A vehicle mileage estimation method based on remote sensing image data and deep learning

Abstract

A vehicle mileage estimation method based on remote sensing imagery and deep learning includes: constructing a self-labeled dataset by dividing high-resolution satellite images into grid cells; estimating vehicle density using a YOLO-based deep learning model, which achieves multi-scale feature fusion through a three-layer pyramid network for accurate vehicle detection; abstracting the urban road network as a graph structure and employing a cross-temporal graph neural network (CrossGraphNet) to build a model of the relationship between vehicle density and speed for accurate speed prediction; and training a semi-supervised graph neural network vehicle mileage estimation model using traffic count data from ground monitoring stations as ground truth. This model combines vehicle density and speed data with road topology relationships and uses a message passing algorithm to accurately estimate vehicle mileage. This invention overcomes the limitations of traditional methods, effectively improves the efficiency and accuracy of traffic monitoring, and provides strong data support for traffic management and planning.

Description

Technical Field

[0001] This invention relates to information and intelligent transportation technology, and in particular to a vehicle mileage estimation method based on remote sensing image data and deep learning. Background Technology

[0002] With the acceleration of urbanization, the development of intelligent transportation systems is becoming increasingly important. Vehicle KM (VKM) modeling, as the core of intelligent transportation systems, provides crucial data support for urban planning, traffic management, and control. VKM refers to the cumulative kilometers a vehicle travels on a specific road and is a standard indicator for measuring road traffic load. It can be further used to estimate traffic-related emissions, energy consumption, potential collisions, and other negative environmental impacts (such as noise and air pollution). It can provide accurate traffic information for transportation departments to formulate traffic planning and road maintenance policies, and reduce the risk of traffic fatalities or injuries caused by poor road conditions. Therefore, accurate and timely VKM estimation is crucial for resource allocation and infrastructure investment, and is directly related to road efficiency and driving safety. However, existing traffic flow modeling methods mainly rely on real-time observation data from road sensors and vehicle-mounted GPS, and these traditional methods have significant limitations. First, the coverage of real-time observation data is often limited, and many road areas lack relevant data; second, the collection and processing of real-time data requires substantial infrastructure investment, resulting in high costs and making large-scale deployment difficult.

[0003] In recent years, with the rapid development of remote sensing technology, vehicle detection based on remote sensing imagery has become a new approach to traffic flow analysis. By detecting vehicle targets in high-resolution remote sensing images, vehicle density on roads can be estimated, providing important parameters for traffic flow modeling. However, existing research on remote sensing-based vehicle detection is mainly limited to static analysis and cannot capture complete spatiotemporal traffic dynamics. This is primarily because the temporal resolution of remote sensing images is relatively low, making it difficult to meet the input requirements of traditional spatiotemporal traffic models. Furthermore, it is difficult to guarantee accuracy for more complex urban spatial relationships. Taking Hong Kong as an example, according to the "Smart City Blueprint," approximately 1200 traffic detectors, including video detectors, automatic license plate recognition detectors, and Bluetooth detectors, have been installed on major roads for real-time monitoring of traffic speed and volume. However, the traffic information provided by these detectors still suffers from incompleteness. Especially at key locations such as intersections in residential areas and highway entrances, the lack of traffic information often leads drivers to make incorrect route planning decisions, thereby reducing road utilization efficiency and increasing the likelihood of traffic accidents.

[0004] It should be noted that the information disclosed in the background section above is only for understanding the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The main objective of this invention is to overcome the deficiencies in the aforementioned background technology and provide a vehicle mileage estimation method based on remote sensing image data and deep learning.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A vehicle mileage estimation method based on remote sensing image data and deep learning includes the following steps:

[0008] S1. Vehicle Density Estimation: High-resolution satellite images are divided into grid cells to identify vehicle targets in each region. A self-labeled dataset containing temporal information is constructed based on the satellite images. A vehicle density estimation model based on the YOLO deep learning framework is built. The YOLO deep learning framework performs multi-scale feature fusion through a three-layer upsampling and downsampling pyramid network to accurately detect vehicles. The model is trained using the dataset. The trained vehicle density estimation model is used to automatically detect vehicles in the satellite images and then estimate the vehicle density.

