A highway route selection design scheme decision method based on multi-source fusion

Abstract

The application discloses a highway route selection design scheme decision-making method based on multi-source fusion. Multi-source data of a highway planning area is collected and standardized pretreated, ground objects are recognized and features are extracted through an improved target detection model, constraint functions with cost, ecology and compliance as targets are constructed by fusing multi-dimensional data, weights are determined by using an analytic hierarchy process, candidate schemes are generated by stage optimization of three-dimensional route selection based on deep reinforcement learning, and the optimal scheme is determined by comprehensive scoring and curvature ratio selection. The method solves the problems of traditional route selection, such as dependence on manual work, low efficiency and one-sided decision-making, improves the scientificity, economy and driving safety of the scheme, and is suitable for the scene of highway route selection design.

Description

Technical Field

[0001] This invention relates to the field of intelligent highway alignment technology, and in particular to a decision-making method for highway alignment design schemes based on multi-source fusion. Background Technology

[0002] With the continuous advancement of national transportation infrastructure construction, highway engineering, as a core carrier connecting regional economies and ensuring people's travel needs, directly determines the construction cost, ecological and environmental impact, and long-term operational safety of its route selection and design scheme. Highway route selection requires comprehensive consideration of multiple dimensions, including terrain conditions, distribution of land features, protection of ecologically sensitive areas, control of construction and land acquisition costs, and compliance with route design standards. It is a complex, multi-objective decision-making task that has a decisive impact on the efficiency and overall benefits of subsequent project implementation.

[0003] Currently, highway alignment design still relies heavily on traditional manual surveys and experience-based decision-making. While some systems incorporate geographic information data to aid analysis, they still face numerous challenges in data integration, intelligent analysis, and optimization decision-making. The data sources required for alignment selection are diverse, with inconsistent formats and significant semantic differences. Traditional data processing methods struggle to achieve deep integration of multi-source information, easily leading to incomplete feature identification, large cost estimation errors, or insufficient consideration of ecological constraints, thus affecting the overall scientific validity of the plan.

[0004] Therefore, how to achieve standardized fusion and intelligent feature recognition of multi-source alignment data, construct a scientific multi-objective constraint model, combine three-dimensional terrain features for efficient alignment optimization, and establish a comprehensive candidate scheme decision-making mechanism has become a key issue in improving the scientificity, economy, and safety of highway alignment design schemes, and is also one of the core directions that urgently need to be broken through in the field of intelligent highway alignment technology. Summary of the Invention

[0005] This invention provides a decision-making method for highway alignment design based on multi-source fusion. By standardizing and preprocessing multi-source data from the highway planning area, and combining this with intelligent feature identification to extract key features, a multi-dimensional data fusion system and a multi-objective constraint model are constructed. Deep reinforcement learning is used to achieve dynamic optimization of three-dimensional alignment. Finally, the optimal route is determined through comprehensive comparison of multiple indicators. This method solves the problems of traditional alignment selection relying on human experience, insufficient data fusion, low optimization efficiency, and biased decision-making, providing scientific and efficient intelligent decision support for highway alignment design.

[0006] To achieve the above objectives, the technical solution adopted by the present invention includes the following steps:

[0007] Multi-source data acquisition and standardized preprocessing: Collect basic geographic data of the highway planning area, including image data, vector data and design data, and perform format standardization and preprocessing on the collected data to generate standardized data in a unified format;

[0008] Ground feature identification and feature extraction: Based on the preprocessed data, a ground feature training set is constructed, and an improved target detection model is used to identify ground features and extract features, outputting ground feature feature data;

[0009] Multi-source data fusion and multi-objective constraint modeling: Associate multi-dimensional data, including land acquisition and demolition costs, terrain attributes, and ecological sensitivity zone determination, to construct a fusion dataset; Define a comprehensive constraint function with the objectives of minimizing construction costs, land acquisition and demolition costs, ecological impact, and maximizing route compliance rate, and use the analytic hierarchy process to determine the weight coefficients of each objective;

[0010] 3D route selection optimization based on deep reinforcement learning: a reinforcement learning framework is built, the state space and action space are defined, a value function is constructed based on a multi-objective constraint function, and candidate route schemes are generated through iterative training;

[0011] Candidate scheme comparison: A comprehensive scoring formula is constructed based on construction cost, land acquisition and demolition cost, ecological impact and compliance rate to calculate the preliminary optimal scheme. If the score difference between the preliminary optimal and the second-best scheme is less than a set threshold, the final scheme is selected by comparing the maximum curvature.

[0012] As a further aspect of the present invention, in the land feature identification and feature extraction, the pre-trained model construction includes: labeling the cropped image with land features, wherein the labeled land feature categories cover at least one or more of the following: buildings, farmland, ponds, woodlands, roads, canals, wasteland, and ecologically sensitive areas.

