constructed, wherein the first simulation case learns a historical expansion strategy based on the historical installation time sequence in the attribute inversion information, and the second simulation case learns a limit filling strategy based on the physical available capacity in the digital twin base space. The first simulation case and the second simulation case are driven to perform multi-round game simulation in the digital twin base space. When the difference between the loss functions of the first simulation case and the second simulation case is less than a preset convergence threshold, it is determined that a Nash equilibrium state has been reached, and the spatial probability distribution when the Nash equilibrium state is reached is used as the spatial deployment density prediction result. The loss function of the first simulation case is used to measure the deviation between its deployment plan and the historical expansion strategy, while the loss function of the second simulation case is used to measure the deviation between its deployment plan and the physical available capacity.

[0012] Furthermore, the step of feeding back data from the spatiotemporal database that meets preset quality requirements to the digital twin base space to update the prior constraint information includes: Extract spatiotemporal data from the spatiotemporal database that have a confidence level greater than a preset confidence threshold, and store the extracted spatiotemporal data as spatiotemporal anchor point samples in the anchor point library of the digital twin base space; If the segmentation accuracy of the weakly supervised segmentation network on the current remote sensing image data decreases by more than a preset threshold relative to the historical accuracy, target anchor point samples that match the style features of the current image are extracted from the anchor point library. The parameters of the weakly supervised segmentation network are adjusted based on the target anchor point samples, and the similarity between the adjusted segmentation result contour and the historical contour shape in the anchor point library is calculated. If the similarity is greater than the preset similarity, the adjusted segmentation result contour and the corresponding attribute inversion information are used as the updated prior constraint information, and the updated prior constraint information is stored in the grid cell of the corresponding remote sensing image coverage area in the digital twin base space.

[0013] On the other hand, embodiments of this application provide a spatiotemporal database construction apparatus that integrates supervised segmentation and digital twins, the apparatus comprising: The acquisition module is used to acquire remote sensing image data and digital twin base space of the target area, wherein the digital twin base space includes prior constraint information characterizing the spatial distribution and physical attributes of geographic entities, and the geographic entities include building vector outlines and permanent feature distributions; The construction module is used to construct the loss function constraint terms of the weakly supervised segmentation network based on the prior constraint information; The training module is used to train the weakly supervised segmentation network based on the remote sensing image data and obtain the initial spatial distribution mask of the photovoltaic system output by the trained weakly supervised segmentation network. The decoupling module is used to decouple the aliased signals in the time series signals of the remote sensing image data based on the digital twin base space, to obtain the attribute inversion information of the photovoltaic system in the time series dimension, and to generate a feedback verification signal for the initial spatial distribution mask based on the attribute inversion information. The input module is used to input the initial spatial distribution mask and the attribute inversion information into a preset scene semantic understanding model to obtain the predicted results of the category label and spatial deployment density of the photovoltaic system. The storage module is used to fuse and store the initial spatial distribution mask, the attribute inversion information, the category label and the prediction result to obtain the spatiotemporal database, and to feed back the data in the spatiotemporal database that meets the preset quality requirements to the digital twin base space to update the prior constraint information.

[0014] In another aspect, embodiments of this application provide an electronic device, the device including: a processor and a memory storing computer program instructions; When the processor executes the computer program instructions, it implements the spatiotemporal database construction method of fusion supervised segmentation and digital twin as described in any one of the first aspects.

[0015] In another aspect, embodiments of this application provide a computer storage medium storing computer program instructions, which, when executed by a processor, implement the spatiotemporal database construction method for fusion supervised segmentation and digital twins as described in any one of the first aspects.

[0016] Compared with existing technologies, the spatiotemporal database construction method and related equipment of this application, which integrates supervised segmentation and digital twins, can add prior constraint information such as building vector contours and permanent feature distributions in the digital twin base to the loss function of the weakly supervised segmentation network. This allows the segmentation process to obtain prior information on geographical location and geometric arrangement, thereby suppressing the confusion between photovoltaic panels and spectrally similar objects such as buildings and water bodies, reducing false positives and false negatives. By utilizing physical attributes such as tilt angle, orientation, and conversion efficiency in the digital twin base to decouple the aliasing of remote sensing time-series signals, the inversion results such as power generation and attenuation rate are freed from atmospheric and sensor interference, improving the accuracy of attribute inversion to a certain extent. The attribute inversion information is used to generate feedback verification signals to correct the initial spatial distribution mask, breaking the limitation of the isolation between segmentation and inversion in traditional serial processing, realizing spatial-attribute collaborative verification. The fused data that meets the quality requirements is fed back to the digital twin base to dynamically update the prior information, forming a closed-loop iterative optimization, thereby improving the accuracy of subsequent segmentation and inversion, and ultimately constructing a highly reliable and highly consistent spatiotemporal database of photovoltaic facilities. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this specification or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the embodiments of this specification. For those skilled in the art, other drawings can be obtained based on these drawings. Figure 1 This is a flowchart illustrating a spatiotemporal database construction method that integrates supervised segmentation and digital twins in an embodiment of this application; Figure 2 This is a structural block diagram illustrating a spatiotemporal database construction device that integrates supervised segmentation and digital twins in this embodiment; Figure 3 This is a structural block diagram illustrating an electronic device in this embodiment. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] To facilitate understanding of the embodiments of this application, further explanation and description will be provided below with reference to the accompanying drawings and specific embodiments. These embodiments do not constitute a limitation on the embodiments of this application. In the drawings, the dimensions and relative dimensions of components may be exaggerated for clarity and / or descriptive purposes. When exemplary embodiments can be implemented differently, a specific process sequence may be performed in a different order than that described. For example, two consecutively described processes may be performed substantially simultaneously or in the reverse order of their description. Furthermore, the same reference numerals denote the same components.

[0020] The terminology used herein is for the purpose of describing particular embodiments and is not intended to be limiting. As used herein, unless the context clearly indicates otherwise, the singular forms “a” and “the” are intended to include the plural forms as well. Furthermore, when the terms “comprising” and / or “including” and variations thereof are used in this specification, it indicates the presence of the stated features, integrals, steps, operations, parts, components, and / or groups thereof, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, parts, components, and / or groups thereof. It should also be noted that, as used herein, the terms “substantially,” “about,” and other similar terms are used as approximate terms rather than as terms of degree, thus explaining the inherent biases in measurements, calculated values, and / or provided values ​​that would be recognized by one of ordinary skill in the art.