[0009] S2. Vehicle Speed ​​Estimation: The urban road network is abstractly represented as a graph structure G = (V, E, X), where V represents a node, i.e., the center point of a road segment, E represents an edge, i.e., the connection between roads, and X represents the node feature matrix, including vehicle density information. Based on this graph structure, a cross-temporal graph neural network (CrossGraphNet) is used to construct a model of the relationship between vehicle density and speed. The cross-temporal graph neural network considers the spatial topological relationship of the road network, realizes the adaptive evaluation of message passing weights between graph nodes, effectively captures the nonlinear relationship between vehicle density and speed, and thus achieves accurate vehicle speed prediction.

[0010] S3. Vehicle Mileage Estimation: Using traffic count data from ground monitoring stations as ground truth, a vehicle mileage estimation model (VKM) based on satellite image data is trained. The VKM model employs a semi-supervised graph neural network. During training, the vehicle density estimated in step S1 and the vehicle speed estimated in step S2 are used as input values ​​to the semi-supervised graph neural network, the road topology is used as the graph structure, and the actual observed mileage of the road is used as the target value. The semi-supervised graph neural network executes a message passing algorithm to achieve accurate estimation of vehicle mileage using limited labeled and unlabeled monitoring data.

[0011] Further, in step S1, the high-resolution satellite image is divided into grid cells to identify vehicle targets within each area, specifically including:

[0012] Vehicle targets are located using grid cells, where each grid cell detects the target vehicle within it. The grid cell detects the target vehicle when its center is located within the grid cell.

[0013] Bounding box regression is performed on the detected vehicle targets to determine the bounding box of each vehicle, which is defined by three attributes: width, length, and center.

[0014] Furthermore, in step S1, during the bounding box regression process, a conditional adjustment strategy is applied. This strategy determines the magnitude of the bounding box adjustment based on the sign of the regression value x. When the regression value x shows a positive deviation, the adjustment magnitude of the bounding box is proportional to the regression value x to ensure that the bounding box closely fits the target vehicle. When the regression value x shows a negative or zero deviation, the adjustment magnitude of the bounding box is reduced.

[0015] Further, in step S1, the three-layer upsampling and downsampling pyramid network specifically includes:

[0016] In the pyramid network with a three-layer upsampling and downsampling structure based on the YOLO model, the upsampling path is used to integrate shallow edge detail features with deep abstract semantic features to enhance the recognition ability of vehicle targets, while the downsampling path is used to refine the details of the feature map to assist in the accurate localization of vehicle targets. The feature maps in the upsampling and downsampling paths are fused through convolution operations to ensure resolution consistency and reduce bias in the detection process. Preferably, the fusion method uses a convolution operation with a kernel size of 1x1, and the fused feature map is convolved with a 3x3 kernel to generate the final feature map used for target regression.

[0017] Furthermore, in step S2, the vehicle density information includes remote sensing image data and vehicle density information extracted from traffic monitoring equipment such as CCTV.

[0018] Furthermore, in step S2, the cross-temporal graph neural network includes a two-layer mask structure and a cross-temporal attention calculation module to process missing data at the road sensor level and graph level, and dynamically process discontinuous data, thereby enhancing the model's ability to learn spatial dependencies and adapt to traffic conditions under different time conditions, and achieving accurate prediction of vehicle speed.

[0019] The dual-layer mask structure includes: a node mask module, which uses ST Graph Layer for feature extraction, identifies and processes missing data at the road sensor level through node masks, and enhances the model's adaptability to changes in the state of individual roads; and a graph mask module, which processes the discontinuity of graph-level data through graph masks, and maintains the overall connectivity of the road network and the stability of information flow.

[0020] The attention calculation module includes: a spatiotemporal attention mechanism, which dynamically adjusts the weight of information transmission between nodes by considering the temporal changes and spatial distribution of vehicle density through an attention layer, in order to capture the dynamic characteristics of traffic flow; and attention weighted fusion, which combines the attention weights of different time steps with node features to achieve accurate prediction of vehicle speed changes, thereby improving the model's real-time response capability to traffic conditions.

[0021] Furthermore, in step S3, the ground monitoring station includes a core station that provides long-term data and a coverage station that provides short-term data. The core station provides long-term traffic count data to construct hourly, daily, and monthly traffic characteristics, while the coverage station provides short-term traffic count data.

[0022] Further, in step S3, the message passing algorithm integrates node features and topological structure information through a message passing mechanism, specifically including: employing a two-layer message passing process, where each layer includes the transformation and aggregation of node features, followed by nonlinear activation, to enhance the model's learning ability for traffic flow features; during the message passing process, effective transmission and updating of information between nodes is achieved through the product of the normalized adjacency matrix and the feature matrix; the initial node features are transformed using a weight matrix, and nonlinear characteristics are introduced through a nonlinear activation function to improve the model's predictive ability; in the second layer of message passing, the weight matrix is ​​again applied to transform the output of the first layer to obtain the final node features.