[0013] As a further aspect of this invention, in multi-source data fusion and multi-objective constraint modeling, the expression for the multi-objective constraint function is: ;in, Construction costs, including roadbed, bridge and tunnel, and pavement costs, are calculated according to highway engineering cost estimation rules. The cost of land acquisition and demolition is the total amount of land acquisition and demolition for all land features. The amount of land acquisition and demolition is calculated based on the area of ​​the land features and the local land acquisition and demolition compensation standards. The ecological impact index is calculated according to preset rules based on the proportion of land features within the ecologically sensitive area. The route compliance rate is the ratio of the number of compliant design indicators to the total number of design indicators. Design indicators include horizontal curve radius, longitudinal slope, and sight distance. , , , The weighting coefficients are determined by the scenario: a weighted knowledge base containing multi-terrain highway projects is constructed. The weighted knowledge base needs to include historical data of highway projects completed by the unit. After inputting the terrain type of the new project, the initial weights of the projects with the most similar terrain features and road grades in the knowledge base are matched by the cosine similarity algorithm. Then, the weights are adjusted by combining the ecological expropriation and demolition requirements of the new project with the analytic hierarchy process.

[0014] As a further aspect of this invention, in the 3D route selection optimization using deep reinforcement learning, the deep reinforcement learning framework includes: adopting a phased optimization strategy; in the coarse selection phase, simplifying the state space based on low-precision DEM data, quickly screening candidate corridors that meet ecological compliance requirements, and narrowing the search range; in the fine selection phase, loading high-precision orthophotos and detailed ground feature data within the candidate corridors, and enabling the complete state space; the action space includes horizontal offset of the horizontal route and elevation adjustment of the slope change points in the longitudinal profile, wherein the horizontal offset must meet the constraints of avoiding ecologically sensitive areas and the radius of the horizontal curve must meet the design specifications; when the distance between the route and the ecologically sensitive area reaches a preset proximity threshold, the weight coefficient of the ecological impact target is adjusted.

[0015] As a further aspect of this invention, the iterative training and route selection optimization includes: initializing an experience cache pool to meet the sample storage requirements; initializing the target neural network and the valuation neural network, with the initial parameters of the target neural network being consistent with those of the valuation neural network; employing an ε-greedy strategy to select actions, randomly selecting actions with ε probability and selecting the action with the largest value function with 1-ε probability; when the number of samples in the experience cache pool reaches a set threshold, extracting the number of samples to meet the network update requirements each time, calculating the MSE loss of the valuation neural network, and updating the parameters of the valuation neural network; synchronizing the parameters of the valuation neural network to the target neural network; during training, prioritizing the extraction of samples within candidate corridors to perform batch sampling optimization; when the fluctuation range of the constraint function value corresponding to the optimal route scheme meets the stability requirements of the scheme in consecutive preset iterations, stopping training and outputting candidate route schemes.

[0016] As a further aspect of the present invention, in the preliminary comparison stage, a comprehensive scoring formula is used to calculate the score of the candidate solutions. The formula is as follows: ;in , , The maximum values ​​of construction cost, land acquisition and demolition cost, and ecological impact index among all candidate schemes are respectively used, and the scheme with the highest score is determined as the preliminary optimal scheme.

[0017] Explanation of beneficial effects:

[0018] 1. This invention achieves unified format of multi-source data for highway planning areas by constructing a standardized preprocessing process for multi-source data and an intelligent feature recognition mechanism for improved target detection models. At the same time, it combines data augmentation and semantic segmentation reasoning to complete feature boundary extraction, effectively overcoming the limitations of traditional route selection, such as chaotic data format, reliance on manual feature recognition, and low feature extraction accuracy. This significantly improves the efficiency and accuracy of early data processing in route selection design.

[0019] 2. This invention constructs a unified fusion dataset through multi-dimensional data association and integration, and combines it with multi-objective constraint modeling to achieve multi-objective collaborative optimization of construction costs, land acquisition and demolition costs, ecological impact, and compliance rate. This mechanism not only solves the problems of fragmented multi-source data and subjective target weight settings in traditional route selection, but also adapts to different terrains and ecological needs through historical project data, ensuring that the objective function is more in line with actual engineering scenarios, effectively balancing economic costs and ecological protection, and improving the comprehensive benefits and scenario adaptability of the route selection scheme.

[0020] 3. This invention constructs a two-layer decision-making mechanism in the intelligent decision-making stage. First, it calculates the preliminary optimal solution through multiple indicators, and then supplements the alignment safety assessment for solutions with similar scores. This mechanism solves the one-sidedness of traditional alignment decision-making, which relies on a single indicator and ignores driving safety. It ensures that the final solution meets both economic and ecological needs and has excellent alignment conditions, providing key guarantees for the long-term safe operation of highways and improving the scientific and comprehensive nature of alignment decision-making. Attached Figure Description

[0021] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0022] Figure 1 This is a flowchart illustrating the steps of a highway alignment design scheme decision-making method based on multi-source fusion according to the present invention.

