Urban green land carbon sink dynamic accounting method and system based on multi-source data fusion and spatial network modeling

Abstract

The invention discloses an urban green land carbon sink dynamic accounting method and system based on multi-source data fusion and spatial network modeling. The method comprises the following steps: collecting urban green space carbon sink original data and preprocessing the data, and constructing a space * time * feature structure carbon sink data cube; identifying green land source plaques based on the spatial dimension data; taking the source ground plaque as an initial node, constructing a resistance surface by using a dynamic resistance factor, simulating a carbon flow in combination with a circuit theory, and generating a carbon flow gallery and a carbon flow density grid; inputting multi-source data into the carbon density dynamic calibration model, and outputting a dynamic grid; constructing a multi-scale prediction engine, and outputting carbon density prediction results under different time scales; and combining GIS and BIM technologies, inputting a prediction result and an urban space structure into a three-dimensional visual platform, generating a carbon density thermodynamic diagram and a carbon flow gallery evolution diagram, and realizing carbon sink dynamic accounting. According to the invention, the precision and space-time continuity of carbon sink accounting can be obviously improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of carbon sink accounting, and more particularly relates to a method and system for dynamic accounting of urban green land carbon sink based on multi-source data fusion and spatial network modeling. BACKGROUND

[0002] With the advancement of the "double carbon" goal, the strategic position of urban green land in the carbon sink system is increasingly prominent. The unit green land carbon sink capacity has been included in the mandatory evaluation index in national policy documents, and urban green land carbon sink accounting has gradually moved from academic research to practical application. However, the current mainstream carbon sink accounting method still faces core bottlenecks such as difficulty in multi-source data fusion, lack of spatial structure cognition, and lack of dynamic prediction ability, making it difficult to meet the needs of carbon sink management at the city level, in real time, and in a systematic manner.

[0003] First, in terms of data processing, existing carbon sink estimation mostly uses a single modal data source (such as remote sensing images or plot surveys), lacking a structured organization and fusion mechanism for multi-source heterogeneous data such as satellite remote sensing, unmanned aerial laser radar, and ground flux observation. Different data are severely fragmented in terms of spatial resolution, time frequency, and feature dimension, resulting in a lack of spatio-temporal continuity and scale coordination in carbon storage estimation, making it difficult to support the depiction of continuous evolution processes.

[0004] Secondly, in terms of spatial modeling, traditional carbon sink accounting methods mostly treat urban green land as isolated units for "patch-level" aggregation, failing to depict the potential carbon flow and connectivity structure between green lands. Although some research has attempted to describe green land connectivity using the corridor theory, it mostly stays at the morphological level, lacking the ability to combine physical spatial elements (such as building density, soil carbon content) with ecological process modeling (such as carbon transport resistance), and has not established a complete spatial network model.

[0005] Thirdly, in terms of dynamic prediction ability, most existing methods are static estimation models, making it difficult to respond to seasonal fluctuations in vegetation growth, extreme climate disturbances, or human interventions. Although some research has introduced machine learning methods, it mostly focuses on single-scale modeling, without constructing a unified multi-time scale carbon sink prediction system, and cannot consider carbon storage changes under short-term climate response, medium-term network regulation, and long-term urban evolution scenarios. Therefore, there is an urgent need for a new method of urban green land carbon sink accounting that has multi-source heterogeneous data fusion capability, can construct a city-level carbon flow network structure, and supports multi-scale dynamic prediction, to realize the technological leap from static estimation to dynamic perception and controllable prediction.

[0006] The patent application with the prior art publication number CN113177744A proposes a method and system for estimating the carbon sink capacity of urban green space system, which includes: constructing a multiple regression equation of the carbon sink capacity of trees and shrubs and its influencing factors such as tree height, crown width, diameter at breast height, and vertical projection area; obtaining vegetation data including tree height, diameter at breast height, crown width, and vertical projection area of trees, shrubs, and grass through sampling and field measurement; calculating the carbon sink capacity in each sample plot using existing methods; obtaining relevant parameters through scheme index query, and multiple regression analysis of measured data and sample plot carbon sink capacity; substituting the relevant parameters of different green space types into the multiple regression equation to obtain the overall planned carbon sink capacity of the urban green space system, and performing visual presentation. This scheme has a single data source, relies only on remote sensing images and a small amount of ground sample plot investigation, lacks structured fusion of multi-source heterogeneous data such as unmanned aerial vehicle laser radar and ground flux observation, and has insufficient spatiotemporal continuity. SUMMARY

[0007] To overcome the problems in the prior art of carbon sink accounting, including a single data source, not considering the carbon flux interaction and spatial network effect between green spaces, and lacking a multi-scale dynamic prediction mechanism, the present application provides a method and system for dynamic accounting of urban green space carbon sink based on multi-source data fusion and spatial network modeling.

[0008] The primary object of the present application is to solve the above technical problems. The technical solution of the present application is as follows: The first aspect of the present application provides a method for dynamic accounting of urban green space carbon sink based on multi-source data fusion and spatial network modeling, comprising the following steps: Collecting carbon sink raw data of urban green space, pre-processing the raw data to obtain pre-processed data; Performing multi-modal feature extraction and cross-modal fusion on the pre-processed data, and constructing a carbon sink data cube with spatial x time x feature structure using a node betweenness centrality weighted spatial interpolation method; Identifying green space source patches in the spatial dimension data of the carbon sink data cube; using dynamic resistance factors to construct a dynamic resistance surface, and combining circuit theory to simulate carbon flow to obtain carbon flow corridors and carbon flow density grids; Inputting multi-source data into a carbon density dynamic calibration model to output a carbon density dynamic grid; using the carbon density dynamic grid to construct a multi-scale prediction engine, and using the prediction engine to output carbon density prediction results at different time scales; Inputting the carbon flow density grid, carbon density prediction results, and urban spatial structure into a three-dimensional visualization platform combining geographic information system and building information modeling technology to generate a carbon density heat map and a carbon flow corridor evolution map, and using the heat map and the evolution map to realize dynamic accounting of carbon sink.

[0009] Further, the original data of urban green carbon sink is collected, and the original data of urban green carbon sink is preprocessed, including remote sensing image enhancement, multi-source data cleaning, coordinate repair, missing data completion, and spatial interpolation, including the following steps: The original data of urban green carbon sink is collected by the sky-ground stereo sensor network; The remote sensing image resolution is reconstructed by using the cycle-consistent generative adversarial network, the image quality and detail expression are improved by using the cycle-consistency loss, and the pseudo-high-resolution image is constructed for the low-resolution image to enhance the feature expression, so as to obtain the enhanced remote sensing image; The SCREEN algorithm is adopted, and the sliding window Z-score detection method is used to remove the pulse noise with an absolute value exceeding a preset threshold in the sensor stream data in real time; The four-quartile-buffer coupling algorithm is adopted to perform buffer topology verification and coordinate repair on the spatial anomaly of unmanned aerial vehicle positioning drift; The time series missing data is completed by using the time series generative adversarial network combined with the long short-term memory network, and the filling value similar to the real data distribution is generated through the generator and discriminator adversarial training mechanism; Combined with the Kriging interpolation method and the complex network neighbor weighting mechanism, the weighted coefficient is determined based on the node betweenness centrality, and high-precision spatial interpolation is performed.