[0023] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the vehicle mileage estimation method based on remote sensing image data and deep learning.

[0024] A computer program product includes a computer program that, when executed by a processor, implements the vehicle mileage estimation method based on remote sensing image data and deep learning.

[0025] The present invention has the following beneficial effects:

[0026] This invention proposes a vehicle mileage estimation algorithm based on remote sensing imagery and deep learning. It innovatively integrates high-resolution remote sensing imagery, deep learning, and graph neural network technologies to achieve accurate vehicle mileage estimation. This invention overcomes the limitations of traditional traffic monitoring systems, such as insufficient spatial coverage, limited sensing capabilities, and high deployment and maintenance costs. By utilizing an improved YOLO framework and a multi-scale feature fusion pyramid network, this invention can automatically identify and locate vehicles from satellite images, thereby achieving large-scale, low-cost traffic data collection. Simultaneously, this invention designs a novel spatiotemporal fusion framework that combines static remote sensing data with dynamic ground observation data through a cross-spatiotemporal graph neural network (CrossGraphNet), effectively capturing the nonlinear relationship between vehicle density and speed, significantly improving the accuracy and timeliness of vehicle mileage estimation. This design not only improves prediction accuracy but also controls the prediction error to within 10% by modeling road network spatial dependencies through graph neural networks, outperforming traditional statistical learning methods. Furthermore, this invention also designs a semi-supervised learning method, which, through knowledge transfer, effectively covers road sections without monitoring equipment, enhancing the algorithm's generalization ability and practical value. Ultimately, the vehicle mileage estimation accuracy using this invention reached 0.92, demonstrating the algorithm's high reliability and accuracy. The performance improvement was particularly significant in sparse data regions. These advantages collectively constitute a major contribution of this invention to the development of intelligent transportation systems, providing strong data support for urban planning, traffic management, and control, and contributing to the intelligent and modern development of transportation.

[0027] Other beneficial effects of the embodiments of the present invention will be further described below. Attached Figure Description

[0028] Figure 1 This is a diagram illustrating the overall framework of the vehicle mileage estimation method based on remote sensing image data and deep learning, as described in an embodiment of the present invention.

[0029] Figure 2 This is a framework diagram of a vehicle density estimation model based on high-resolution satellite data, as described in an embodiment of the present invention.

[0030] Figure 3 This is a framework diagram of the vehicle speed estimation model according to an embodiment of the present invention;

[0031] Figure 4 This is a diagram of the CrossGraphNet model framework according to an embodiment of the present invention.

[0032] Figure 5 This is a framework diagram of the vehicle mileage estimation model according to an embodiment of the present invention. Detailed Implementation

[0033] The embodiments of the present invention will be described in detail below. It should be emphasized that the following description is merely exemplary and not intended to limit the scope and application of the present invention.

[0034] To address the limitations of traditional ground traffic monitoring systems, such as limited spatial coverage, insufficient sensing capabilities, and high deployment and maintenance costs, this invention proposes an innovative solution: a vehicle mileage estimation method based on remote sensing image data and deep learning. The core contributions of this invention can be summarized as follows: 1. A vehicle detection and density estimation method based on remote sensing imagery is proposed. Utilizing the constructed deep learning model, vehicles can be automatically and accurately identified and located from high-resolution satellite images, achieving wide-area and cost-effective traffic data acquisition. 2. An innovative spatiotemporal fusion framework is designed, integrating static remote sensing data with dynamic ground observation data. A deep learning model is used to capture and analyze spatiotemporal dependencies, significantly improving the accuracy and timeliness of vehicle mileage estimation. 3. This invention implements a comprehensive algorithm that integrates data acquisition, processing, analysis, and visualization functions. This algorithm not only has good scalability but also universality, meeting the needs of traffic monitoring in different cities.

[0035] See Figure 1 This invention provides a vehicle mileage estimation method based on remote sensing image data and deep learning, comprising the following steps:

[0036] Step S1, Vehicle Density Estimation: Divide the high-resolution satellite image into grid cells to identify vehicle targets in each region. Based on the satellite image, construct a self-labeled dataset containing temporal information, and build a vehicle density estimation model based on the YOLO deep learning framework (see...). Figure 2 The YOLO deep learning framework uses a three-layer upsampling pyramid network to perform multi-scale feature fusion to accurately detect vehicles; it trains a model using the dataset; and it uses the trained vehicle density estimation model to automatically detect vehicles in satellite images and then estimate vehicle density.