[0023] Figure 2 This is a schematic diagram of a highway alignment design scheme based on a multi-source fusion decision-making method. Detailed Implementation

[0024] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0025] Example 1

[0026] S100: Multi-source data acquisition and standardized preprocessing

[0027] Multi-source data acquisition: First, the core data types required for highway route selection design are identified, mainly including basic geographic data of the highway planning area. This type of data encompasses three main categories: image data, vector data, and engineering design data. Image data reflects the visual features of the planning area, vector data presents the spatial location and attribute relationships of features (such as land use type, existing road distribution, etc.), and engineering design data includes preliminary route design concepts and related technical parameters. During the acquisition process, it is essential to ensure that the data covers the complete area of ​​the planning region and that the data timeliness meets current engineering design requirements to avoid design deviations due to data lag.

[0028] Furthermore, data format standardization processing is implemented: For different types of collected data, a unified format operation is performed. For image data, professional geographic information processing tools are used for stitching to eliminate overlap errors and coordinate offsets between images, ultimately outputting orthophotos that conform to engineering standards, ensuring the spatial accuracy and visual coherence of the images. For vector data, it is converted according to vector formats commonly used in highway engineering, unifying the data's coordinate system, attribute field structure, and encoding rules to facilitate subsequent data overlay analysis and related calls. For engineering design data, it is exported from the original design software to a universally applicable engineering design data format, retaining key information such as alignment coordinates, design elevations, and technical indicators, ensuring that the design data can be effectively integrated with geographic data.

[0029] Further data preprocessing involves optimizing the standardized data, with a focus on performing a series of quality enhancement operations on image data. Noise reduction eliminates noise generated during image acquisition due to equipment interference and environmental factors, preventing noise from interfering with subsequent feature identification. Enhancement adjusts visual parameters such as contrast and brightness to highlight edge features and details, making feature outlines clearer and more discernible. Grayscale conversion transforms color images into grayscale images, reducing data dimensionality and computational load while preserving core visual features, laying the foundation for efficient computation of subsequent intelligent feature identification models.

[0030] S200: Intelligent Identification and Feature Extraction of Ground Features

[0031] Construction of the Ground Feature Training Set: First, representative image segments are selected from the preprocessed image data and sliced ​​at fixed sizes to form a sample set for model training. Then, using professional image annotation tools, ground feature classification and annotation are performed on the image slices in the sample set according to the actual needs of highway engineering alignment design. The labeled ground feature categories must at least cover types closely related to alignment selection, such as buildings, farmland, ponds, woodlands, roads, canals, wasteland, and ecologically sensitive areas, ensuring coverage of ground features that need to be considered during alignment selection. During the annotation process, the boundaries of ground features must be accurately delineated, and each ground feature is assigned a corresponding category label. Finally, a ground feature training set containing image samples and label information is formed for subsequent model training.

[0032] Further, the pre-trained model was constructed and trained: a ground feature recognition network was built based on the improved object detection model. The improvement mainly focused on enhancing the model's recognition accuracy and generalization ability for specific ground features in highway engineering, such as optimizing the model's feature extraction network structure and enhancing its ability to recognize small-sized ground features and ground features in complex backgrounds. Subsequently, various parameters required for model training were set, including training batch size, initial learning rate, and number of iterations. To avoid overfitting, data augmentation operations were performed, such as flipping, rotating, adjusting brightness, and mosaicking the training samples to expand sample diversity and improve the model's adaptability to ground features in different scenes and imaging conditions. Finally, the model was pre-trained using the constructed ground feature training set, and the model parameters were optimized through multiple iterations to achieve the recognition accuracy required for engineering projects.

[0033] Further, feature identification and feature extraction: The pre-processed and optimized image data is input into a pre-trained model to perform semantic segmentation and inference operations. The model achieves accurate feature identification through pixel-by-pixel analysis and classification of the image pixels, and outputs feature category information and pixel-level masks. Based on the mask results output by the model, professional image processing algorithms are used to extract the polygonal boundaries of the features, ensuring the accuracy and integrity of the boundaries. Subsequently, based on the spatial resolution of the image and the boundary coordinates, the actual area of ​​each feature is calculated. Simultaneously, using the spatial overlay analysis function of the geographic information system, the features are associated with the route stationing layer to determine the route stationing interval of each feature. Finally, information such as feature category, area, and stationing interval is integrated to output structured feature feature data, providing basic data support for subsequent cost calculation and ecological assessment.