[0010] Further, multi-modal feature extraction and cross-modal fusion are performed on the preprocessed data, and the spatial interpolation method combined with the node betweenness centrality weighting is used to construct a carbon sink data cube with a spatial-time-feature structure, including the following steps: Multi-modal features are extracted from the preprocessed remote sensing image data, GIS spatial data, sensor time series data, and hyperspectral images, respectively; The multi-modal features are extracted and compressed by convolution and max-pooling operations, and redundant information is removed, and the multi-modal features after removing the redundant information are constructed into a spatial-time correlation matrix by feature outer product, and the fusion features are obtained; The fusion features are dynamically weighted by using the Sigmoid gating mechanism, and the error contribution of each modal feature in carbon sink prediction is calculated by using the classification boosting algorithm model, and each modal feature is inversely weighted according to the error size, and the feature fusion structure parameter is optimized by using the genetic algorithm, and the optimized fusion features are obtained; Based on the optimized fusion features, the spatial grid is divided by using the geographic hash coding, and the carbon sink data cube with a spatial-time-feature three-dimensional structure is constructed by using the dynamic update engine.

[0011] Further, the pre-processed remote sensing image data, GIS spatial data, sensor time series data and hyperspectral image are respectively subjected to multi-modal feature extraction to obtain multi-modal features, including the following steps: The U-Net segmentation network is used to process the remote sensing image to extract the city green land patch boundary information features; The spatial autocorrelation index is used to analyze the GIS spatial data to obtain the aggregation degree features of the city green land in space I , and the expression is as follows:

[0012] Among them, n is the total number of spatial units; S0 is the sum of the spatial weight matrix; W ij is the spatial weight matrix, representing the adjacency relationship between region i and region j ; The attribute value of region i ; The average value of all region attribute values; The multi-scale one-dimensional convolution method is used to process the sensor time series data to extract the carbon sink data change trend features at the minute to day scale; The hyperspectral image target most sensitive spectral band is identified using the band attention mechanism to obtain the key spectral band features sensitive to the characteristics of vegetation carbon sinks.

[0013] Further, the spatial dimension data in the carbon sink data cube is subjected to green land source patch recognition, including the following steps: Based on the spatial dimension data in the carbon sink data cube, the morphological spatial pattern analysis method is used to extract the city green land core patch; The patch importance index in the landscape connectivity index is used to measure the relative contribution of each city green land core patch in the overall carbon sink connectivity network, and the expression is as follows:

[0014] Among them, PC represents the overall possible connectivity index of all patches in the landscape, PC remove，k represents the possible connectivity index of the remaining patches after removing patch k, dPC k represents the relative importance of patch k to the overall landscape connectivity; Based on the patch importance index result, the core patch is sorted, and the patch with a patch importance index value greater than a set threshold is selected as the green land source patch.

[0015] Further, taking the source patch as the starting node, a dynamic resistance surface is constructed by using a dynamic resistance factor, and a carbon flow is simulated by combining circuit theory to obtain a carbon flow corridor and a carbon flow density grid, including the following steps: The research area is divided into grid units of a uniform scale by using a dynamic resistance factor, a basic resistance value of each grid unit is calculated, and a resistance surface is constructed, the dynamic resistance factor including one or more of the following: building density, road grade, soil organic carbon content, and human activity intensity; A parameter optimization method is used to adjust the model parameters in the resistance surface to construct a dynamic resistance surface of urban carbon flow; A carbon flow is simulated by using circuit theory in combination with the dynamic resistance surface, taking the source patch as a carbon flow injection node to obtain a carbon flow flux distribution result, and the carbon flow flux calculation expression is as follows:

[0016] Wherein, V xy is the carbon potential difference between patches x and y, R xy is the cumulative resistance of the dynamic resistance surface, I xy is the carbon flow flux; A carbon flow density grid map is output by using the carbon flow flux distribution result, and a carbon flow corridor with high transmission efficiency is identified from the carbon flow density grid map.

[0017] Further, the parameter optimization method is a genetic algorithm, a response surface method, or a particle swarm optimization.

[0018] Further, the multi-source data includes one or more of the following: normalized vegetation index data, laser radar inverted aboveground biomass data, meteorological data, and vegetation radial growth, the multi-source data is input into a carbon density dynamic calibration model, and the expression of the calibration model is as follows:

[0019] Wherein, represents the carbon storage of ecosystem type i at time t; is a reference value, representing the biomass inverted by laser radar multiplied by the carbon content coefficient; is a remote sensing correction factor, represents the normalized vegetation index value corresponding to the same grid at time t, represents the normalized vegetation index value at the reference time, used to correct the difference in normalized vegetation index at different time points to improve the accuracy of carbon storage estimation; is the radial growth, representing the influence of the radial expansion of vegetation on carbon storage over time; For the microclimate correction factor, the effects of temperature T and humidity H on vegetation growth and carbon storage are considered, and the expression is as follows: .

[0021] Further, a multi-scale prediction engine is constructed using the carbon density dynamic grid, and the prediction engine outputs carbon density prediction results at different time scales, including the following steps: The continuous time series slice of the past 2n days in the carbon density dynamic grid is input into the ConvLSTM model, and the input is a 2n*m-dimensional spatiotemporal sequence data, and the attention gate mechanism is used to focus on the spatial units in the carbon density grid that exhibit abnormal temporal variation characteristics, to obtain high-frequency spatiotemporal fluctuation characteristics, and the high-frequency spatiotemporal fluctuation characteristics are used to predict the carbon density value in the future 0 to n days; The carbon flow network topology index is calculated using the carbon flow corridor and the carbon flow density grid, the carbon flow network topology index and the historical carbon density data are input into the graph neural network model, the Node2Vec algorithm is used to embed the nodes in the carbon flow network into a low-dimensional space, and the network features and adjacency relationships are captured, to obtain the vector representation of the nodes; the node vector representation and the historical carbon density features of the corresponding nodes are combined with the graph convolution network to realize carbon flow propagation modeling, and a carbon flow propagation model is obtained; A knowledge constraint loss function is constructed by introducing a photosynthesis mechanism into the carbon flow propagation model, and the expression of the knowledge constraint loss function is as follows:

[0022] wherein, and are weight coefficients of the loss function, is the predicted carbon flow value of the model, is the carbon flow value output by the photosynthesis mechanism model, and MSE is the mean square error; The knowledge constraint loss function is used to train the model by combining the mean square error and the absolute error between the model calculated value and the real historical observed value, to suppress the overfitting risk of the model to the data bias, and output the carbon density prediction value of the main hub node in the future 1 to p months; The urban spatial structure, built-up area range, green land distribution, building density, and plant physiological parameters monitored by sensors are input into the Transformer model embedded with the Farquhar photosynthesis model, and the carbon density trend prediction value and carbon sink saturation year limit judgment result at the scale of 1 to q years are output.