[0037] In a preferred embodiment, step S1, dividing the high-resolution satellite image into grid cells to identify vehicle targets in each region, specifically includes: locating vehicle targets using grid cells, wherein each grid cell detects target vehicles within it, and the grid cell detects a target vehicle when the center of the target vehicle is located within the grid cell; performing bounding box regression on the detected vehicle targets to determine the bounding box of each vehicle, the bounding box being defined by three attributes: the width, length, and center of the vehicle.

[0038] In a preferred embodiment, a conditional adjustment strategy is applied during the bounding box regression process. This strategy determines the magnitude of the bounding box adjustment based on the sign of the regression value x. When the regression value x shows a positive deviation, the adjustment magnitude of the bounding box is proportional to the regression value x to ensure that the bounding box closely fits the target vehicle; when the regression value x shows a negative or zero deviation, the adjustment magnitude of the bounding box is reduced. A specific example of the bounding box regression strategy is shown in Formula 1:

[0039]

[0040] In a preferred embodiment, the three-layer upsampling and downsampling pyramid network specifically includes: in the three-layer upsampling and downsampling pyramid network based on the YOLO model, the upsampling path is used to integrate shallow edge detail features and deep abstract semantic features to enhance the recognition ability of vehicle targets, and the downsampling path is used to refine the details of the feature map to assist in the accurate localization of vehicle targets; the feature maps in the upsampling path and the downsampling path are fused through convolution operation to ensure resolution consistency and reduce deviations in the detection process; preferably, the fusion method uses a convolution operation with a kernel size of 1x1, and the fused feature map is generated by 3x3 kernel convolution to generate the final feature map for target regression.

[0041] Step S2, Vehicle Speed ​​Estimation: The urban road network is abstractly represented as a graph structure G = (V, E, X), where V represents a node, i.e., the center point of a road segment, and E represents an edge, i.e., the connection between roads. A node feature matrix X is defined, and the elements of the feature matrix contain vehicle density information. Based on this graph structure, a cross-temporal graph neural network (CrossGraphNet) is used to construct a model of the relationship between vehicle density and speed (see...). Figure 3 The spatiotemporal graph neural network considers the spatial topology of the road network, realizes adaptive evaluation of message passing weights between graph nodes, effectively captures the nonlinear relationship between vehicle density and speed, and thus achieves accurate vehicle speed prediction.

[0042] In some embodiments, the vehicle density information may include remote sensing image data and vehicle density information extracted from traffic monitoring equipment such as CCTV.

[0043] See Figure 4In a preferred embodiment, the spatiotemporal graph neural network includes a two-layer mask structure and a spatiotemporal attention computation module to handle missing road sensor-level and graph-level data, and dynamically process discontinuous data, thereby enhancing the model's ability to learn spatial dependencies and adapt to traffic conditions under different time conditions, and achieving accurate prediction of vehicle speed. The two-layer mask structure includes: a node mask module, which uses an ST Graph Layer for feature extraction, identifies and processes missing road sensor-level data through node masks, enhancing the model's adaptability to changes in individual road conditions; and a graph mask module, which processes discontinuities in graph-level data through graph masks, maintaining the overall connectivity of the road network and the stability of information flow. The attention computation module includes: a spatiotemporal attention mechanism, which, through an attention layer, considers the temporal changes and spatial distribution of vehicle density, dynamically adjusts the weights of information transmission between nodes to capture the dynamic characteristics of traffic flow; and attention weighted fusion, which combines the attention weights at different time steps with node features to achieve accurate prediction of vehicle speed changes, thereby improving the model's real-time response to traffic conditions.

[0044] Step S3, Vehicle Mileage Estimation: Using traffic count data from ground monitoring stations as ground truth, a vehicle mileage estimation model (VKM) based on satellite image data is trained. The VKM model employs a semi-supervised graph neural network. During training, the vehicle density estimated in step S1 and the vehicle speed estimated in step S2 are used as input values ​​to the semi-supervised graph neural network, the road topology is used as the graph structure, and the actual observed mileage of the road is used as the target value. The semi-supervised graph neural network executes a message passing algorithm, using limited labeled data (ground truth) and a large amount of unlabeled data (satellite and other monitoring data) to achieve accurate estimation of vehicle mileage.