[0034] S300 Multi-Source Data Fusion and Multi-Objective Constraint Modeling

[0035] Multi-dimensional data association processing: Association logic is established for multi-dimensional data, including feature data, land acquisition and compensation standards, topographic data, and vector data of ecologically sensitive areas. In the land acquisition and compensation calculation dimension, based on the feature category and area information in the feature data, the latest local land acquisition and compensation standards are matched to determine the unit compensation amount for different categories of features, and then the land acquisition and compensation amount for each feature is calculated. In the topographic attribute matching dimension, the center point coordinates of the features are calculated and spatially overlaid with topographic data (such as digital elevation models) to obtain topographic attribute information such as altitude, slope, and aspect of the feature's location, providing a basis for route longitudinal profile design and construction difficulty assessment. In the ecologically sensitive area determination dimension, a vector data intersection algorithm is used to overlay the polygonal boundary of the feature with the vector boundary of the ecologically sensitive area. If there is an intersection and the intersection area reaches a set threshold, the feature is determined to be located within an ecologically sensitive area, providing a basis for subsequent ecological impact assessment.

[0036] Furthermore, the fusion dataset is constructed by integrating various data sources after multi-dimensional correlation processing, unifying the data format and storage structure, and building a multi-source fusion dataset. During the integration process, it is crucial to ensure that the attribute fields of different dimensions of data correspond one-to-one. For example, information such as land acquisition and demolition fees, terrain attributes, and ecologically sensitive area affiliation are linked and bound to basic information in the feature data, such as feature ID, category, and station number, forming a complete attribute record for each feature. The fusion dataset must contain key information required for the full lifecycle management of features, facilitating data retrieval and analysis during subsequent multi-objective constraint modeling and route selection optimization.

[0037] Furthermore, the construction and weight determination of the multi-objective constraint function are as follows: First, the core objective of the multi-objective constraint function is defined, namely, minimizing construction costs, land acquisition and demolition costs, and ecological impact, while maximizing the route compliance rate. The meaning and calculation logic of each parameter in the function are clarified. Construction costs encompass the cost of roadbed, bridge and tunnel, and pavement engineering, calculated according to highway engineering cost estimation rules. Land acquisition and demolition costs are the sum of the land acquisition and demolition amounts for all features. The ecological impact index is calculated based on the proportion of land features within ecologically sensitive areas, according to preset quantification rules. The route compliance rate is the ratio of the number of indicators conforming to design specifications to the total number of design indicators. Design indicators include key technical parameters such as horizontal curve radius, longitudinal slope, and sight distance. The expression for the comprehensive constraint function is: ;in, Construction costs, including roadbed, bridge and tunnel, and pavement costs, are calculated according to highway engineering cost estimation rules. The cost of land acquisition and demolition is the total amount of land acquisition and demolition for all land features. The amount of land acquisition and demolition is calculated based on the area of ​​the land features and the local land acquisition and demolition compensation standards. The ecological impact index is calculated according to preset rules based on the proportion of land features within the ecologically sensitive area. The route compliance rate is the ratio of the number of compliant design indicators to the total number of design indicators. Design indicators include horizontal curve radius, longitudinal slope, and sight distance. , , , For the weighting coefficients, satisfying This is used to characterize the importance of different targets in line selection.

[0038] Furthermore, the weight coefficients of each objective are determined using the analytic hierarchy process (AHP). The specific process is as follows: First, a weight knowledge base containing multi-terrain highway projects is constructed. This knowledge base must include historical data of completed highway projects within the organization, covering weight configuration experience for projects of different terrain types and road grades. After inputting key information such as the terrain type and road grade of the new project, the cosine similarity algorithm is used to match the most similar historical projects in the knowledge base based on terrain features and road grades to obtain initial weights. Then, considering the ecological protection requirements and land acquisition difficulties of the new project, the initial weights are adjusted and optimized through the judgment matrix construction and consistency check of the AHP, ultimately determining weight coefficients that conform to the actual scenario of the new project.

[0039] S400 3D Line Selection Optimization Based on Deep Reinforcement Learning

[0040] The reinforcement learning framework is constructed as follows: First, the state space of the reinforcement learning is defined. This space must comprehensively cover the key influencing factors in the route selection process, including the current location information of the route, the terrain features of the area (such as altitude and slope), the distribution characteristics of land features (such as land feature categories and distance from sensitive areas), cost characteristics (such as construction cost per unit length and land acquisition and demolition costs), and compliance characteristics (such as the compliance status of various design indicators). This ensures that the state can accurately describe the current environment and engineering conditions of the route selection. Second, the action space is defined, specifying the route selection adjustment actions that the agent can execute. These mainly include lateral offset adjustments (to avoid undesirable land features or sensitive areas) and elevation adjustments of longitudinal profile slope change points (to adapt to terrain changes and meet longitudinal slope specifications). At the same time, constraints must be set for the actions, such as lateral offsets needing to avoid ecologically sensitive areas and horizontal curve radii needing to meet design specifications, to prevent actions from exceeding the engineering feasibility range. Finally, based on the multi-objective constraint function constructed above, a value function is built. The calculation result of the constraint function is used as the core input of the value function, so that the value function can quantitatively evaluate the benefits after each action is executed, that is, the comprehensive impact of the action on cost control, ecological protection and compliance improvement, and provide a decision-making basis for the action selection of the intelligent agent.