[0023] The second aspect of the application provides a city green land carbon sink dynamic accounting system based on multi-source data fusion and spatial network modeling, which is used to realize the steps of the city green land carbon sink dynamic accounting method based on multi-source data fusion and spatial network modeling, and the system comprises: A data perception module is configured to collect original carbon sink data of urban green land, pre-process the original data, and output pre-processed data; A fusion calculation module is configured to perform multi-modal feature extraction and cross-modal fusion on the pre-processed data output by the data perception module, combine node betweenness centrality weighted spatial interpolation methods, and output a carbon sink data cube with a spatial-time-feature structure. A dynamic modeling module is configured to identify green land source patches in the spatial dimension data of the carbon sink data cube output by the fusion calculation module, use the source patches as starting nodes, construct a dynamic resistance surface using a dynamic resistance factor, simulate carbon flow based on circuit theory, output carbon flow corridors and carbon flow density grids, input multi-source data into a carbon density dynamic calibration model, output a carbon density dynamic grid, and construct a multi-scale prediction engine using the carbon density dynamic grid to output carbon density prediction results at different time scales. An application service module is configured to combine the carbon flow density grid and the carbon density prediction results output by the dynamic modeling module with urban spatial structure input into a three-dimensional visualization platform based on geographic information system (GIS) and building information modeling (BIM) technology, generate a carbon density heat map and a carbon flow corridor evolution map, and realize dynamic accounting of carbon sinks using the heat map and the evolution map.

[0024] Compared with the prior art, the technical scheme of the present application has the following advantages: The present application constructs a carbon sink data cube with a spatial-time-feature structure, realizes standardized fusion and structured expression of multi-source heterogeneous data such as satellite remote sensing, unmanned aerial vehicle laser radar, ground flux tower, and meteorological monitoring, improves the data precision and spatio-temporal continuity of carbon sink accounting, and provides a unified data support framework for dynamic estimation; a carbon flow spatial network is constructed using circuit theory, the interaction and spatial network effect of carbon flux between urban green lands are depicted based on simulation of electronic flow, carbon flow corridors and carbon flow density grids are output, the connectivity and transmission efficiency of the green land system are quantitatively reflected, high-flux corridors and carbon sink hub nodes are identified, and precise structural analysis basis is provided for urban green space pattern optimization and low-carbon planning; a multi-scale prediction engine is constructed, which can respectively cope with carbon density change prediction at the hour, month, and year levels, significantly improve the adaptability and foresight of the system to carbon sink trends under multiple scenarios such as climate fluctuations, seasonal evolution, and urban expansion, and make the prediction results more credible. BRIEF DESCRIPTION OF DRAWINGS

[0025] In order to make the purpose and technical scheme of the present application clearer, the present application provides the following drawings and makes the following description: Figure 1 A flowchart of a method for dynamic accounting of urban green land carbon sinks based on multi-source data fusion and spatial network modeling is provided for the embodiments of the present application. Figure 2 A schematic diagram of unmanned aerial vehicle and ground sensor position offset calibration is provided for an embodiment of the present application. Figure 3 A structural diagram of a city green land carbon sink dynamic accounting system based on multi-source data fusion and spatial network modeling is provided for an embodiment of the present application. DETAILED DESCRIPTION

[0026] In order to more clearly understand the above-mentioned purposes, features and advantages of the present application, the present application will be further described in detail below in combination with the drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0027] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application, however, the present application can also be implemented in other ways different from those described herein, therefore, the scope of protection of the present application is not limited by the specific embodiments disclosed below.

[0028] Embodiment 1: The present application provides a city green land carbon sink dynamic accounting method based on multi-source data fusion and spatial network modeling, as shown in Figure 1 As shown in the flowchart of the city green land carbon sink dynamic accounting method based on multi-source data fusion and spatial network modeling, the specific steps are as follows: S1: Collecting city green land carbon sink original data, pre-processing the original data to obtain pre-processed data.

[0029] More specifically, in the process of collecting city green land carbon sink original data, the city green land carbon sink original data is collected by using a sky-ground three-dimensional sensing network.

[0030] The sky-ground three-dimensional sensing network collects carbon sink related parameters and constructs a data base. Using the multi-source data acquisition and processing technology of "sky-ground-air" integration, the key parameters related to carbon sink in the ecological system are obtained, and they are integrated into a unified data base to support the development of carbon sink monitoring, accounting, evaluation and management. This process involves the collaborative application of various technical means, including satellite remote sensing, unmanned aerial vehicle remote sensing, ground sensors, Internet of Things, etc., forming a carbon sink monitoring system covering a wide range, high precision and multiple dimensions.

[0031] Satellite remote sensing is the space-based part, which mainly obtains large-scale, periodic and high-resolution carbon sink related data through satellite observation. For example, satellites can monitor the vegetation index, vegetation coverage and other parameters of forest, grassland, wetland and other ecosystems. The advantage of satellite remote sensing is that it has wide coverage, long time series and high spatial resolution, which can make up for the shortcomings of ground monitoring stations.

[0032] The space-based part mainly uses unmanned aerial vehicles (UAVs) equipped with LiDAR, hyperspectral imagers, and other sensors to obtain high-precision vegetation structural parameters such as tree height and crown width, as well as leaf nitrogen content and vegetation information. UAVs can flexibly monitor complex terrains, making up for the shortcomings of satellite remote sensing in local details and dynamic changes.

[0033] The ground-based part mainly relies on ground sensors, carbon flux towers, weather stations, and other equipment to obtain high-precision carbon sink data at the sample plot level. For example, ground-based carbon flux observation towers can monitor the carbon exchange flux of forests or grasslands in real time, providing dynamic data on carbon absorption and release. In addition, ground sensors can also monitor key parameters such as soil carbon storage and vegetation biomass.

[0034] Using satellite data, we mainly use a combination of Sentinel-2 (10m resolution) and Gaofen-5 (spectral resolution 5nm) satellites to cover the entire city daily, extract NDVI, chlorophyll fluorescence, and other vegetation index data, which is a combination of high spatial resolution and high spectral resolution remote sensing technology for urban vegetation monitoring and environmental assessment. Through multi-satellite remote sensing data assimilation technology, we fuse data from different overpasses to generate continuous observation sequence data with a time resolution of 6 hours.

[0035] The space-based data uses autonomous cruise UAVs equipped with hyperspectral imagers (400-2500nm) and LiDAR, with a point cloud density of >200 points / m 2 , accurately extracting structural parameters such as tree height and crown width; hyperspectral data can be used to estimate leaf nitrogen content (accuracy >90%), which is directly related to plant carbon sequestration efficiency. The deployment strategy of UAVs is based on a complex network centrality algorithm, which identifies carbon sink hub patches (such as urban forest core areas) and dynamically adjusts the cruise path, combining complex network theory with ecological network analysis, traffic optimization, and other multi-disciplinary knowledge.

[0036] The complex network centrality algorithm is used to identify important nodes in the network, which have high influence or control in the network. This is mainly achieved through degree centrality, betweenness centrality, and closeness centrality.

[0037] The high-degree centrality node is usually connected to multiple other nodes, indicating that it has a high functional connection in the ecological network. The high-degree centrality expression is as follows:

[0038] where k represents the number of existing edges connected to node i, and N-1 represents the number of edges connecting node i to other nodes.

[0039] High betweenness centrality nodes usually act as "bridges" in the network, connecting multiple sub-networks, indicating their intermediary role in carbon emission and carbon absorption processes. The high betweenness centrality expression is as follows:

[0040] wherein, represents the number of paths passing through node i and being the shortest path; represents the number of shortest paths connecting s and t.