[0021] Existing methods for constructing spatiotemporal databases for photovoltaic facilities mainly rely on high-resolution remote sensing imagery and deep learning models. They employ a sequential processing approach, first spatial segmentation and extraction, then temporal attribute inversion. This approach suffers from spectral confusion between photovoltaic panels and buildings / water bodies, and the lack of geographical prior constraints in the model leads to frequent errors in extraction and omission. Furthermore, interference with remote sensing signals results in low accuracy and insufficient reliability of attribute inversion. The sequential processing mode of segmentation and inversion prevents attribute information from correcting segmentation errors, thus limiting overall accuracy and reliability.

[0022] To address the problems in the prior art, embodiments of this application provide a method and related equipment for constructing a spatiotemporal database that integrates supervised segmentation and digital twins.

[0023] The following section first introduces a spatiotemporal database construction method that integrates supervised segmentation and digital twin, as provided in the embodiments of this application.

[0024] Figure 1 This illustration shows a flowchart of a spatiotemporal database construction method integrating supervised segmentation and digital twins, according to an embodiment of this application. Figure 1 As shown in the embodiment of this application, a method for constructing a spatiotemporal database that integrates supervised segmentation and digital twins includes the following steps: S101. Acquire remote sensing image data and digital twin base space of the target area, wherein the digital twin base space includes prior constraint information characterizing the spatial distribution and physical attributes of geographic entities, and geographic entities include building vector outlines and permanent feature distributions.

[0025] In this embodiment, remote sensing image data can be acquired through commercial satellite platforms or drone platforms. For example, high-resolution multispectral images can be obtained by subscribing to data providers. The digital twin base space can be constructed through manual modeling, lidar scanning, or importing data from existing geographic information systems. Among these methods, building vector outlines can be drawn manually or extracted using traditional image processing methods, and the distribution of permanent features can be generated based on historical maps or manual annotations.

[0026] S102. Construct the loss function constraint term for the weakly supervised segmentation network based on prior constraint information.

[0027] Specifically, a twin collision loss function is constructed, which is used to calculate the spatial overlap penalty value between the initial spatial distribution mask and the negative sample distribution map of permanent features. If the spatial overlap penalty value is greater than the preset spatial overlap penalty threshold, the weakly supervised segmentation network is forced to set the activation value of the region corresponding to the region with a value greater than the preset spatial overlap penalty threshold to zero.

[0028] The permanent feature negative sample distribution map is a set of geographic entities in the digital twin base space that should not contain photovoltaic systems. These areas include permanent buildings, water bodies, roads, etc., which may be similar to photovoltaic systems in spectral characteristics, but cannot physically install photovoltaic systems. In this embodiment, the permanent feature negative sample distribution map serves as a negative constraint, clearly indicating which areas are non-photovoltaic areas, thereby guiding the weakly supervised segmentation network to avoid identifying photovoltaic systems in these areas. The permanent feature negative sample distribution map can be generated through manual annotation, extraction from existing geographic information system data such as building vector data, water system data, road network data, or automated identification and classification based on high-precision remote sensing images. For example, all known non-photovoltaic areas, such as building roofs, water surfaces, and bare land, can be marked as negative sample areas on the map to form a binary mask map, where the value of the negative sample area is 1, and the value of other areas is 0. In addition, by combining historical data and expert knowledge, geographical entities in the digital twin base space can be classified, and areas unsuitable for photovoltaic installation, such as historical buildings, shaded areas of high-rise buildings, and vegetation-covered areas, can be identified as permanent negative samples and corresponding distribution maps can be generated.

[0029] The twin collision loss function calculates the spatial overlap between the initial spatial distribution mask output by the weakly supervised segmentation network and the permanent feature negative sample distribution map. This twin collision loss function transforms the spatial overlap into an optimizable penalty term, guiding the network to reduce erroneous activations in negative sample regions during training. In this embodiment, the twin collision loss function can be constructed by calculating the sum or average of the pixel-level products of the initial spatial distribution mask and the permanent feature negative sample distribution map. For example, if both are high activation values ​​at a certain pixel (representing photovoltaic and negative samples), a higher penalty is applied. Alternatively, a more complex spatial overlap metric can be used, such as a variant of the Jaccard index or a negative Dice coefficient, but the penalty is applied only to the overlapping regions.

[0030] The spatial overlap penalty value can be obtained by multiplying the activation values ​​of the initial spatial distribution mask with the corresponding values ​​of the permanent feature negative sample distribution map pixel by pixel, and then summing or averaging the products over all pixels to obtain a scalar value representing the total degree of overlap. Alternatively, it can be obtained by calculating the logical AND operation of the two mask images and counting the number or area of ​​non-zero pixels in the AND operation result, which can then be used as the spatial overlap penalty value.

[0031] In this embodiment, after calculating the spatial overlap penalty value in each training iteration, if the spatial overlap penalty value exceeds a preset spatial overlap penalty threshold, the specific regions in the initial spatial distribution mask that cause the overlap are identified, and the activation values ​​of these regions in the mask are directly set to 0. For example, the probability of predicting photovoltaic is set to 0. This operation can be applied as a post-processing step to the output of the weakly supervised segmentation network before or after the loss function calculation. The preset spatial overlap penalty threshold is determined through validation set experiments before training, such as selecting the inflection point of the false positive rate curve. Alternatively, this forced zeroing operation can be integrated into the loss function. For example, by adding a penalty term, when a pixel is simultaneously predicted as photovoltaic and belongs to a negative sample region, its loss will increase sharply, thereby forcing the weakly supervised segmentation network to reduce the activation value of that pixel to zero.

[0032] In this embodiment, the negative sample distribution map of permanent features is used as a negative constraint, and a twin collision loss function is constructed to calculate the overlap between the initial spatial distribution mask and known non-photovoltaic regions. This allows for real-time perception and punishment of unreasonable segmentation results. When the spatial overlap penalty value exceeds the preset spatial overlap penalty threshold, a forced zeroing strategy directly suppresses the activation response of the weakly supervised segmentation network in these known non-photovoltaic regions. This reduces the probability of mis-extraction due to spectral similarity to a certain extent, enabling the weakly supervised segmentation network to accurately identify the real photovoltaic region when extracting the spatial distribution of photovoltaic systems, thus reducing the possibility of misidentifying permanent features as photovoltaic systems.

[0033] S103. Train the weakly supervised segmentation network based on remote sensing image data, and obtain the initial spatial distribution mask of the photovoltaic system output by the trained weakly supervised segmentation network.

[0034] In this embodiment, the weakly supervised segmentation network can adopt a deep learning architecture, such as U-Net or DeepLab. During the training of the weakly supervised segmentation network, remote sensing image data is used as input. The weakly supervised segmentation network outputs a probability map of photovoltaic systems. In the weakly supervised mode, the training data only contains image-level labels, such as indicating whether a photovoltaic system or sparse point labels exist in the image. The weakly supervised segmentation network learns to identify the spatial features of photovoltaic systems from remote sensing images through iterative optimization and generates an initial spatial distribution mask. The initial spatial distribution mask is represented in the form of a binary image, where the pixel value indicates whether the location belongs to a photovoltaic system. For example, the value of a photovoltaic system is 1, and the value of other areas is 0.