[0041] Furthermore, phased iterative training and route selection optimization are employed: a phased optimization strategy is used to improve the efficiency and accuracy of route selection. In the coarse selection stage, the state space is simplified based on low-precision terrain data (such as low-resolution digital elevation models), reducing computational load and quickly filtering out candidate corridors that meet basic ecological compliance requirements (such as avoiding core sensitive areas), narrowing the search range for subsequent fine optimization and avoiding inefficiency caused by ineffective searches. In the fine selection stage, high-precision orthophotos and detailed survey data are loaded into the candidate corridors, enabling the full state space to ensure that the optimization process can fully consider detailed terrain and feature information, thereby improving route selection accuracy. During iterative training, an experience cache pool (used to store state-action-reward-next state samples during training) and dual neural networks (target neural network and valuation neural network) are initialized first, with the parameters of the two networks remaining consistent in the initial state. An ε-greedy strategy is used to select actions, i.e., actions are randomly selected with a partial probability to explore new route selection possibilities, and actions with the highest value function value are selected with a partial probability to utilize the current best experience, balancing the relationship between exploration and utilization. When the number of samples in the experience buffer pool reaches a set threshold, a certain number of samples are drawn from the buffer pool each time (prioritizing samples from candidate corridors to improve sample relevance) for network training. The mean squared error loss of the estimating neural network is calculated, and the parameters of the estimating neural network are updated using the gradient descent algorithm. The parameters of the estimating neural network are periodically synchronized to the target neural network to ensure the stability of the target network. Training continues iteratively. When the fluctuation range of the constraint function value corresponding to the optimal route solution meets the stability requirements of the solution (i.e., the performance of the solution no longer changes significantly) during a preset number of iterations, training stops, and multiple candidate route solutions that meet engineering requirements are generated.

[0042] S500: Intelligent Decision-Making for Highway Route Selection

[0043] Candidate scheme comparison: A comprehensive scoring formula based on four core indicators—construction cost, land acquisition and demolition cost, ecological impact, and compliance rate—is used to determine the preliminary optimal scheme. The comprehensive scoring formula is as follows: ,in , , These are the maximum values ​​of construction cost, land acquisition and demolition cost, and ecological impact index among all candidate schemes. The weights (0.4, 0.3, 0.2, 0.1) in the formula represent the contribution of each indicator to the comprehensive score. By calculating the comprehensive score of each candidate scheme, the scheme with the highest score is determined as the preliminary optimal scheme.

[0044] If the score difference between the preliminary optimal solution and the second-best solution is less than a set threshold, it indicates that the two solutions have little difference in overall benefits. In this case, it is necessary to further calculate the maximum curvature value of the route alignment safety index for judgment. The curvature calculation method is as follows: Discretize the route alignment into a series of continuous coordinate points, calculate the curvature value of each point based on these continuous coordinate points, compare the curvature values ​​of all coordinate points, and find the maximum value as the maximum curvature. The curvature calculation formula is: ,in These are the coordinates of three consecutive points on the curve; iterate through all points. Find the maximum value among them. The smaller the maximum curvature, the smoother the route alignment and the higher the driving safety. Therefore, the scheme with the smaller maximum curvature is selected as the final scheme to ensure that the final scheme meets both economic and ecological requirements and has excellent driving safety performance.

[0045] Example 2

[0046] In the route selection and design of a lakeside hilly highway project, which is to Class I highway standards, the route passes through low hilly areas, river valley farmland, and a provincial nature reserve experimental zone. It is necessary to balance construction economics, land acquisition and resettlement costs, and ecological protection needs, while also meeting route compliance requirements and driving safety requirements. Please refer to [link / reference needed]. Figure 2 The image shown is a schematic diagram of the line selection effect.

[0047] Step 1: Multi-source data acquisition and standardized preprocessing

[0048] Collect multi-source data from the project area and perform standardization processing:

[0049] Image data: Orthophotos are generated by stitching together, and then subjected to Gaussian filtering for noise reduction, histogram equalization for enhancement, and weighted average grayscale conversion to meet the accuracy requirements for ground feature identification;

[0050] Vector data: converted to SHP format, covering land use, ecologically sensitive areas, and existing road data;

[0051] Engineering design data: Export preliminary route alignment, bridge and tunnel plans, and other CAD format data, and associate them with route stationings. The preprocessed data quality meets the following requirements: image signal-to-noise ratio ≥30dB, vector coordinate error ≤0.1m, providing highly reliable input for subsequent feature identification.