[0041] High closeness centrality nodes usually have a shorter average path length, indicating that they have higher accessibility in the network and are key nodes in the carbon emission and carbon absorption processes. The high closeness centrality expression is as follows:

[0042]

[0043] wherein, represents the average distance from node i to the remaining points, and the inverse of the average distance is the closeness centrality.

[0044] The complex network centrality algorithm is used to calculate high centrality nodes, high betweenness centrality nodes and high closeness centrality nodes, so as to identify key hub patches and dynamically adjust the cruise path of the unmanned aerial vehicle, ensuring efficient execution of the cruise task and maximizing the coverage of carbon sink hub patches, thereby improving the accuracy of carbon sink monitoring and management.

[0045] The ground Internet of Things sensor network includes soil multi-parameter sensors, photosynthetically active radiation meters, CO2 flux towers, etc., for collecting CO2 flux, photosynthetically active radiation (PAR), soil temperature and humidity, etc.

[0046] Among them, the edge computing node is embedded in the sensor, TensorFlow Lite is introduced to run the machine learning model on the edge device, reducing the dependence on the cloud, and real-time monitoring of abnormal data. Sparse sensing technology is adopted, which utilizes the sparsity of signals to reduce the number of data samples to reduce the sensing cost, storage demand and bandwidth requirement. The original data is compressed to 10% of the original volume before uploading, saving bandwidth by more than 70%.

[0047] As shown in Figure 2 , the position offset between the unmanned aerial vehicle and the ground sensor is calibrated by UWB ultra-wideband positioning (accuracy ±10 cm), and the target position is calculated by measuring the time of flight (TOF) or time difference of arrival (TDOA) of the signal, thereby achieving centimeter-level positioning accuracy between the unmanned aerial vehicle and the ground sensor.

[0048] In the sensor deployment process, the sensor is deployed flexibly based on the importance of the node. Specifically, based on the complex network node betweenness centrality, the topological importance of green space patches in ecological corridors is calculated, and key nodes are preferentially deployed. Patches with a centrality > 0.3 are deployed with laser radar + soil sensor dual-mode equipment, and secondary areas only need single-mode equipment.

[0049] In the sensor re-embedded federated learning client, without uploading the original data, the carbon sink feature extraction model is optimized collaboratively; in the damaged device area, the alternative data is generated through the LSTM virtual sensor model; through the blockchain, the device credit is empowered to ensure the credibility of the carbon sink data.

[0050] The preprocessing of the original data of urban green space carbon sink includes remote sensing image enhancement, multi-source data cleaning, coordinate repair, missing data completion, and spatial interpolation, which includes the following steps: The CycleGAN is used for remote sensing image resolution reconstruction, the cycle consistency loss is used to improve the image quality and detail expression, and the pseudo high-resolution image is constructed for low-resolution image to enhance the feature expression, so as to obtain the enhanced remote sensing image. In order to ensure the consistency of the generated image between the two domains, the CycleGAN introduces the cycle consistency loss (Cycle Consistency Loss), which is expressed as follows:

[0051] Where x and y are the images of the source domain and the target domain respectively. This loss function ensures that the generated image does not lose key information during the conversion process.

[0052] The specific process is to use low-resolution 10m satellite images as input, which usually contain less detailed information; the generator learns the mapping relationship between low-resolution images and high-resolution images, and converts the input low-resolution images into high-resolution pseudo images. This process uses the mechanism of generative adversarial network, and the discriminator is used to evaluate the authenticity of the generated image; in order to ensure that the generated high-resolution image can be consistent with the original image when converted back to low-resolution image, the CycleGAN introduces the cycle consistency loss, which makes the generator not only pay attention to the visual effect of the image when generating high-resolution image, but also ensure that it can match the original image when converted back to low-resolution image; finally, the high-resolution pseudo image generated by the generator can be used as a substitute for high-resolution image for subsequent remote sensing analysis, target detection and other tasks.

[0053] The SCREEN algorithm is used, and the sliding window Z-score detection method is used to remove the absolute value of the sensor stream data that exceeds the preset threshold (in this embodiment, the preset threshold is ±3 The pulse noise of the Z-score standardization is removed to reduce the interference of noise on subsequent analysis. Among them, the Z-score standardization is a commonly used data preprocessing method, which is used to convert the data into a standard normal distribution with a mean of 0 and a standard deviation of 1. Its calculation formula is:

[0054] Where X is the original data point, is the mean of the data in the window, is the standard deviation of the data in the window.

[0055] Set a threshold (±3 ), when the absolute value of the Z-score of the data point exceeds the threshold, it is considered that the data point is an outlier. The data points marked as outliers are removed from the data set to reduce the impact of noise on subsequent analysis.

[0056] The four-quartile-buffer coupling algorithm is used to perform buffer topology verification and coordinate repair on the spatial anomaly of unmanned aerial vehicle positioning drift. The four-quartile statistical method is used to identify abnormal values or drift points in the data, and the buffer analysis technology is used to perform topology verification and coordinate repair on these abnormal points, thereby improving the accuracy and integrity of the spatial data; The time series missing value completion is performed by using the time series generative adversarial network combined with the long short-term memory network. The generator and the discriminator are trained in an adversarial training mechanism to generate filling values similar to the real data distribution. The LSTM time series generative adversarial network (LSTM-GAN) combines the time series modeling capability of the long short-term memory network (LSTM) and the adversarial training mechanism of the generative adversarial network (GAN), and is used to generate high-quality and realistic time series data. Through the adversarial training mechanism between the generator and the discriminator, time series data similar to the real data distribution can be generated, and it performs well in time series data missing value filling. The average absolute error (MAE) is usually less than 0.05, and the expression of the average absolute error is as follows:

[0057] Where, is the true value, is the generated filling value, and n is the number of samples. In practical applications, the MAE of LSTM-GAN is usually less than 0.05, indicating that its filling effect is very close to the real data.

[0058] Fourth, combined with the Kriging interpolation method and the complex network neighbor weighting mechanism, the weighted coefficient is determined based on the node betweenness centrality, and high-precision spatial interpolation is performed. It mainly includes three steps: ① Spatial interpolation is performed by Kriging interpolation method:

[0059] wherein, is the interpolation value at the position point , is the observation value of the known point , is the weight coefficient, is the terrain elevation at the position point , determined by the following conditions:

[0060]

[0061] wherein, is the semi-variogram model, describing the spatial autocorrelation between points and .

[0062] Newly added , solve the problem of radiation distortion caused by urban terrain, wherein is usually related to the height of the terrain, which can be expressed as:

[0063] wherein, is the solar incident angle. By multiplying the radiation value of each pixel of the image by , the brightness change caused by the terrain undulation can be corrected, and the geometric accuracy and interpretability of the image can be improved.

[0064] ②The weights are weighted and adjusted by using the betweenness centrality of the nodes in the complex network to improve the accuracy and rationality of the interpolation results.

[0065] ③Weight adjustment, in Kriging interpolation, the weight coefficient is usually determined by the semi-variogram model. However, in this method, the weight not only considers the spatial autocorrelation, but also considers the importance of the node in the network. Specifically, the weight of each node will be adjusted according to its betweenness centrality, that is:

[0066] wherein, is the normalized weight of node i. The normalization method can be simple linear normalization or normalization based on maximum and minimum value to ensure that the total weight is 1.