[0035] The weakly supervised segmentation network is trained using an overall loss function that includes the twin collision loss function.

[0036] S104. Based on the digital twin base space, the aliased signals in the time series signals of the remote sensing image data are decoupled to obtain the attribute inversion information of the photovoltaic system in the time series dimension, and a feedback verification signal for the initial spatial distribution mask is generated based on the attribute inversion information.

[0037] The process of decoupling aliased signals from time-series remote sensing image data based on a digital twin base space includes: simulating the time-series characteristic curves of the surface response of a photovoltaic system under different physical states in the digital twin base space; transforming the time-series characteristic curves to the frequency domain space to generate a twin frequency domain fingerprint database; performing sparse decomposition of the pixel-level signals of low spatial resolution time-series images in the frequency domain space with the twin frequency domain fingerprint database to obtain sparse decomposition coefficients; determining the installation time of the photovoltaic system based on the step abrupt change time nodes of the sparse decomposition coefficients; and inferring the sub-pixel area ratio information of the photovoltaic system in the low spatial resolution time-series images based on the amplitude of the sparse decomposition coefficients.

[0038] In this embodiment, digital twin technology is used to reproduce the changes in the spectral and thermodynamic response characteristics of a photovoltaic system over time in the real physical world within a virtual environment. For example, simulations are performed based on physical models, using radiative transfer models, energy balance equations, and spectral characteristic parameters of photovoltaic materials to calculate the impact of the photovoltaic system on surface reflectivity or emissivity under different tilt angles, orientations, attenuation levels, and environmental conditions. By adjusting the tilt angle and orientation parameters of the photovoltaic panels in the digital twin model, the impact of different installation methods on the time-series curve of surface reflectivity is simulated, thereby generating a time-series response curve. Simulations can also be performed based on data-driven models, acquiring historical high-resolution remote sensing image data and corresponding ground truth information to train a machine learning model to learn the surface response patterns of the photovoltaic system under different physical states, thereby predicting and generating its time-series characteristic curve.

[0039] In this embodiment, the time-domain signal is decomposed into a superposition of different frequency components by applying Fourier transform, thereby obtaining its amplitude-frequency characteristics and phase-frequency characteristics. The amplitude-frequency characteristics and phase-frequency characteristics constitute a twin frequency-domain fingerprint database of the photovoltaic system in the frequency domain. In addition, time-frequency analysis methods such as wavelet transform can be used to analyze the characteristics of the signal at different frequency scales while preserving time information, thereby generating a twin frequency-domain fingerprint database.

[0040] In this embodiment, sparse representation theory can be used to treat the frequency domain signal of each low spatial resolution pixel as a linear combination of a small number of basis elements in the twin frequency domain fingerprint database, namely photovoltaic systems and other land cover types. For example, a nonnegative matrix factorization algorithm can be used to decompose the pixel signal matrix into a basis matrix (twin frequency domain fingerprint database) and a coefficient matrix (sparse decomposition coefficients). The sparsity constraint of the coefficient matrix ensures that each pixel signal is mainly composed of a few land cover types. In addition, optimization algorithms such as basis pursuit or Lasso regression can be used to find the sparsest representation of the pixel signal given the twin frequency domain fingerprint database as a dictionary, thereby obtaining the corresponding sparse decomposition coefficients.

[0041] Regarding the determination of installation time, since the installation of photovoltaic (PV) systems causes significant and persistent changes in the surface response signal, the step abrupt change points in the time series can be identified by monitoring the PV-related components in the sparse decomposition coefficients. For example, cumulative sum control charts or statistical change point detection algorithms can be used to analyze the time series of the PV system's corresponding coefficients. When the mean or variance of the coefficients changes significantly and persistently, the installation time is determined. For the inverse calculation of sub-pixel area proportion information, the amplitude of the sparse decomposition coefficients reflects the contribution of the PV system to the pixel. By establishing a quantitative mapping relationship between the amplitude of the sparse decomposition coefficients and the actual sub-pixel area proportion, for example, through a linear regression model or nonlinear function, the coverage ratio of the PV system within low spatial resolution pixels can be inverted from the coefficient amplitudes. In this embodiment, the mapping relationship can be pre-set by simulating PV systems with different area proportions in the digital twin substrate space and calculating their corresponding sparse decomposition coefficients.

[0042] In this embodiment, by simulating the time-series characteristic curve of the surface response of the photovoltaic system in the digital twin base space and converting it into a frequency domain fingerprint library, the contribution of the photovoltaic system is separated from the mixed pixel signal through sparse decomposition, thereby overcoming the limitation of traditional methods in distinguishing the photovoltaic system from background features in low-resolution images. Based on the step change point and amplitude of the sparse decomposition coefficients, the installation time of the photovoltaic system is determined, and its area proportion information at the sub-pixel level is inverted, which improves the accuracy and reliability of the time-series attribute inversion of the photovoltaic system to a certain extent.

[0043] The process of generating a feedback verification signal for the initial spatial distribution mask based on attribute inversion information includes: mapping the sub-pixel area ratio information to the output space of the weakly supervised segmentation network and constructing an area consistency loss function; using the area consistency loss function as a feedback verification signal and backpropagating it to the weakly supervised segmentation network to constrain the consistency between the global total area and the sub-pixel area ratio information of the initial spatial distribution mask.

[0044] The feedback verification signal is used during the training of a machine learning or deep learning model to evaluate the difference between the model output and the real target and to guide the adjustment of model parameters. In this embodiment, the feedback verification signal exists in the form of a loss function, which provides feedback by calculating the deviation between the model's predicted result and the expected result. The area consistency loss term is a component of the loss function and is used to measure the consistency between the area represented by the segmentation mask output by the model and the area information obtained by other independent methods, such as physical inversion. When the two are inconsistent, the loss term will generate a penalty value, prompting the model to adjust its output to reduce this inconsistency. During the training process, the area consistency loss function is used as an additional supervision signal to optimize the network parameters together with the main loss function.

[0045] Mapping the subpixel area proportion information to the output space of the weakly supervised segmentation network involves transforming or interpolating the subpixel area proportion information from its original representation, such as the area ratio value of each low-resolution pixel, to a representation with the same resolution and spatial dimension as the initial spatial distribution mask output by the weakly supervised segmentation network. In this embodiment, this can be achieved through various interpolation methods, such as bilinear interpolation, nearest neighbor interpolation, or more complex upsampling techniques, to ensure that the two types of area information can be compared on the same spatial scale.