[0045] In a preferred embodiment, the ground monitoring station includes a core station providing long-term data and a coverage station providing short-term data. The core station provides long-term traffic count data to construct hourly, daily, and monthly traffic characteristics, while the coverage station provides short-term traffic count data.

[0046] In a preferred embodiment, in step S3, the message passing algorithm integrates node features and topological information through a message passing mechanism, specifically including: employing a two-layer message passing process, where each layer includes the transformation and aggregation of node features, followed by nonlinear activation, to enhance the model's learning ability for traffic flow features; in the first-layer message passing process, effective transmission and updating of information between nodes is achieved through the product of the normalized adjacency matrix and the feature matrix; the initial node features are transformed using a weight matrix, and nonlinear characteristics are introduced through a nonlinear activation function to improve the model's predictive ability. In the second-layer message passing, the weight matrix is ​​again applied to transform the output of the first layer to obtain the final node features. A specific expression of this message passing algorithm is Equation 2:

[0047]

[0048] The method of this invention effectively solves the shortcomings of traditional traffic monitoring systems in terms of spatial coverage, sensing capabilities, and cost, improves the efficiency and accuracy of traffic monitoring, and provides strong data support for traffic management and planning.

[0049] The following further describes specific embodiments and experimental verifications of the present invention.

[0050] A vehicle mileage estimation method based on remote sensing image data and deep learning, the overall flowchart is as follows: Figure 1 As shown, the main steps include:

[0051] Step 1, as follows Figure 2 As shown, a vehicle density estimation model is constructed based on high-resolution satellite data. By building a self-labeled satellite remote sensing dataset containing temporal information, and combining an improved YOLO deep learning framework and a multi-scale feature fusion pyramid network, the detection accuracy for small vehicles such as cars and taxis is significantly improved, laying a data foundation for subsequent analysis.

[0052] Step 2, as follows Figure 3 As shown, a graph neural network is innovatively introduced for vehicle speed estimation. This network fully considers the spatial topology of the road network and, through the design of a cross-temporal graph neural network (CrossGraphNet), achieves adaptive evaluation of message passing weights between graph nodes, effectively capturing the nonlinear relationship between vehicle density and speed, thereby achieving accurate vehicle speed prediction. Figure 4 ).

[0053] Step 3, as follows Figure 5As shown, a semi-supervised graph neural network model is used to estimate vehicle mileage based on acquired vehicle density and speed data. Vehicle density and speed serve as the input x-values ​​to the semi-supervised graph neural network, road topology serves as the graph structure, and the actual observed mileage of the road serves as the y-values ​​of the model. This method overcomes the geographical limitations of traditional traffic counting stations, achieves effective coverage of road sections without monitoring equipment, and provides a new technical means for urban traffic monitoring and management.

[0054] Example

[0055] Yuen Long District in Hong Kong was selected as the research area. In recent years, Yuen Long District has experienced rapid development, but traffic congestion has become increasingly severe. Besides peak hours, roads such as Gong Kang Road, Shap Sha Heung Road, Tai Tong Road, and Tai Shu Ha East-West Road are frequently congested, especially in certain sections. As a typical traffic-congested area in Hong Kong, Yuen Long District provides rich experimental case studies. First, the street network of Yuen Long District was obtained from the OSM platform, and its spatial topology was analyzed as the basis for constructing the GNN network. Specifically, each street was treated as a node, and edges were constructed based on the topological relationships. Simultaneously, the ground infrastructure in Yuen Long District (such as CCTV and vehicle counting stations) is well-distributed and numerous, providing objective validation data for this project. Furthermore, based on factors such as resolution, imaging band, and historical storage, three optical sensors—QuickBird, WorldView-3, and GF-2—were selected to collect the remote sensing data required for the experiment. Their imaging resolution of less than 1 meter is more suitable for vehicle detection, avoiding errors. In addition, the rail service has been in operation for over a year, and its data acquisition and preprocessing techniques are mature. In this example, high-resolution satellite remote sensing images were collected for 50 time periods from 2010 to 2022.

[0056] Step 1: Construct a vehicle density estimation model based on high-resolution satellite data. First, the image is divided into multiple grid cells. Each grid cell is responsible for detecting targets within it. If the center of a target appears within a grid cell, that cell is responsible for detecting that target. Then, bounding box regression is performed (see Equation 1), with each bounding box containing three attributes: width, length, and center. Next, a three-layer pyramid network with upsampling and downsampling structures is designed based on the YOLO model. The upsampling path integrates shallow edge detail features with deep abstract semantic features. Feature maps of the same size from the upsampling and downsampling paths are fused; the former is used for localization, and the latter for refining feature map details. Finally, nearest neighbor sampling is used as the upsampling strategy to maintain consistent resolution between the upsampling and downsampling paths. The fusion method uses a 1x1 kernel convolution operation. The fused feature map is then convolved with a 3x3 kernel to generate the final feature map used for target regression. This step is used to eliminate the jagged edges caused by sub-pixel offsets resulting from the fusion of the upsampling and downsampling paths.