[0052] Step 2: Intelligent identification and feature extraction of ground features

[0053] Training set construction: 2000 256×256 pixel samples were cropped from the preprocessed images and labeled according to highway engineering requirements using the LabelMe annotation tool, covering 8 core land features:

[0054] Buildings (152 in total, covering an area of ​​128,000 square meters, including 1,422 square meters of brick and tile houses and 446 square meters of brick and concrete houses).

[0055] Farmland (34.94 mu of paddy fields and 52.59 mu of dry land, totaling 87.53 mu, or approximately 5.83 hectares);

[0056] Ponds (89.19 mu of intensive fish farming ponds, approximately 5.95 hectares);

[0057] Forest land (721.78 mu, approximately 48.12 hectares, including 28 hectares of forest land within ecologically sensitive areas).

[0058] Existing roads (1.1km of expressway section);

[0059] Irrigation canal (7.71 mu, approximately 0.51 hectares);

[0060] Wasteland (93.29 mu, approximately 6.22 hectares of sandy land);

[0061] Ecologically sensitive area (experimental area of ​​provincial nature reserve, covering an area of ​​120 hectares).

[0062] Further, model training and inference: With 16 training batches and an initial learning rate of 0.001, data augmentation such as flipping and rotation was performed. After pre-training, the model's ground feature recognition accuracy reached 92.3%. The optimized images were input into the model, semantic segmentation and inference were used to extract polygonal boundaries of ground features, calculate the area of ​​the ground features and associate them with route station numbers, and output a structured ground feature feature dataset.

[0063] Step 3: Multi-source data fusion and multi-objective constraint modeling

[0064] 3.1 Multi-dimensional data correlation

[0065] Related to land acquisition and demolition costs: Calculated based on land feature type and local land acquisition compensation standards.

[0066] Land acquisition compensation: Paddy fields 46,000 yuan / mu × 34.94 mu = 1,607,200 yuan, forest land 32,200 yuan / mu × 721.78 mu = 23,241,300 yuan, total land acquisition compensation 35,628,000 yuan;

[0067] Compensation for standing crops: 2061 yuan / mu × 812.88 mu = 1,675,300 yuan;

[0068] Compensation for house expropriation: RMB 1,200 / ㎡ for brick and tile houses × 1,422㎡ = RMB 1,706,400; RMB 1,314 / ㎡ for brick and concrete houses × 446㎡ = RMB 586,000; including RMB 934,000 for decoration compensation, the total compensation for house expropriation is RMB 20,224,800.

[0069] Total land acquisition and demolition costs =Land acquisition compensation + crop compensation + house expropriation + electricity and telecommunications compensation = 3562.80 + 167.53 + 2022.48 + 416.62 = 8665.46 million yuan.

[0070] Terrain attribute matching: By overlaying 10m precision DEM data with the coordinates of the center point of the ground features, the slope of the roadbed section, the elevation of the bridge site, and the terrain of the tunnel entrance and exit are obtained.

[0071] Determination of ecologically sensitive areas: Vector intersection calculation shows that 28 hectares of the 721.78 mu of forest land are located within ecologically sensitive areas, accounting for 38.2%.

[0072] 3.2 Construction of Multi-Objective Constraint Functions

[0073] Define the synthesis constraint function as follows:

[0074] ;

[0075] Parameter calculation: Construction cost According to the "Highway Engineering Estimation Index", the estimated costs are as follows: Roadbed: 6.193 km × 1.2 million yuan / km = 7.4316 million yuan; Bridges: 1.486 km (Bridge No. 1: 488m × 8.1143 million yuan / km + Bridge No. 2: 648m × 8.1456 million yuan / km = 98.6923 million yuan); Tunnels: 0.781 km (Tunnel No. 1: 344.7m × 2.1775 million yuan / m + Tunnel No. 2: 436m × 2.4767 million yuan / m = 183.0424 million yuan); Pavement: 8.46 km × 8.3747 million yuan / km = 70.8501 million yuan. Total... =644.3355 million yuan;

[0076] Ecological Impact Index Based on the ratio of "land area within the ecologically sensitive area / total project land area", the total project land area is 1237.6 mu (approximately 82.51 hectares), of which 28 hectares are within the ecologically sensitive area. =28 / 82.51≈0.339;

[0077] Route compliance rate The test covered 5 core indicators (horizontal curve radius, longitudinal slope, sight distance, roadbed width, and bridge / tunnel clearance), and 4 of them met the requirements. =4 / 5=0.8.