[0046] The area consistency loss function is used to reflect the difference between the mapped sub-pixel area ratio information and the area represented by the initial spatial distribution mask output by the weakly supervised segmentation network. In this embodiment, the mean square error, cross-entropy loss or other forms of distance metric are used to calculate this difference. The role of the area consistency loss function is to generate a large loss value when the area output by the segmentation network does not match the sub-pixel area ratio information obtained by inversion.

[0047] The area consistency loss function is used as a feedback verification signal and backpropagated to the weakly supervised segmentation network. Thus, during deep learning model training, optimization algorithms such as gradient descent are used to propagate the error signal calculated from the area consistency loss function from the output layer to the input layer layer by layer using a chain rule. This calculates the contribution of each parameter in the network to the total loss. Based on this gradient information, the optimizer updates the network parameters, thereby reducing the loss function value.

[0048] In this embodiment, the sub-pixel area ratio information obtained by decoupling the time-series signal inversion is mapped to the output space of the segmentation network, and an area consistency loss function is constructed. The physical constraint on the total area is transformed into the gradient optimization objective for network training, so that the segmentation result not only depends on the visual features of the image, but also needs to meet the area consistency constraint. The area consistency loss function is used as a feedback verification signal for backpropagation, forcing the network to automatically adjust the boundary and coverage of the segmentation mask when updating parameters, ensuring that the segmented total photovoltaic area is consistent with the physical inversion result, thereby solving the problem of the segmentation result being out of touch with physical facts, and improving the physical credibility of spatial distribution data in the spatiotemporal database to a certain extent.

[0049] S105. Input the initial spatial distribution mask and attribute inversion information into the preset scene semantic understanding model to obtain the prediction results of the photovoltaic system's category label and spatial deployment density.

[0050] The process involves inputting the initial spatial distribution mask and attribute inversion information into a preset scene semantic understanding model. This includes: constructing an attribute topology graph based on the spatial distribution mask and the spatial attribution relationship of building vector contours in the digital twin base space, with individual buildings as nodes and spatial adjacency relationships between buildings as edges; and using a graph neural network to perform message passing and aggregation on the attribute topology graph to obtain the residential or non-residential category label corresponding to the building. The node features of the attribute topology graph include photovoltaic coverage, photovoltaic panel main axis dispersion, and building height.

[0051] The initial spatial distribution mask is a binary distribution map of the photovoltaic system extracted from remote sensing image data using a weakly supervised segmentation network. Pixel values ​​indicate the presence of a photovoltaic system at that location. In this embodiment, the association between the photovoltaic system and the building can be determined through pixel-level overlap analysis or spatial inclusion relationships. The building vector contour in the digital twin base space is pre-stored geometric data describing the building boundaries, existing in the form of polygons or sets of line segments in this embodiment. The vector contour provides the location and shape information of the building in geographic space and is key to identifying individual building entities and their spatial adjacency relationships. For example, these contours can be obtained from high-precision map data, CAD drawings, or LiDAR scanning data. Spatial attribution relationships determine the correspondence between the photovoltaic system and a specific building vector contour. This relationship can be the geometric overlap, inclusion, or proximity of the photovoltaic system mask and the building contour. For example, the building contour with the largest intersection area can be used as the building to which the photovoltaic system belongs by calculating the intersection area between the photovoltaic system mask and each building contour; or, for a photovoltaic system located on a building roof, it can be directly determined whether it is completely included by the building contour.

[0052] Constructing an attribute topology map using individual buildings as nodes and spatial adjacency relationships between buildings as edges applies a graph-structured data model to geographic entities and their interrelationships. Using individual buildings as nodes indicates that each node in the attribute topology map represents an independent building entity, carrying its associated attribute information. Using spatial adjacency relationships as edges indicates that if two buildings are geographically adjacent or share a boundary, an edge exists between them in the topology map. This edge can represent their spatial interaction or influence. For example, the distance between building outlines can be calculated; if the distance is less than a preset threshold, an adjacency relationship is considered to exist. Alternatively, adjacent buildings can be identified based on the spatial topology analysis function of a geographic information system.

[0053] Graph Neural Networks (GNNs) are deep learning models specifically designed for processing graph-structured data. They learn node representations by passing and aggregating information between nodes and edges in the graph, and update the feature representation of the current node using information from neighboring nodes, thereby capturing complex relationships in the graph structure. In this embodiment, graph convolutional networks, graph attention networks, or graph sampling and aggregation models can be used. Message passing and aggregation are the core mechanisms for learning in GNNs. Message passing is the process by which a node sends its own feature information to its neighboring nodes. Aggregation is the process by which a node receives messages from its neighboring nodes and integrates them, such as by performing operations like summation, averaging, or max pooling, to update its own feature representation. Through multiple rounds of message passing and aggregation, nodes can fuse information from their local neighborhood and even the global graph structure, thereby learning more expressive features. For example, in each layer of a GNN, each node collects features from its neighboring nodes, then combines these features with its own features using an aggregation function such as averaging, and finally generates new node features using a nonlinear transformation such as ReLU.

[0054] Among them, node features are the numerical or vector information carried by each node, i.e., each building, used to describe its own attributes. These features are the basic inputs for graph neural networks to learn and classify. For example, node features can be numerical, categorical, or embedded vectors. Photovoltaic coverage rate is the proportion of the area covered by photovoltaic systems on the roof or facade of a single building to the total usable area of ​​the building. This feature reflects the degree of utilization of photovoltaic deployment in a building and is one of the important indicators for judging the functional attributes of a building. In this embodiment, it is obtained by calculating the ratio of the number of pixels belonging to the building in the initial spatial distribution mask to the total number of pixels within the building outline. The photovoltaic panel principal axis dispersion is the degree of concentration or variation of the distribution of photovoltaic panel arrays installed on a single building along their principal axis. High dispersion indicates diverse installation directions of photovoltaic panels, which is related to the complexity or multifunctionality of building structures. For example, dispersion can be determined by statistically analyzing the main axis directions of all photovoltaic panels within a building and calculating their standard deviation or entropy value. Building height is the vertical height of a single building, and it is a physical attribute that affects lighting conditions, roof area, and building type. For example, it can be calculated by the difference between digital surface models and digital elevation models, or obtained from pre-stored building attribute data in a digital twin base space. Among these, residential buildings have relatively low photovoltaic coverage and are scattered, while industrial parks or commercial buildings have higher photovoltaic coverage and are more regularly arranged. By aggregating neighborhood information through graph neural networks, the category of a building can be accurately identified based on the similarity of its functional zoning.