[0057]

[0058] Step 2: Construct a model of the relationship between vehicle density and speed based on a graph neural network (GNN). First, the complex urban road network is summarized as a graph G = (V, E, X), where V = {v1, ..., v...} n} represents a road segment (with the center point of the road segment as a node), E = {e 0,1 ,...,e ij} represents the connection between roads (the topological relationship between roads is represented as edges), and the node feature matrix X = {x1,...,x...} is defined. n The feature matrix elements contain feature information such as vehicle density associated with nodes. Secondly, to make the model spatially dependent, this invention designs a cross-spatiotemporal graph neural network (CrossGraphNet) traffic flow reconstruction method. It includes two key modules: the first is a two-layer mask structure, and the second is cross-spatiotemporal attention computation. These are used to handle missing road sensor-level and graph-level data and dynamically process discontinuous data, thereby improving the model's adaptability and robustness (Table 1).

[0059] Table 1: Performance Comparison of Graph Neural Network Models

[0060]

[0061]

[0062] Step 3: Based on the acquired vehicle density and speed data, develop a vehicle mileage estimation algorithm. Traffic count data from core stations and coverage stations are used as ground truth. Core stations provide long-term traffic count data to construct hourly, daily, and monthly factors to summarize traffic characteristics. Coverage stations only provide short-term (one to two days per year) traffic count data. The VKM observed at each station is used as ground truth to train a satellite-based VKM estimation model. The learned relationships are transferred to major roads without counting stations, and VKM estimation is performed using only satellite-based traffic statistics. The satellite-based VKM model is evaluated using existing station VKM as ground truth to ensure its effectiveness. Vehicle density and speed are used as input x-values ​​to the semi-supervised graph neural network, road topology is used as the graph structure, and the actual observed mileage of the road is used as the y-value of the model. Figure 5 Finally, through vehicle mileage estimation algorithms, professional business functions including estimation services and visualization analysis can be realized. Figure 1 This can provide traffic management departments with more accurate decision-making basis and promote the intelligent and modern development of the transportation system.

[0063] For a detailed expression of the message passing algorithm for semi-supervised graphical neural networks, please refer to Formula 2:

[0064]

[0065] A two-layer message passing process is adopted, where each layer process includes the node feature H. (0) Transformation and aggregation, along with subsequent nonlinear activation σ, are used to enhance the model's ability to learn traffic flow features; during the first-layer message passing process, the adjacency matrix is ​​normalized. and characteristic matrix H (0) The product of W and W enables efficient information transfer and updating between nodes; the weight matrix W is applied. (0) The initial node features are transformed, and a nonlinear characteristic is introduced through a nonlinear activation function σ to improve the model's predictive ability. In the second-layer message passing, the weight matrix W is applied again. (1) The output of the first layer is transformed to obtain the final node features H. (2) .

[0066] This invention has the following advantages and positive effects:

[0067] (1) An innovative technical architecture combines high-resolution remote sensing imagery, deep learning, and graph neural networks for vehicle mileage estimation. An improved YOLO framework and a multi-scale feature fusion pyramid network enhance vehicle detection accuracy, addressing the limitations of traditional methods that rely on ground sensors and have limited coverage. The vehicle mileage estimation results provide a reliable data foundation for assessing and controlling traffic emissions, and also provide a basis for environmental protection departments to formulate precise emission reduction policies.

[0068] (2) An innovative spatiotemporal fusion framework based on graph neural networks was designed, fully considering the topological relationships and temporal characteristics of the road network. By designing a cross-spatiotemporal graph neural network (CrossGraphNet), adaptive evaluation of message passing weights between road network nodes was achieved, effectively capturing the complex nonlinear relationship between vehicle density and speed. By modeling the spatial dependencies of the road network using graph neural networks, the prediction error was controlled within 10%, significantly outperforming traditional statistical learning methods.