[0078] Furthermore, the weights are determined:

[0079] Construct a weighted knowledge base: incorporate historical data from 10 mountain road projects within the organization;

[0080] Cosine similarity matching: The new project's terrain (low hills) and road grade (Level 1) have a similarity of 0.88 with a certain level of highway in the knowledge base, and the initial weights are obtained as α=0.32, β=0.23, γ=0.30, and δ=0.15.

[0081] Adjustment of the analytic hierarchy process: Due to the high proportion of ecologically sensitive areas in this project (38.2%), γ is increased to 0.28, and the final weights are α=0.32, β=0.23, γ=0.28, δ=0.17 (α+β+γ+δ=1).

[0082] Constraint function calculation:

[0083] minF=0.32×64433.55+0.23×8665.46+0.28×0.339-0.17×0.8=22644.57.

[0084] Step 4: 3D Line Selection Optimization Based on Deep Reinforcement Learning

[0085] 4.1 Building a Reinforcement Learning Framework

[0086] State space: includes route location (X / Y / Z coordinates), terrain features (slope, altitude), land feature features (distance from sensitive areas, land feature type), cost features (construction / land acquisition cost per unit length), and compliance features (compliance status of indicators).

[0087] Operational space: horizontal offset (±5m, to avoid ecologically sensitive areas), elevation adjustment of longitudinal profile slope change points (±0.5m, to meet longitudinal slope ≤5%).

[0088] Value function: (ω is the normalization coefficient, making the value function take values ​​in the range [0,100]).

[0089] 4.2 Phased Optimization Training

[0090] Coarse selection stage: Based on a simplified state space using a 30m precision DEM, three candidate corridors were quickly selected (avoiding core sensitive areas), reducing the search range from 8.46km×3km to 8.46km×0.5km;

[0091] Selection phase: Load 0.5m orthophoto and detailed ground feature data, enable the full state space, initialize the experience cache pool (capacity 100,000 samples) and dual neural networks, and use the ε-greedy strategy (ε=0.1) to select actions;

[0092] Iteration stopping condition: In 50 consecutive iterations, the value of the constraint function fluctuates by ≤0.5%, and two candidate solutions (Solution 1 and Solution 2) are output.

[0093] Step 5: Comparison of candidate solutions

[0094] The maximum value of the indicator is: , , =0.35;

[0095] Score for Option 1: ;

[0096] Score for Option 2: ;

[0097] Furthermore, the preliminary optimal solution is Solution 1 (score difference 14.36 > threshold 5%).

[0098] Curvature verification (curvature comparison is not required due to score difference exceeding the limit)

[0099] Scheme 1: Discretize the planar linear shape into 1000 coordinate points and calculate the maximum curvature. (The horizontal curve radius is 320m, which meets the requirement of ≥200m for Class I highways), the alignment is gentle, and the driving safety is excellent.

[0100] The final optimal solution is solution 1.

[0101] Example 3

[0102] The route is 4.3km long, with a design speed of 60km / h, and is built to first-class highway standards. It passes through river valleys, plains, and low hills, and has no tunnels. It includes one 50m medium-sized bridge. The amount of land acquisition and demolition needs to be controlled, and basic farmland must be avoided.

[0103] Multi-source data association

[0104] Construction costs: Roadbed engineering 4.18 million yuan, medium bridge engineering 4.75 million yuan, pavement engineering 34.4 million yuan. ;

[0105] Land acquisition and demolition costs: Basic farmland compensation of 12.568 million yuan (240 mu × 46,000 yuan / mu), house expropriation of 1.44 million yuan (1200m² × 1200 yuan / m²). .

[0106] Ecological impact: No sensitive areas. E Life = 0.05;

[0107] Furthermore, the multi-objective constraint function is calculated as follows:

[0108] ;

[0109] (Weight matching for plains area projects:) α =0.40, β =0.20, γ =0.15, δ =0.25)

[0110] Furthermore, the selection of alternatives

[0111] Option A (near the village): , , E The yield was 0.05, indicating a compliance rate of 100%.

[0112] Option B (Avoiding Villages): , , E The yield was 0.06, and the compliance rate was 100%.

[0113] Furthermore, the overall score is: Scheme A = 18.72, Scheme B = 12.32, the difference is 6.4, which is greater than the preset threshold of 5%, so Scheme A is selected as the optimal scheme.