[0055] In this embodiment, based on the correspondence between the initial spatial distribution mask and the building vector outline in the digital twin, an attribute topology graph with buildings as nodes and spatial adjacency as edges is constructed. A graph neural network is used for message passing and neighborhood aggregation. By combining the local features of the building and neighborhood information, the building attributes are determined, thereby identifying the photovoltaic category. This overcomes the problem of ambiguous attribute judgment in complex urban environments using traditional methods.

[0056] The process of obtaining the category label and spatial deployment density prediction results of photovoltaic systems includes: constructing a first simulation case and a second simulation case in a digital twin base space, wherein the first simulation case learns a historical expansion strategy based on the historical installation time sequence in attribute inversion information, and the second simulation case learns a limit filling strategy based on the physical available capacity in the digital twin base space; driving the first simulation case and the second simulation case to conduct multi-round game simulations in the digital twin base space, and determining that a Nash equilibrium state has been reached when the difference between the loss functions of the first simulation case and the second simulation case is less than a preset convergence threshold, and using the spatial probability distribution when the Nash equilibrium state is reached as the spatial deployment density prediction result; wherein the loss function of the first simulation case is used to measure the deviation of its deployment scheme from the historical expansion strategy, and the loss function of the second simulation case is used to measure the deviation of its deployment scheme from the physical available capacity.

[0057] In this embodiment, the first and second simulation cases are virtual agents or models used to simulate the deployment behavior of photovoltaic systems. They can be instantiated in the digital twin base space as software modules, intelligent agents, or simulators. For example, a set of rules, parameters, and behavioral patterns can be defined to construct the first and second simulation cases, or their behavior can be characterized through machine learning models. The historical expansion strategy refers to the pattern or rule of how the photovoltaic system has been gradually installed and expanded spatially over a past period. Learning the historical expansion strategy can be achieved by analyzing the photovoltaic system history recorded in attribute inversion information. Installation time, location, scale, and other time-series data are used to capture historical expansion trends and preferences. For example, time series analysis, Markov chain models, reinforcement learning, or agent-based simulation methods can be used. The limit-filling strategy is the maximum deployment density and spatial layout pattern that a photovoltaic system can achieve under given physical space constraints. Learning the limit-filling strategy requires comprehensive consideration of physical available capacity information such as building roof area, orientation, slope, shading, and structural load-bearing capacity in the digital twin base space. For example, optimization algorithms such as linear programming, integer programming, or physical rule-based simulation can be used to determine the maximum deployment scheme that satisfies all physical constraints.

[0058] Multi-round game simulation is a process in which the first and second simulation examples interact and compete in a digital twin-based space. In each round of simulation, the first and second simulation examples predict or adjust photovoltaic deployment according to their respective strategies and evaluate each other's actions. This simulation can be implemented through iterative optimization, Monte Carlo simulation, or agent-based multi-agent systems. The difference in the loss function is used to quantify the degree of disagreement or conflict between the first and second simulation examples on the current deployment scheme. For example, it can be set as a measure of the difference between the spatial deployment schemes predicted by the first and second simulation examples, or the difference in the objective function values ​​of their respective strategies. The preset convergence threshold is set as needed and is not specifically limited here. When the difference in the loss function is less than the preset convergence threshold, it is determined that the strategies of the first and second simulation examples have become consistent or reached a stable compromise state.

[0059] In game theory, Nash equilibrium is a state in which each participant has chosen their optimal strategy, and with the strategies of other participants remaining unchanged, no one can gain a greater benefit by changing their own strategy. In this scenario, it represents a dynamic balance between historical expansion trends and physical limits. Spatial probability distribution is the likelihood that each potential deployment location in the digital twin's base space will be occupied by a photovoltaic system. In Nash equilibrium, this probability distribution reflects the most likely deployment density and spatial layout of photovoltaic systems under the combined influence of historical expansion trends and physical limits. For example, it could be a two-dimensional grid, where each grid cell stores a probability value between 0 and 1, representing the likelihood of deploying photovoltaics in that area.

[0060] For example, the multi-round game simulation adopts the following approach: In the first round of iteration, the first simulation case generates an initial deployment scheme S1 based on the historical installation density, and the second simulation case generates an initial deployment scheme S2 based on the physical capacity constraint. In each subsequent round of iteration, the pixel-level mean square error of S1 and S2 is calculated as the difference in the loss function. When the difference in the loss function is less than 0.01, the iteration stops, and the pixel-level weighted average of S1 and S2 is used as the final spatial probability distribution.

[0061] In this embodiment, a first and a second computational instance are constructed in the digital twin base space, and the first and second computational instances are driven to play a multi-round game. The convergence is determined by the difference in the loss function, and finally a spatial probability distribution that takes into account both historical patterns and physical constraints is obtained. This overcomes the shortcomings of relying solely on static semantic models, which are difficult to capture dynamic evolution and lack physical capacity considerations, and makes the prediction results closer to actual planning and physical environment.

[0062] S106. The initial spatial distribution mask, attribute inversion information, category labels and prediction results are fused and stored to obtain a spatiotemporal database. Data in the spatiotemporal database that meets the preset quality requirements are fed back to the digital twin base space to update the prior constraint information.

[0063] In this embodiment, converged storage can integrate all relevant data, such as photovoltaic system ID, spatial coordinates, installation time, area, category, deployment density, etc., into a relational database or NoSQL database.

[0064] The process of feeding back data from the spatiotemporal database that meets preset quality requirements to the digital twin base space to update the prior constraint information includes: extracting spatiotemporal data from the spatiotemporal database with a confidence level greater than a preset confidence threshold, and storing the extracted spatiotemporal data as spatiotemporal anchor point samples in the anchor point library of the digital twin base space; when the segmentation accuracy of the weakly supervised segmentation network on the current remote sensing image data decreases by more than a preset threshold compared to historical accuracy, extracting target anchor point samples that match the style features of the current image from the anchor point library; adjusting the parameters of the weakly supervised segmentation network based on the target anchor point samples, and calculating the similarity between the adjusted segmentation result contour and the historical contour shape in the anchor point library; when the similarity is greater than a preset similarity, using the adjusted segmentation result contour and the corresponding attribute inversion information as updated prior constraint information, and storing the updated prior constraint information in the grid cells of the corresponding remote sensing image coverage area in the digital twin base space.

[0065] The confidence level can be evaluated based on metrics such as pixel-level probability values ​​output by the weakly supervised segmentation network, cross-entropy loss predicted by the model, or voting consistency in ensemble learning. The pre-set confidence threshold can be set as needed, for example, 0.95 or 0.99. The anchor library can be a relational database, a NoSQL database, or a distributed file system. The anchor library includes, but is not limited to, segmentation masks, attribute inversion information, timestamps, and geographic coordinates.