[0067] Further, a data cleaning method based on dynamic rule configuration is established, and the rule base architecture includes logical rules, statistical planning and semantic rules.

[0068] S2: Multimodal feature extraction and cross-modal fusion are performed on the preprocessed data, combined with node betweenness centrality weighted spatial interpolation method, to construct a carbon sink data cube with space-time-feature structure, including the following steps: Multimodal feature extraction is performed on the preprocessed remote sensing image data, GIS spatial data, sensor time series data, and hyperspectral image, respectively, to obtain multimodal features, including the following steps: The U-Net segmentation network is used to process remote sensing images to extract urban green space patch boundary information features, which are used to represent spatial range and morphological structure. Moran's I spatial autocorrelation index is used to analyze GIS spatial data to obtain the aggregation degree of urban green space in space I , which is expressed as follows:

[0069] wherein, n is the total number of spatial units; S0 is the sum of the spatial weight matrix; W ij is the spatial weight matrix, representing the adjacency relationship between region i and region j ; is the attribute value of region i ; is the average value of all region attribute values; Through Moran's I analysis, the spatial distribution pattern of urban green space is identified, such as whether it presents high aggregation, low aggregation, or random distribution.

[0070] A multi-scale one-dimensional convolution method is used to process sensor time series data to extract carbon sink data change trend features at minute to day scales. This method extracts different levels of features from data through different scale convolution kernels, thereby better capturing multi-scale changes in time series.

[0071] The band attention mechanism is used to identify the most sensitive spectral bands of hyperspectral image targets (such as chlorophyll content or fluorescence features), obtaining key spectral band features sensitive to vegetation carbon sink characteristics. This method analyzes the importance of different bands in hyperspectral images to reduce redundant information and improve the efficiency and accuracy of subsequent analysis.

[0072] Convolution (Conv) and max pooling (MaxPooling) operations are used to extract and compress multimodal features and remove redundant information. Convolution operation is used to extract local features, while max pooling operation is used to reduce the spatial dimension of feature maps while retaining key information. This combination can effectively reduce the amount of calculation and improve the generalization ability of the model. The multimodal features after removing redundant information are constructed into a space-time correlation matrix through feature outer product, obtaining fusion features.

[0073] The Sigmoid gating mechanism is used to dynamically allocate weights to the fused features. The purpose is to filter out the most useful cross-modal information for the task, thereby improving the performance of the model. The Sigmoid function can map the features to the range of 0~1 as weights to measure the importance of the features, thereby achieving dynamic control of the features.

[0074] The error contribution of each modality feature in carbon sink prediction is calculated using the Classification Algorithm based on Contribution (CATBoost) model, and each modality feature is inversely weighted according to the error size. The genetic algorithm is used to optimize the feature fusion structure parameters, improve the fusion precision and generalization ability, and obtain the optimized fusion features. The training includes the following steps: ①First, input the single modality features (such as remote sensing images, sensor data, etc.) into the base model. For example, remote sensing images may be input into a random forest model, while sensor data may be input into an LSTM model. These models process different types of features and output prediction results.

[0075] ②For each modality, calculate the error between its prediction result and the actual value. The error calculation method can be mean square error (MSE) or other appropriate loss function. The smaller the error, the better the prediction effect of the modality.

[0076] ③According to the prediction error of each modality, assign the corresponding weight. The weight calculation formula is:

[0077] where E i is the prediction error of the i-th modality, and W i is the weight of the modality. The higher the weight, the better the prediction effect of the modality, so the contribution of the modality in the final weighted prediction is greater.

[0078] Genetic algorithm is used to tune parameters and edge-cloud collaborative computing method is used for multi-modal data fusion and optimization.

[0079] The genetic algorithm is used to optimize the feature fusion dimension of the carbon sink prediction model, aiming to minimize the fusion feature dimension while maximizing the carbon sink prediction accuracy. This method has a significant advantage in parameter tuning of the carbon sink prediction model, which can effectively balance the goal of minimizing feature dimension and maximizing prediction accuracy. Through reasonable gene coding, fitness function design and parameter optimization, the genetic algorithm can find the optimal combination of model parameters, thereby improving the performance and efficiency of the model.

[0080] Edge-cloud collaborative computing leverages the strengths of both edge and cloud computing to achieve efficient data processing and intelligent decision-making. In this model, the edge is responsible for lightweight feature fusion, while the cloud handles deep fusion optimization, enabling global and local collaborative optimization.

[0081] The main task of the edge is real-time data processing and lightweight feature extraction, using MobileNetV3 for feature extraction and preliminary processing. The main task of the cloud is global data processing, model training, and deep learning optimization. In edge-cloud collaborative computing, the cloud can further optimize the features extracted by the edge. The cloud can also use federated learning (Federated Learning) to perform distributed training on multiple edge devices, protecting data privacy while improving the model's generalization ability.

[0082] Based on the optimized fusion features, a spatial grid is divided using Geohash encoding, and a dynamic update engine is used to construct a spatial-time-feature three-dimensional structure carbon sink data cube. The carbon sink data cube is divided into spatial grids using Geohash encoding, and continuous observation slices are constructed based on a sliding window mechanism in the time dimension. The feature dimension includes vegetation physiological index (12 dimensions), soil parameters (8 dimensions), and meteorological factors (6 dimensions), which are used to support subsequent carbon flow modeling and carbon sink dynamic prediction tasks.

[0083] The dynamic update engine is a system architecture that combines incremental update mechanisms and version management, aiming to improve data processing efficiency, reduce computing resource consumption, and enhance data traceability and security. Its core idea is to update only the changed part of the data (incremental update) and use technologies such as blockchain to evidence and manage data versions, achieving efficient and secure data updating and auditing.

[0084] S3: Identify green space source patches in the spatial dimension data of the carbon sink data cube using the MSPA method. Use the source patch as the starting node and construct a dynamic resistance surface using dynamic resistance factors. Simulate carbon flow using circuit theory to obtain carbon flow corridors and carbon flow density grids.

[0085] More specifically, identifying green space source patches in the spatial dimension data of the carbon sink data cube using the MSPA method includes the following steps: Based on the spatial dimension data in the carbon sink data cube, the morphological spatial pattern analysis (MSPA) method is used to extract urban green space core patches. The core patch is a green area with continuous distribution, area compliance, and high vegetation coverage, which is used as a candidate unit for subsequent carbon flow simulation. The patch importance index in the landscape connectivity index is used to measure the relative contribution of each urban green space core patch to the overall carbon sink connectivity network, and the expression is as follows:

[0086] Wherein, PC represents the overall probability of connectivity index of all patches in the landscape (Probability of Connectivity), PC remove，k represents the probability of connectivity index of the remaining patches after removing patch k, dPC k represents the relative importance of patch k to the overall landscape connectivity; Based on the patch importance index results, the core patches are sorted, and the patches with patch importance index values greater than a set threshold (in this embodiment, the threshold is set to 10%) are selected as green source patches, which are used for subsequent resistance surface construction and starting node setting in carbon flow corridor simulation.