[0069] (3) A groundbreaking semi-supervised learning method is adopted to combine limited road observation data with a large amount of unlabeled data. Through knowledge transfer from graph structures, effective coverage of road sections without monitoring equipment is achieved, significantly improving the algorithm's generalization ability and practical value. The final vehicle mileage estimation accuracy reaches 0.92, proving the reliability and accuracy of the algorithm of this invention. Especially in data-sparse regions, the improvement in estimation performance is more significant.

[0070] This invention also provides a storage medium for storing a computer program, which, when executed, performs at least the methods described above.

[0071] This invention also provides a control device, including a processor and a storage medium for storing a computer program; wherein the processor executes the computer program by performing at least the method described above.

[0072] This invention also provides a processor that executes a computer program, at least performing the methods described above.

[0073] The storage medium can be implemented by any type of non-volatile storage device, or a combination thereof. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), magnetic random access memory (FRAM), flash memory, magnetic surface memory, optical disc, or compact disc read-only memory (CD-ROM); the magnetic surface memory can be a disk drive or magnetic tape drive. The storage media described in the embodiments of this invention are intended to include, but are not limited to, these and any other suitable types of memory.

[0074] In the several embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0075] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0076] In addition, in the various embodiments of the present invention, each functional unit can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0077] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0078] Alternatively, if the integrated units of this invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

[0079] The methods disclosed in the several method embodiments provided by this invention can be arbitrarily combined without conflict to obtain new method embodiments.

[0080] The features disclosed in the several product embodiments provided by this invention can be arbitrarily combined without conflict to obtain new product embodiments.

[0081] The features disclosed in the several method or device embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method or device embodiments.

[0082] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various equivalent substitutions or obvious modifications can be made without departing from the concept of the present invention, and all such modifications, achieving the same performance or application, should be considered within the scope of protection of the present invention.

[0083] The background section of this invention may include background information about the problems or environment in which the invention is being developed, and is not necessarily a description of prior art. Therefore, the content included in the background section does not constitute an admission of prior art by the applicant.

[0084] The above description provides a further detailed explanation of the present invention in conjunction with specific / preferred embodiments, and it should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various substitutions or modifications can be made to these described embodiments without departing from the concept of the present invention, and all such substitutions or modifications should be considered within the scope of protection of the present invention. In the description of this specification, the reference to terms such as "an embodiment," "some embodiments," "preferred embodiment," "example," "specific example," or "some examples," etc., indicates that the specific features, structures, materials, or characteristics described in connection with that embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in any suitable manner in one or more embodiments or examples. Without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification and the features of different embodiments or examples. Although the embodiments of the present invention and their advantages have been described in detail, it should be understood that various changes, substitutions, and modifications can be made herein without departing from the scope of protection of the patent application.

Claims

1. A vehicle mileage estimation method based on remote sensing image data and deep learning, characterized in that, Includes the following steps: S1. Vehicle Density Estimation: High-resolution satellite images are divided into grid cells to identify vehicle targets in each region. A self-labeled dataset containing temporal information is constructed based on the satellite images. A vehicle density estimation model based on the YOLO deep learning framework is built. The YOLO deep learning framework performs multi-scale feature fusion through a three-layer upsampling and downsampling pyramid network to accurately detect vehicles. The model is trained using the dataset. The trained vehicle density estimation model is used to automatically detect vehicles in the satellite images and then estimate the vehicle density. S2. Vehicle Speed ​​Estimation: The urban road network is abstractly represented as a graph structure G = (V, E, X), where V represents a node, i.e., the center point of a road segment, and E represents an edge, i.e., the connection between roads. A node feature matrix X is defined, and the elements of the feature matrix contain vehicle density information. Based on this graph structure, a cross-temporal graph neural network (CrossGraphNet) is used to construct a model of the relationship between vehicle density and speed. The cross-temporal graph neural network considers the spatial topological relationship of the road network, realizes the adaptive evaluation of message passing weights between graph nodes, captures the nonlinear relationship between vehicle density and speed, and thus achieves accurate vehicle speed prediction. S3. Vehicle Mileage Estimation: Using traffic count data from ground monitoring stations as ground truth, a vehicle mileage estimation model based on satellite image data is trained. The vehicle mileage estimation model adopts a semi-supervised graph neural network. During training, the vehicle density estimated in step S1 and the vehicle speed estimated in step S2 are used as input values ​​of the semi-supervised graph neural network, the road topology is used as the graph structure, and the actual observed mileage of the road is used as the target value. The semi-supervised graph neural network executes a message passing algorithm to achieve accurate estimation of vehicle mileage using limited labeled and unlabeled monitoring data. 2.The vehicle mileage estimation method based on remote sensing image data and deep learning according to claim 1, wherein, In step S1, the high-resolution satellite image is divided into grid cells to identify vehicle targets within each area, specifically including: Vehicle targets are located using grid cells, where each grid cell detects the target vehicle within it. The grid cell detects the target vehicle when its center is located within the grid cell. Bounding box regression is performed on the detected vehicle targets to determine the bounding box of each vehicle, which is defined by three attributes: width, length, and center. 3.The vehicle mileage estimation method based on remote sensing image data and deep learning according to claim 2, wherein, In step S1, during the bounding box regression process, a conditional adjustment strategy is applied. This strategy determines the magnitude of the bounding box adjustment based on the sign of the regression value x. When the regression value x shows a positive deviation, the adjustment magnitude of the bounding box is proportional to the regression value x to ensure that the bounding box closely fits the target vehicle. When the regression value x shows a negative or zero deviation, the adjustment magnitude of the bounding box is reduced. 4.The vehicle mileage estimation method based on remote sensing image data and deep learning according to any one of claims 1 to 3, characterized in that, In step S1, the three-layer upsampling and downsampling pyramid network specifically includes: In the pyramid network with a three-layer upsampling and downsampling structure based on the YOLO model, the upsampling path is used to integrate shallow edge detail features with deep abstract semantic features to enhance the recognition ability of vehicle targets, while the downsampling path is used to refine the details of the feature map to assist in the accurate localization of vehicle targets. The feature maps in the upsampling and downsampling paths are fused through convolution operations to ensure resolution consistency and reduce bias in the detection process. Preferably, the fusion method uses a convolution operation with a kernel size of 1x1, and the fused feature map is convolved with a 3x3 kernel to generate the final feature map used for target regression.