[0066] The detection of decreased segmentation accuracy can be achieved by periodically evaluating the performance metrics of the weakly supervised segmentation network on an independent validation set, such as Intersection over Union (IoU) and F1-score, or by comparing it with known high-precision geographic entity data in the digital twin base space. When the decrease in these metrics exceeds a preset threshold, a model update mechanism is triggered. The preset threshold can be set as needed, such as 5% or 10%. Image style features include, but are not limited to, texture features, color histograms, spectral features, or deep features extracted by a pre-trained deep learning feature extraction network. By calculating the distance between these feature vectors, such as Euclidean distance or cosine similarity, the anchor point sample that is closest in style to the current remote sensing image data is identified and extracted.

[0067] Parameter tuning can employ fine-tuning strategies from transfer learning, such as freezing the parameters of the first few layers of the network and training only the later layers, or performing end-to-end training of the entire network with a small learning rate. Furthermore, incremental learning or online learning methods can be used to gradually absorb new knowledge while retaining the model's existing knowledge. The similarity of contour shapes can be quantified using various geometric metrics, such as intersection-over-union ratio, Dice coefficient, and Hausdorff distance, or by comparing shape descriptors such as Fourier descriptors or Zernike moments.

[0068] The preset similarity threshold can be set according to the actual application's requirements for update accuracy. For example, if the IoU is greater than 0.8 or 0.9, the updated prior constraint information can be stored in vector data format, such as GeoJSON, Shapefile, or raster data format, and associated with geographic grid cells in the digital twin base space through spatial indexes, such as quadtrees or R-trees, to facilitate subsequent rapid retrieval and application.

[0069] In this embodiment, an anchor point sample library is constructed by extracting high-confidence spatiotemporal data to ensure the reliability of the data source used for model updates. When a decrease in the segmentation accuracy of the weakly supervised segmentation network on the current remote sensing image is detected, the network parameters are adjusted by matching the target anchor point samples most relevant to the style features of the current image from the anchor point library. This enables adaptive learning for specific environmental changes, overcoming the problems of insufficient generalization ability and performance degradation of the model in different spatiotemporal scenarios. After adjustment, the similarity between the segmentation result contour and the historical contour in the anchor point library is calculated for verification, ensuring the consistency and stability of the updated prior constraint information in spatial geometric features.

[0070] Based on the spatiotemporal database construction method integrating supervised segmentation and digital twins provided in the above embodiments, this application also provides a specific implementation of a spatiotemporal database construction apparatus integrating supervised segmentation and digital twins. Please refer to the following embodiments.

[0071] First see Figure 2 The spatiotemporal database construction apparatus 200 integrating supervised segmentation and digital twin provided in this application embodiment includes: The acquisition module 201 is used to acquire remote sensing image data and digital twin base space of the target area. The digital twin base space includes prior constraint information that characterizes the spatial distribution and physical attributes of geographic entities. The geographic entities include building vector outlines and permanent feature distributions. Module 202 is used to construct the loss function constraint terms of the weakly supervised segmentation network based on prior constraint information; Training module 203 is used to train a weakly supervised segmentation network based on remote sensing image data and obtain the initial spatial distribution mask of the photovoltaic system output by the trained weakly supervised segmentation network. The decoupling module 204 is used to decouple the aliased signals in the time series signals of the remote sensing image data based on the digital twin base space, to obtain the attribute inversion information of the photovoltaic system in the time series dimension, and to generate a feedback verification signal for the initial spatial distribution mask based on the attribute inversion information. The input module 205 is used to input the initial spatial distribution mask and attribute inversion information into the preset scene semantic understanding model to obtain the prediction results of the category label and spatial deployment density of the photovoltaic system; Storage module 206 is used to fuse and store the initial spatial distribution mask, attribute inversion information, category labels and prediction results to obtain a spatiotemporal database, and to feed back the data in the spatiotemporal database that meets the preset quality requirements to the digital twin base space to update the prior constraint information.

[0072] As an optional implementation of this embodiment, the decoupling module 204 is specifically used for: The time-series characteristic curves of the surface response of a photovoltaic system under different physical states are simulated in a digital twin base space. The time-series characteristic curves are transformed into the frequency domain space to generate a twin frequency domain fingerprint database. The pixel-level signals of low spatial resolution time-series images in remote sensing image data are sparsely decomposed in the frequency domain space with the twin frequency domain fingerprint database to obtain sparse decomposition coefficients. The installation time of the photovoltaic system is determined based on the step change time nodes of the sparse decomposition coefficients, and the sub-pixel area ratio information of the photovoltaic system in the low spatial resolution time-series images is inferred based on the amplitude of the sparse decomposition coefficients.

[0073] As an optional implementation of this embodiment, the decoupling module 204 is further specifically used for: The subpixel area proportion information is mapped to the output space of the weakly supervised segmentation network to construct an area consistency loss function. The area consistency loss function is used as a feedback verification signal and backpropagated to the weakly supervised segmentation network to constrain the global total area of ​​the initial spatial distribution mask and the subpixel area proportion information to maintain consistency.

[0074] As an optional implementation of this embodiment, the construction module 202 is specifically used for: A twin collision loss function is constructed, which is used to calculate the spatial overlap penalty value of the initial spatial distribution mask and the negative sample distribution map of permanent features. If the spatial overlap penalty value is greater than the preset spatial overlap penalty threshold, the weakly supervised segmentation network is forced to set the activation value of the region corresponding to the region with a spatial overlap penalty value greater than the preset spatial overlap penalty threshold to zero.

[0075] As an optional implementation of this embodiment, the input module 205 is specifically used for: Based on the initial spatial distribution mask and the spatial attribution relationships of building vector outlines in the digital twin base space, an attribute topology graph is constructed with individual buildings as nodes and spatial adjacency relationships between buildings as edges. A graph neural network is used to perform message passing and aggregation on the attribute topology graph to obtain the residential or non-residential category labels corresponding to the buildings. The node features of the attribute topology graph include photovoltaic coverage, the dispersion of the photovoltaic panel's main axis, and the building height.

[0076] As an optional implementation of this embodiment, the input module 205 is further specifically used for: In a digital twin base space, a first simulation case and a second simulation case are constructed. The first simulation case learns a historical expansion strategy based on the historical installation time sequence from attribute inversion information, while the second simulation case learns a limit filling strategy based on the physical available capacity in the digital twin base space. The first and second simulation cases are driven to conduct multi-round game simulations in the digital twin base space. When the difference between the loss functions of the first and second simulation cases is less than a preset convergence threshold, a Nash equilibrium state is determined to be reached, and the spatial probability distribution at the Nash equilibrium state is used as the spatial deployment density prediction result. The loss function of the first simulation case is used to measure the deviation between its deployment plan and the historical expansion strategy, while the loss function of the second simulation case is used to measure the deviation between its deployment plan and the physical available capacity.