[0087] Taking the source patch as the starting node, a dynamic resistance surface is constructed using a dynamic resistance factor, and carbon flow is simulated using circuit theory to obtain carbon flow corridors and carbon flow density grids, including the following steps: The research area is divided into uniform scale grid cells using a dynamic resistance factor, and the basic resistance value of each grid cell is calculated and a resistance surface is constructed, the dynamic resistance factor including one or more of the following: building density, road grade, soil organic carbon content, human activity intensity.

[0088] The model parameters (weight coefficients, integration range, parameter combination, etc.) in the resistance surface are adjusted using a parameter optimization method to optimize the modeling accuracy of the resistance surface, improve the rationality and adaptability of carbon flow simulation, and construct a dynamic resistance surface of urban carbon flow, wherein the parameter optimization method is genetic algorithm, response surface method (RSM) or particle swarm optimization (PSO).

[0089] Carbon flow is simulated using circuit theory (Circuit Theory) combined with the dynamic resistance surface with the source patch as the carbon flow injection node, analogous to the transmission process of electronic current in a resistance network, to obtain carbon flow flux distribution results, and the carbon flow flux calculation expression is as follows:

[0090] Wherein, V xy is the carbon potential difference between patches x and y, R xy is the cumulative resistance of the dynamic resistance surface, I xy is the carbon flow flux; The carbon flow density grid map is output by using the carbon flow flux distribution result, and a carbon flow corridor with high transmission efficiency is identified from the carbon flow density grid map, which is used to construct an efficient carbon connection network structure between urban green spaces.

[0091] Further, the topological indexes such as the node betweenness centrality and the network vulnerability index are calculated, and a dynamic reconnection mechanism is established. Specifically, the fragmentation degree of the green space landscape is monitored in real time, and the structure and connection mode of the green space system are automatically adjusted according to the preset threshold value, so as to restore or improve the ecological function and connectivity of the green space.

[0092] S4: inputting the multi-source data into the carbon density dynamic calibration model to output a carbon density dynamic grid; constructing a multi-scale prediction engine by using the carbon density dynamic grid, and outputting a carbon density prediction result at different time scales by using the prediction engine.

[0093] More specifically, the multi-source data includes one or more of the following: normalized vegetation index data (NDVI), above-ground biomass data inverted by laser radar (LIDAR), meteorological data (such as temperature and humidity), and vegetation radial growth, the multi-source data is input into the carbon density dynamic calibration model, and the expression of the calibration model is as follows:

[0094] wherein, represents the carbon storage of the ecosystem type i at time t; is a baseline value, which represents the biomass inverted by LIDAR multiplied by the carbon content coefficient (0.5 for trees and 0.45 for shrubs); is a remote sensing correction factor, represents the normalized vegetation index value corresponding to the same grid at time t, represents the normalized vegetation index value at the baseline time, which can be selected as the typical average value / maximum value of vegetation growth, and is used to correct the difference of the normalized vegetation index (NDVI) at different time points, so as to improve the accuracy of the carbon storage estimation; is the radial growth, which represents the influence of the radial expansion of vegetation on the carbon storage over time; is a microclimate correction factor, which considers the influence of temperature T and humidity H on vegetation growth and carbon storage, and the expression is as follows: .

[0096] Further, considering the influence of extreme climate, an abnormal event compensation mechanism is introduced, including the response to extreme climate, insect pest disturbance, etc., to output the corrected carbon density grid data, in this embodiment, the spatial resolution of the carbon density grid data is 1m, and the time resolution is 1h.

[0097] A multi-scale prediction engine is constructed by using a carbon density dynamic grid, and a carbon density prediction result in different time scales is output by using the prediction engine, including the following steps: The continuous time slice of the past 2n (in the embodiment, n is 7) days in the carbon density dynamic grid is input into the ConvLSTM model, the input is a 2n*m (in the embodiment, m is 26) dimensional spatio-temporal sequence data, the attention gate mechanism is used to focus on the spatial units in the carbon density grid that show abnormal time variation characteristics, and high-frequency spatio-temporal fluctuation characteristics are obtained, the high-frequency spatio-temporal fluctuation characteristics are used to predict the carbon density value from 0 to n days in the future, the prediction time step is 1 hour, and the prediction time step is used to respond to microclimate disturbance and short-term carbon sink fluctuation. The carbon flow network topology indicators such as node betweenness centrality and network vulnerability index are calculated by using the carbon flow corridor and the carbon flow density grid, the carbon flow network topology indicators and the historical carbon density data are input into the graph neural network model, the Node2Vec algorithm is used to embed the nodes in the carbon flow network in a low dimension, the network features and adjacency relationships are captured, and the vector representation of the nodes is obtained; the node vector representation and the historical carbon density features of the corresponding nodes are combined with the graph convolution network GCN to realize carbon flow propagation modeling, the message passing mechanism of the GCN allows the nodes to receive information from their neighbor nodes and aggregate these information into their own feature representation, and the carbon flow propagation model is obtained; the knowledge constraint loss function is constructed by introducing the photosynthesis mechanism in the carbon flow propagation model; the mean square error (MSE) and the absolute error (L1) between the model calculation value and the real historical observation value are combined with the knowledge constraint loss function to suppress the overfitting risk of the model to data bias, and the carbon density prediction value of the main hub node from 1 to p (in the embodiment, the value of p is 3) months in the future is output, wherein the expression of the knowledge constraint loss function is as follows:

[0098] wherein, and are weight coefficients of the loss function, in the embodiment, the value of is 0.65, the value of is 0.35, is the carbon flow value predicted by the model, is the carbon flow value output by the photosynthesis mechanism model, and MSE is the mean square error. The urban spatial structure, built-up area range, green space distribution, building density, and sensor-monitored plant physiological parameters (such as leaf nitrogen content, maximum photosynthetic rate, light saturation point, etc.) are input into the Transformer model embedded with the Farquhar photosynthesis model. Since the Transformer model itself is based on the self-attention mechanism, which is inherently position-invariant, it is necessary to introduce the sequential information in the sequence through position encoding. The position encoding layer uses sine and cosine functions to represent the periodic changes of different positions, thereby helping the model better capture the positional relationships in the sequence. This makes the model not only focus on the absolute position of the time series, but also consider the periodic changes of carbon sinks in different seasons (such as plant growth in spring and enhanced transpiration in summer, etc.).

[0099] In the feedforward layer, the Farquhar model is used to simulate the photosynthetic rate mechanism. This model is a classic leaf-scale photosynthesis model that describes the photosynthetic rate of plants under different CO2 concentrations. By embedding the Farquhar model into the feedforward layer of the Transformer, the carbon sink capacity under different urban expansion scenarios can be simulated based on the characteristics of urban expansion and vegetation coverage. The output is the carbon sink saturation state. For example, when urban expansion leads to a decrease in green space and an increase in buildings, the photosynthetic rate of plants will decrease, thereby affecting the carbon sink capacity. To adjust the predictive behavior of the model under different carbon sink states (such as carbon saturation or non-saturation), a dynamic weight adjustment mechanism is introduced. When the carbon sink system is detected to be in a carbon saturation state, the weight of the mechanism module is increased. Specifically, when the carbon saturation state is detected (carbon_saturation > threshold), the model automatically increases the weight of the mechanism model to ensure that under carbon saturation conditions, the model relies more on the physical laws of carbon cycle processes for prediction. Using the adjusted model to predict the carbon sink saturation period, the carbon density trend prediction values and carbon sink saturation period judgment results are output for 1 to q (in this embodiment, the value of q is 5) years.