5. The vehicle mileage estimation method based on remote sensing image data and deep learning as described in any one of claims 1 to 4, characterized in that, In step S2, the vehicle density information includes remote sensing image data and vehicle density information extracted from traffic monitoring equipment such as CCTV.

6. The vehicle mileage estimation method based on remote sensing image data and deep learning as described in any one of claims 1 to 5, characterized in that, In step S2, the cross-temporal graph neural network includes a two-layer mask structure and a cross-temporal attention calculation module to process missing data at the road sensor level and graph level, and dynamically process discontinuous data, thereby enhancing the model's ability to learn spatial dependencies and adapt to traffic conditions under different time conditions, and achieving accurate prediction of vehicle speed. The dual-layer mask structure includes: a node mask module, which uses ST Graph Layer for feature extraction, identifies and processes missing data at the road sensor level through node masks, and enhances the model's adaptability to changes in the state of individual roads; and a graph mask module, which processes the discontinuity of graph-level data through graph masks, and maintains the overall connectivity of the road network and the stability of information flow. The attention calculation module includes: a spatiotemporal attention mechanism, which dynamically adjusts the weight of information transmission between nodes by considering the temporal changes and spatial distribution of vehicle density through an attention layer, in order to capture the dynamic characteristics of traffic flow; and attention weighted fusion, which combines the attention weights of different time steps with node features to achieve accurate prediction of vehicle speed changes, thereby improving the model's real-time response capability to traffic conditions.

7. The vehicle mileage estimation method based on remote sensing image data and deep learning as described in any one of claims 1 to 6, characterized in that, In step S3, the ground monitoring station includes a core station that provides long-term data and a coverage station that provides short-term data. The core station provides long-term traffic count data to construct hourly, daily, and monthly traffic characteristics, while the coverage station provides short-term traffic count data.

8. The vehicle mileage estimation method based on remote sensing image data and deep learning as described in any one of claims 1 to 7, characterized in that, In step S3, the message passing algorithm integrates node features and topological structure information through a message passing mechanism. Specifically, it includes: employing a two-layer message passing process, where each layer includes the transformation and aggregation of node features, followed by nonlinear activation, to enhance the model's learning ability for traffic flow features; in the first-layer message passing process, the effective transmission and updating of information between nodes is achieved through the product of the normalized adjacency matrix and the feature matrix; the initial node features are transformed using a weight matrix, and nonlinear characteristics are introduced through a nonlinear activation function to improve the model's predictive ability; in the second-layer message passing process, the weight matrix is ​​applied again to transform the output of the first layer to obtain the final node features.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the vehicle mileage estimation method based on remote sensing image data and deep learning as described in any one of claims 1 to 8.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the vehicle mileage estimation method based on remote sensing image data and deep learning as described in any one of claims 1 to 8.

Images