[0077] As an optional implementation of this embodiment, the storage module 206 is specifically used for: Spatiotemporal data with confidence levels greater than a preset confidence threshold are extracted from the spatiotemporal database and stored as spatiotemporal anchor point samples in the anchor point library of the digital twin base space. When the segmentation accuracy of the weakly supervised segmentation network on the current remote sensing image data decreases by more than a preset threshold compared to historical accuracy, target anchor point samples matching the style features of the current image are extracted from the anchor point library. Based on the target anchor point samples, the parameters of the weakly supervised segmentation network are adjusted, and the similarity between the adjusted segmentation result contour and the historical contour shape in the anchor point library is calculated. If the similarity is greater than a preset similarity, the adjusted segmentation result contour and the corresponding attribute inversion information are used as updated prior constraint information, and the updated prior constraint information is stored in the grid cell of the corresponding remote sensing image coverage area in the digital twin base space.

[0078] Figure 3 A schematic diagram of the hardware structure of the electronic device provided in an embodiment of this application is shown.

[0079] The electronic device may include a processor 301 and a memory 302 storing computer program instructions.

[0080] Specifically, the processor 301 may include a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0081] Memory 302 may include mass storage for data or instructions. For example, and not limitingly, memory 302 may include a hard disk drive (HDD), floppy disk drive, flash memory, optical disk, magneto-optical disk, magnetic tape, or Universal Serial Bus (USB) drive, or a combination of two or more of these. In one instance, memory 302 may include removable or non-removable (or fixed) media, or memory 302 may be a non-volatile solid-state memory. Memory 302 may be internal or external to an electronic device.

[0082] In one instance, memory 302 may be read-only memory (ROM). In one instance, the ROM may be a mask-programmed ROM, a programmable ROM (PROM), an erasable PROM (EPROM), an electrically erasable PROM (EEPROM), an electrically rewritable ROM (EAROM), or flash memory, or a combination of two or more of these.

[0083] Memory 302 may include read-only memory (ROM), random access memory (RAM), disk storage media device, optical storage media device, flash memory device, electrical, optical, or other physical / tangible memory storage device. Therefore, typically, memory includes one or more tangible (non-transitory) computer-readable storage media (e.g., memory devices) encoded with software including computer-executable instructions, and when the software is executed (e.g., by one or more processors), it is operable to perform the operations described with reference to a method for constructing a spatiotemporal database integrating fusion supervised segmentation and digital twins according to a first aspect of this disclosure.

[0084] The processor 301 reads and executes computer program instructions stored in the memory 302 to achieve... Figure 1 The embodiment shown illustrates a method for constructing a spatiotemporal database that integrates supervised segmentation and digital twins.

[0085] In one example, the electronic device may also include a communication interface 303 and a bus 304. For example, Figure 3 As shown, the processor 301, memory 302, and communication interface 303 are connected through bus 304 and complete communication with each other.

[0086] The communication interface 303 is mainly used to realize communication between various modules, devices, units and / or equipment in the embodiments of this application.

[0087] Bus 304 includes hardware, software, or both, that couples components of an electronic device together. For example, and not as a limitation, the bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Extended Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), a HyperTransport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an Infinite Bandwidth Interconnect, a Low Pin Count (LPC) bus, a memory bus, a Microchannel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local (VLB) bus, or other suitable buses, or combinations of two or more of these. Where appropriate, bus 304 may include one or more buses. Although specific buses are described and illustrated in embodiments of this application, this application contemplates any suitable bus or interconnect.

[0088] This electronic device can execute a spatiotemporal database construction method that integrates supervised segmentation and digital twins, as described in this application embodiment, thereby achieving a combination of... Figures 1-2 This paper describes a method and apparatus for constructing a spatiotemporal database that integrates supervised segmentation and digital twins.

[0089] Furthermore, in conjunction with the spatiotemporal database construction method integrating supervised segmentation and digital twins in the above embodiments, this application embodiment can provide a computer storage medium for implementation. The computer storage medium stores computer program instructions; when these computer program instructions are executed by a processor, they implement any of the spatiotemporal database construction methods integrating supervised segmentation and digital twins in the above embodiments.

[0090] It should be clarified that this application is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of this application is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application.

[0091] The functional blocks shown in the above block diagram can be implemented as hardware, software, firmware, or a combination thereof. When implemented in hardware, they can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this application are programs or code segments used to perform the required tasks. Programs or code segments can be stored on a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried on a carrier wave. "Machine-readable medium" can include any medium capable of storing or transmitting information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, ROM, flash memory, erasable ROM (EROM), floppy disks, CD-ROMs, optical disks, hard disks, fiber optic media, radio frequency (RF) links, etc. Code segments can be downloaded via computer networks such as the Internet, intranets, etc.

[0092] It should also be noted that the exemplary embodiments mentioned in this application describe methods or systems based on a series of steps or apparatus. However, this application is not limited to the order of the above steps; that is, the steps can be performed in the order mentioned in the embodiments, or in a different order, or several steps can be performed simultaneously.

[0093] The aspects of this disclosure have been described above with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It should be understood that each block in the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that these instructions, executable via the processor of the computer or other programmable data processing apparatus, enable the implementation of the functions / actions specified in one or more blocks of the flowchart illustrations and / or block diagrams. Such a processor can be, but is not limited to, a general-purpose processor, a special-purpose processor, a special application processor, or a field-programmable logic circuit. It is also understood that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can also be implemented by special-purpose hardware performing the specified functions or actions, or can be implemented by a combination of special-purpose hardware and computer instructions.

[0094] The above description is merely a specific implementation of this application. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. It should be understood that the protection scope of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the protection scope of this application.

Claims

1. A method for constructing a spatiotemporal database integrating supervised segmentation and digital twins, characterized in that, include: Acquire remote sensing image data and digital twin base space of the target area, wherein the digital twin base space includes prior constraint information characterizing the spatial distribution and physical attributes of geographic entities, and the geographic entities include building vector outlines and permanent feature distributions; Based on the prior constraint information, a loss function constraint term for a weakly supervised segmentation network is constructed. The weakly supervised segmentation network is trained based on the remote sensing image data, and the initial spatial distribution mask of the photovoltaic system output by the trained weakly supervised segmentation network is obtained. Based on the digital twin base space, the temporal signals in the remote sensing image data are decoupled by aliasing to obtain the attribute inversion information of the photovoltaic system in the temporal dimension, and a feedback verification signal for the initial spatial distribution mask is generated based on the attribute inversion information. The initial spatial distribution mask and the attribute inversion information are input into a preset scene semantic understanding model to obtain the predicted results of the category label and spatial deployment density of the photovoltaic system. The initial spatial distribution mask, the attribute inversion information, the category label, and the prediction result are fused and stored to obtain the spatiotemporal database. Data in the spatiotemporal database that meets the preset quality requirements are fed back to the digital twin base space to update the prior constraint information.