[0100] S5: Combine the carbon flux density grid, carbon density prediction results, and urban spatial structure input into a three-dimensional visualization platform that integrates geographic information system (GIS) and building information modeling (BIM) technologies. Use WebGL technology for grid rendering to generate carbon density heat maps and carbon flow corridor evolution maps. Use the heat maps and evolution maps to realize dynamic accounting and visual display of carbon sinks.

[0101] Furthermore, a multi-objective system is introduced, including green energy planning, ecological carbon sequestration planning, and transportation carbon reduction planning. Combined with the demand for carbon asset life cycle management and risk response, a multi-agent model is constructed to deduce and verify the mechanism and build an intelligent decision engine to realize intelligent decision support.

[0102] Embodiment 2: The embodiment provides a city green land carbon sink dynamic accounting system based on multi-source data fusion and spatial network modeling, which is used for implementing the city green land carbon sink dynamic accounting method based on multi-source data fusion and spatial network modeling in the embodiment 1, as shown in the following formula: Figure 3 As shown in the formula, the system comprises: A data sensing module is configured to collect city green land carbon sink original data, pre-process the original data, and output pre-processed data. A fusion calculation module is configured to perform multi-modal feature extraction and cross-modal fusion on the pre-processed data output by the data sensing module, combine a node betweeness centrality weighted spatial interpolation method, and output a carbon sink data cube with spatial, time and feature structure. A dynamic modeling module is configured to perform green land source patch recognition on the spatial dimension data in the carbon sink data cube output by the fusion calculation module, use the source patch as a starting node, construct a dynamic resistance surface by using a dynamic resistance factor, simulate carbon flow by combining a circuit theory, output carbon flow corridors and carbon flow density grids, input multi-source data into a carbon density dynamic calibration model, output carbon density dynamic grids, construct a multi-scale prediction engine by using the carbon density dynamic grids, and output carbon density prediction results at different time scales by using the prediction engine. An application service module is configured to combine the carbon flow density grids and the carbon density prediction results output by the dynamic modeling module with city space structure input into a three-dimensional visualization platform of geographic information system (GIS) and building information model (BIM) technology, generate a carbon density heat map and a carbon flow corridor evolution map, and realize carbon sink dynamic accounting and visual display by using the heat map and the evolution map.

[0103] Obviously, the above embodiments of the present application are only examples for clearly illustrating the present application, and are not intended to limit the implementation modes of the present application. Based on the above description, other different forms of changes or variations can be made by those skilled in the art. Here, all the implementation modes are not required or can not be exhausted. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the claims of the present application.

Claims

1. A dynamic accounting method for urban green space carbon sequestration based on multi-source data fusion and spatial network modeling, characterized in that, Includes the following steps: Collect raw data on carbon sequestration in urban green spaces, and preprocess the raw data to obtain preprocessed data; Multimodal feature extraction and cross-modal fusion are performed on the preprocessed data, and a spatial interpolation method with node betweenness centrality weighted is used to construct a carbon sink data cube with a spatial × temporal × feature structure. Green source patches are identified in the spatial dimension data of the carbon sink data cube; with the source patches as the starting nodes, dynamic resistance surfaces are constructed using dynamic resistance factors, and carbon flow is simulated by circuit theory to obtain carbon flow corridors and carbon flow density grids. Multi-source data is input into the carbon density dynamic calibration model, and a carbon density dynamic raster is output. The carbon density dynamic raster is used to construct a multi-scale prediction engine, and the prediction engine is used to output carbon density prediction results at different time scales. By combining the carbon flow density grid, carbon density prediction results, and urban spatial structure input with a three-dimensional visualization platform based on geographic information system and building information modeling technology, a carbon density heat map and a carbon flow corridor evolution map are generated. The heat map and the evolution map are then used to realize dynamic carbon sink accounting.

2. The method for dynamic accounting of urban green space carbon sequestration based on multi-source data fusion and spatial network modeling as described in claim 1, characterized in that, The process involves collecting raw data on urban green space carbon sequestration and preprocessing it, including remote sensing image enhancement, multi-source data cleaning, coordinate restoration, missing data completion, and spatial interpolation. This includes the following steps: Collect raw data on carbon sequestration in urban green spaces using a three-dimensional sky-ground sensor network; Remote sensing image resolution reconstruction is performed using a cycle consistency generative adversarial network. The cycle consistency loss is used to improve image quality and detail representation. Pseudo-high resolution images are constructed from low-resolution images to enhance feature representation, resulting in enhanced remote sensing images. The SCREEN algorithm is used to remove impulse noise in the sensor stream data whose absolute value exceeds a preset threshold in real time based on the sliding window Z-score detection method; The quartile-buffer coupling algorithm is used to perform buffer topology verification and coordinate repair for UAV positioning drift spatial anomalies; Temporal generative adversarial networks combined with long short-term memory networks are used to complete missing time series data. Imputation values ​​similar to the distribution of real data are generated through an adversarial training mechanism between the generator and the discriminator. By combining the Kriging interpolation method with the neighbor weighting mechanism of complex networks, weighting coefficients are determined based on node betweenness centrality to perform high-precision spatial interpolation.

3. The method for dynamic accounting of urban green space carbon sequestration based on multi-source data fusion and spatial network modeling as described in claim 1, characterized in that, Multimodal feature extraction and cross-modal fusion are performed on the preprocessed data. A spatial × temporal × feature structure carbon sink data cube is constructed by combining a node betweenness centrality-weighted spatial interpolation method, including the following steps: Multimodal features were extracted from the preprocessed remote sensing image data, GIS spatial data, sensor time-series data, and hyperspectral images to obtain multimodal features; Multimodal features are extracted and compressed and redundant information is removed by convolution and max pooling operations. The multimodal features after removing redundant information are used to construct a spatial-temporal correlation matrix by feature outer product to obtain fused features. The Sigmoid gating mechanism is used to dynamically assign weights to the fused features, and the classification boosting algorithm model is used to calculate the error contribution of each modality feature in carbon sink prediction. Based on the error magnitude, each modality feature is back-weighted, and the feature fusion structure parameters are optimized by combining a genetic algorithm to obtain the optimized fused features. Based on the optimized fusion features, geohashing encoding is used for spatial grid division, and a dynamic update engine is used to construct a spatiotemporally continuous, dimension-aligned three-dimensional carbon sink data cube with a spatial × temporal × feature structure.

4. The method for dynamic accounting of urban green space carbon sequestration based on multi-source data fusion and spatial network modeling according to claim 3, characterized in that, Multimodal feature extraction is performed on preprocessed remote sensing image data, GIS spatial data, sensor time-series data, and hyperspectral images to obtain multimodal features, including the following steps: The U-Net segmentation network was used to process remote sensing images and extract the boundary information features of urban green space patches. Spatial autocorrelation indicators were used to analyze GIS spatial data to obtain the spatial aggregation characteristics of urban green spaces. I The expression is as follows: in, n S0 is the total number of spatial units; S0 is the sum of the spatial weight matrices; W ij This is a spatial weight matrix, representing the region. i With the region j The adjacency relationship between them; area i The attribute value; The average of all regional attribute values; Multi-scale one-dimensional convolution method is used to process sensor time-series data to extract the changing trend features of carbon sink data at the minute to day scale; By utilizing a band attention mechanism, the most sensitive spectral bands of targets in hyperspectral images are identified, thereby obtaining key spectral band features that are sensitive to the carbon sink characteristics of vegetation.