[0114] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A decision-making method for highway alignment design schemes based on multi-source fusion, characterized in that, include: Multi-source data acquisition and standardized preprocessing: Collect basic geographic data of the highway planning area, including image data, vector data and design data, and perform format standardization and preprocessing on the collected data to generate standardized data in a unified format; Ground feature identification and feature extraction: Based on the preprocessed data, a ground feature training set is constructed, and an improved target detection model is used to identify ground features and extract features, outputting ground feature feature data; The improved target detection model involves building a ground feature recognition network, extracting the network structure, and enhancing the ability to recognize small-sized ground features and ground features in complex backgrounds. Multi-source data fusion and multi-objective constraint modeling: Associate multi-dimensional data, including: land feature data, land acquisition and demolition compensation standards, topographic data, and vector data of ecologically sensitive areas; calculate land acquisition and demolition amounts for multi-dimensional data, obtain topographic attributes and determine ecologically sensitive areas; define a comprehensive constraint function with the objectives of minimizing construction costs, land acquisition and demolition costs, ecological impact, and maximizing route compliance rate, and use the analytic hierarchy process (AHP) to determine the weight coefficients of each objective; 3D route selection optimization based on deep reinforcement learning: a deep reinforcement learning framework is built, the state space and action space are defined, a value function is constructed based on a comprehensive constraint function, and candidate route schemes are generated through iterative training; The deep reinforcement learning framework includes: a phased optimization strategy; a coarse selection phase that simplifies the state space based on low-precision DEM data to quickly filter candidate corridors that meet ecological compliance requirements and narrow the search range; a fine selection phase that loads high-precision orthophotos and detailed ground feature data into the candidate corridors and enables the complete state space; and an action space that includes lateral offset of the horizontal route and elevation adjustment of slope change points in the longitudinal profile, wherein the lateral offset of the horizontal route meets the constraints of avoiding ecologically sensitive areas and ensuring that the radius of the horizontal curve meets the design specifications; and when the distance between the route and the ecologically sensitive area reaches a preset proximity threshold, the weight coefficient of the ecological impact target is adjusted. Candidate scheme comparison: Based on construction costs, land acquisition and demolition costs, ecological impact and compliance rate of route specifications, a comprehensive scoring formula is constructed to calculate the preliminary optimal scheme. If the score difference between the preliminary optimal and the second-best scheme is less than a set threshold, the final scheme is selected by comparing the maximum curvature.

2. The highway alignment design scheme decision-making method based on multi-source fusion according to claim 1, characterized in that, In the land feature recognition and feature extraction, the target detection model construction includes: labeling the cropped image with land features, and the labeled land feature categories at least cover one or more of the following: buildings, farmland, ponds, woodlands, roads, canals, wasteland, and ecologically sensitive areas.

3. The highway alignment design scheme decision-making method based on multi-source fusion according to claim 1, characterized in that, In the multi-source data fusion and multi-objective constraint modeling, the expression for the multi-objective constraint function is: ;in, Construction costs, including roadbed, bridge and tunnel, and pavement costs, are calculated according to highway engineering cost estimation rules. The cost of land acquisition and demolition is the total amount of land acquisition and demolition for all land features. The amount of land acquisition and demolition is calculated based on the area of ​​the land features and the local land acquisition and demolition compensation standards. The ecological impact index is calculated according to preset rules based on the proportion of land features within the ecologically sensitive area. The route compliance rate is the ratio of the number of compliant design indicators to the total number of design indicators. Design indicators include horizontal curve radius, longitudinal slope, and sight distance. , , , The weighting coefficients are determined by the scenario: a weighted knowledge base containing multi-terrain highway projects is constructed. The weighted knowledge base needs to include historical data of highway projects completed by the unit. After inputting the terrain type of a new project, the initial weights of the projects with the most similar terrain features and road grades in the knowledge base are matched by the cosine similarity algorithm. Then, the weights are adjusted by combining the ecological expropriation and demolition requirements of the new project with the analytic hierarchy process.

4. The highway alignment design scheme decision-making method based on multi-source fusion according to claim 1, characterized in that, The iterative training generates candidate route schemes including: Initialize an experience cache pool to meet the sample storage requirements, and initialize the target neural network and the valuation neural network, keeping the initial parameters of the target neural network consistent with those of the valuation neural network. Employ an ε-greedy strategy to select actions, randomly choosing actions with ε probability and selecting the action with the highest value function with 1-ε probability. When the number of samples in the experience cache pool reaches a set threshold, extract the number of samples required for each network update, calculate the MSE loss of the valuation neural network, and update the parameters of the valuation neural network. Synchronize the parameters of the valuation neural network to the target neural network. During training, prioritize sampling samples within candidate corridors to perform batch sampling optimization. When the fluctuation range of the constraint function value corresponding to the optimal route solution meets the stability requirements of the solution in consecutive preset iterations, stop training and output the candidate route solution.

5. The highway alignment design scheme decision-making method based on multi-source fusion according to claim 1, characterized in that, The preliminary optimal solution is calculated by using a comprehensive scoring formula to calculate the score of the candidate solutions. The formula is as follows: ,in , , The maximum values ​​of construction cost, land acquisition and demolition cost, and ecological impact index among all candidate schemes are respectively used, and the scheme with the highest score is determined as the preliminary optimal scheme.

Images