2. The spatiotemporal database construction method integrating supervised segmentation and digital twins according to claim 1, characterized in that, The aliasing signal decoupling based on the digital twin base space in the temporal signals of the remote sensing image data includes: Simulate the time-series characteristic curves of the surface response of the photovoltaic system under different physical states in the digital twin substrate space; The time-series feature curves are transformed into the frequency domain space to generate a twin frequency domain fingerprint database; The pixel-level signals of the low spatial resolution temporal images in the remote sensing image data are sparsely decomposed in the frequency domain space with the twin frequency domain fingerprint database to obtain sparse decomposition coefficients. The installation time of the photovoltaic system is determined based on the step abrupt change time node of the sparse decomposition coefficient, and the sub-pixel area ratio information of the photovoltaic system in the low spatial resolution time series image is inferred based on the amplitude of the sparse decomposition coefficient.

3. The spatiotemporal database construction method integrating supervised segmentation and digital twins according to claim 2, characterized in that, The feedback verification signal includes an area consistency loss term, and the generation of the feedback verification signal for the initial spatial distribution mask based on the attribute inversion information includes: The sub-pixel area ratio information is mapped to the output space of the weakly supervised segmentation network to construct an area consistency loss function; The area consistency loss function is used as the feedback verification signal and backpropagated to the weakly supervised segmentation network to constrain the global total area of ​​the initial spatial distribution mask and the sub-pixel area ratio information to maintain consistency.

4. The spatiotemporal database construction method integrating supervised segmentation and digital twins according to claim 1, characterized in that, The prior constraint information includes a distribution map of negative samples of permanent land features, and the loss function constraint terms for constructing a weakly supervised segmentation network based on the prior constraint information include: Construct a twin collision loss function, wherein the twin collision loss function is used to calculate the spatial overlap penalty value between the initial spatial distribution mask and the permanent feature negative sample distribution map; If the spatial overlap penalty value is greater than a preset spatial overlap penalty threshold, the weakly supervised segmentation network is forced to set the activation value of the region corresponding to the region with a spatial overlap penalty value greater than the preset spatial overlap penalty threshold to zero.

5. The spatiotemporal database construction method integrating supervised segmentation and digital twins according to claim 1, characterized in that, The step of inputting the initial spatial distribution mask and the attribute inversion information into the preset scene semantic understanding model includes: Based on the initial spatial distribution mask and the spatial attribution relationship of building vector contours in the digital twin base space, an attribute topology graph is constructed with individual buildings as nodes and spatial adjacency relationships between buildings as edges. The attribute topology graph is processed by a graph neural network to perform message passing and aggregation, thereby obtaining the residential category label or non-residential category label corresponding to the building. The node characteristics of the attribute topology graph include photovoltaic coverage, photovoltaic panel main axis dispersion, and building height.

6. The spatiotemporal database construction method integrating supervised segmentation and digital twins according to claim 1, characterized in that, The predicted results of obtaining the category label and spatial deployment density of the photovoltaic system include: In the digital twin base space, a first simulation case and a second simulation case are constructed, wherein the first simulation case learns a historical expansion strategy based on the historical installation time sequence in the attribute inversion information, and the second simulation case learns a limit filling strategy based on the physical available capacity in the digital twin base space. The first simulation case and the second simulation case are driven to perform multi-round game simulation in the digital twin base space. When the difference between the loss functions of the first simulation case and the second simulation case is less than a preset convergence threshold, it is determined that a Nash equilibrium state has been reached, and the spatial probability distribution when the Nash equilibrium state is reached is used as the spatial deployment density prediction result. The loss function of the first simulation case is used to measure the deviation between its deployment plan and the historical expansion strategy, while the loss function of the second simulation case is used to measure the deviation between its deployment plan and the physical available capacity.

7. The spatiotemporal database construction method integrating supervised segmentation and digital twins according to claim 1, characterized in that, The step of feeding back data from the spatiotemporal database that meets preset quality requirements to the digital twin base space to update the prior constraint information includes: Extract spatiotemporal data from the spatiotemporal database that have a confidence level greater than a preset confidence threshold, and store the extracted spatiotemporal data as spatiotemporal anchor point samples in the anchor point library of the digital twin base space; If the segmentation accuracy of the weakly supervised segmentation network on the current remote sensing image data decreases by more than a preset threshold relative to the historical accuracy, target anchor point samples that match the style features of the current image are extracted from the anchor point library. The parameters of the weakly supervised segmentation network are adjusted based on the target anchor point samples, and the similarity between the adjusted segmentation result contour and the historical contour shape in the anchor point library is calculated. If the similarity is greater than the preset similarity, the adjusted segmentation result contour and the corresponding attribute inversion information are used as the updated prior constraint information, and the updated prior constraint information is stored in the grid cell of the corresponding remote sensing image coverage area in the digital twin base space.

8. A spatiotemporal database construction device integrating supervised segmentation and digital twins, characterized in that, include: The acquisition module is used to acquire remote sensing image data and digital twin base space of the target area, wherein the digital twin base space includes prior constraint information characterizing the spatial distribution and physical attributes of geographic entities, and the geographic entities include building vector outlines and permanent feature distributions; The construction module is used to construct the loss function constraint terms of the weakly supervised segmentation network based on the prior constraint information; The training module is used to train the weakly supervised segmentation network based on the remote sensing image data and obtain the initial spatial distribution mask of the photovoltaic system output by the trained weakly supervised segmentation network. The decoupling module is used to decouple the aliased signals in the time series signals of the remote sensing image data based on the digital twin base space, to obtain the attribute inversion information of the photovoltaic system in the time series dimension, and to generate a feedback verification signal for the initial spatial distribution mask based on the attribute inversion information. The input module is used to input the initial spatial distribution mask and the attribute inversion information into a preset scene semantic understanding model to obtain the predicted results of the category label and spatial deployment density of the photovoltaic system. The storage module is used to fuse and store the initial spatial distribution mask, the attribute inversion information, the category label and the prediction result to obtain the spatiotemporal database, and to feed back the data in the spatiotemporal database that meets the preset quality requirements to the digital twin base space to update the prior constraint information.

9. An electronic device, characterized in that, include: Processor and memory storing computer program instructions; When the processor executes the computer program instructions, it implements the spatiotemporal database construction method that integrates supervised segmentation and digital twins as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions, which, when executed by a processor, implement the spatiotemporal database construction method for fusion supervised segmentation and digital twins as described in any one of claims 1-7.

Images