5. The method for dynamic accounting of urban green space carbon sequestration based on multi-source data fusion and spatial network modeling according to claim 1, characterized in that, Identifying green space source patches in the spatial dimension data of the carbon sink data cube includes the following steps: Based on the spatial dimension data in the carbon sink data cube, the core patches of urban green space are extracted using morphological spatial pattern analysis. The relative contribution of each core urban green space patch to the overall carbon sink connectivity network is measured using the patch importance index from the landscape connectivity index, as shown in the following expression: PC represents the overall potential connectivity index of all patches in the landscape. remove，k dPC represents the probability connectivity index of the remaining patches after removing patch k. k This indicates the relative importance of patch k to the overall landscape connectivity; The core patches are ranked based on the patch importance index results, and patches with a patch importance index value greater than a set threshold are selected as green space source patches.

6. The method for dynamic accounting of urban green space carbon sequestration based on multi-source data fusion and spatial network modeling according to claim 1, characterized in that, Starting with source patches as the initial nodes, a dynamic drag surface is constructed using dynamic drag factors. Carbon flow is then simulated using circuit theory to obtain carbon flow corridors and carbon flow density grids. The process includes the following steps: The study area is divided into grid cells of uniform scale using dynamic resistance factors. The basic resistance value of each grid cell is calculated and a resistance surface is constructed. The dynamic resistance factors include one or more of the following: building density, road grade, soil organic carbon content, and intensity of human activities. By using parameter optimization methods to adjust the model parameters in the resistance surface, a dynamic resistance surface for urban carbon flow is constructed. Using circuit theory and the dynamic resistance surface, carbon flow is simulated with source patches as carbon flow injection nodes to obtain carbon flux distribution results. The carbon flux calculation expression is as follows: in, V xy The carbon potential difference between patches x and y. R xy For the cumulative resistance of the dynamic resistance surface, I xy Carbon flux; The carbon flux density raster map is output using the carbon flux distribution results, and high-efficiency carbon flux corridors are identified from the carbon flux density raster map.

7. The method for dynamic accounting of urban green space carbon sequestration based on multi-source data fusion and spatial network modeling according to claim 6, characterized in that, The parameter optimization method is a genetic algorithm, response surface methodology, or particle swarm optimization.

8. The method for dynamic accounting of urban green space carbon sequestration based on multi-source data fusion and spatial network modeling according to claim 1, characterized in that, The multi-source data includes one or more of the following: normalized vegetation index data, aboveground biomass data retrieved from lidar, meteorological data, and vegetation radial growth. The multi-source data is input into the carbon density dynamic calibration model, and the expression of the calibration model is as follows: in, This represents the carbon storage of ecosystem type i at time t; The baseline value represents the biomass retrieved by lidar multiplied by the carbon content coefficient. It is a remote sensing correction factor. This represents the normalized vegetation index value corresponding to the same grid cell at time t. The normalized vegetation index (NVI) value represents the baseline time and is used to correct for differences in NVI at different time points in order to improve the accuracy of carbon storage estimation. Radial growth represents the impact of the radial expansion of vegetation on carbon storage over time. The microclimate correction factor, which considers the effects of temperature T and humidity H on vegetation growth and carbon storage, is expressed as follows: 。 9. The method for dynamic accounting of urban green space carbon sequestration based on multi-source data fusion and spatial network modeling according to claim 1, characterized in that, A multi-scale prediction engine is constructed using a dynamic carbon density raster, and the prediction engine outputs carbon density prediction results at different time scales, including the following steps: The continuous temporal slices of the past 2n days in the carbon density dynamic grid are input into the ConvLSTM model. The input is 2n×m dimensional spatiotemporal sequence data. The attention gating mechanism is used to focus on the spatial cells in the carbon density grid that exhibit abnormal temporal variation characteristics to obtain high-frequency spatiotemporal fluctuation characteristics. The high-frequency spatiotemporal fluctuation characteristics are used to predict the carbon density value for the next 0 to n days. Carbon flow network topology indices are calculated using carbon flow corridors and carbon flow density grids. These indices, along with historical carbon density data, are input into a graph neural network model. The Node2Vec algorithm is used to perform low-dimensional embedding of nodes in the carbon flow network, capturing network features and adjacency relationships to obtain vector representations of the nodes. These node vector representations and the corresponding historical carbon density features are then combined with a graph convolutional network to model carbon flow propagation, resulting in a carbon flow propagation model. In the carbon flow propagation model, a knowledge-constrained loss function based on the photosynthesis mechanism is constructed. The expression of the knowledge-constrained loss function is as follows: in, and These are the weighting coefficients of the loss function. It is the carbon flow value predicted by the model. It is the carbon flux value output by the photosynthesis mechanism model, and MSE is the mean square error. The model is trained by combining the knowledge-constrained loss function with the mean square error and absolute error between the model's estimated values ​​and the actual historical observation values, which suppresses the risk of overfitting the model to the data bias and outputs the carbon density prediction values ​​of the main hub nodes for the next 1 to p months. The Transformer model, which incorporates urban spatial structure, built-up area, green space distribution, building density, and plant physiological parameters monitored by sensors into the Farquhar photosynthesis model, outputs carbon density trend predictions and carbon sink saturation year predictions at scales of 1 to q years.

10. A dynamic accounting system for urban green space carbon sequestration based on multi-source data fusion and spatial network modeling, employing the dynamic accounting method for urban green space carbon sequestration based on multi-source data fusion and spatial network modeling as described in any one of claims 1-9, characterized in that, include: The data sensing module is used to collect raw data on carbon sequestration in urban green spaces, preprocess the raw data, and output the preprocessed data. The fusion computing module is used to perform multimodal feature extraction and cross-modal fusion on the preprocessed data output by the data perception module. It combines the spatial interpolation method with node betweenness centrality weighting to output a carbon sink data cube with spatial × time × feature structure. The dynamic modeling module is used to identify green space source patches in the spatial dimension data of the carbon sink data cube output by the fusion computing module; using the source patches as starting nodes, a dynamic resistance surface is constructed using dynamic resistance factors, and carbon flow is simulated by combining circuit theory to output carbon flow corridors and carbon flow density grids; multi-source data is input into the carbon density dynamic calibration model to output a carbon density dynamic grid; a multi-scale prediction engine is constructed using the carbon density dynamic grid, and the prediction engine outputs carbon density prediction results at different time scales. The application service module is used to combine the carbon flow density raster and carbon density prediction results output by the dynamic modeling module with the urban spatial structure input to a three-dimensional visualization platform using geographic information system and building information modeling technology to generate carbon density heat maps and carbon flow corridor evolution maps. The heat maps and evolution maps are then used to realize dynamic carbon sink accounting.

Images