Pavement maintenance intelligent decision-making method and system based on deep learning and GIS (Geographic Information System)

Abstract

The invention relates to a pavement maintenance intelligent decision-making method and system based on deep learning and GIS, and belongs to the field of artificial intelligence traffic science. The method comprises the following steps: collecting road surface images and road geographic data, constructing a GIS image data set GISVision after manual labeling, preprocessing the GIS image data set GISVision, and dividing a training set, a verification set and a test set; inputting road image data into an image feature extractor, inputting road geographic data into a GIS feature encoder, and respectively obtaining image features and GIS encoding features; the two images are processed by a geographic alignment image encoder to obtain joint features, the image features and the joint features are added and then input into a disease form modeling device to obtain disease modeling features, and then the disease modeling features are respectively input into a road risk grade scoring device and a maintenance visualization generator to obtain predicted risk grades and maintenance visualization images; and constructing total loss function training, iterating parameters by using an Adam optimizer, finally inputting to-be-predicted data into the trained model, and combining the maintenance prediction condition of the GIS pavement maintenance scientific decision system. According to the invention, the accuracy of road disease identification can be improved, and scientific early warning and decision making are realized.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence in transportation, specifically relating to an intelligent decision-making method and system for road maintenance based on deep learning and GIS. Background Technology

[0002] Against the backdrop of rapid urbanization, the scale of road infrastructure construction is constantly expanding, and the demand for road maintenance is also increasing. Traditional road maintenance work mainly relies on manual inspections and periodic checks, which is not only inefficient but also easily affected by subjective human factors, resulting in some road defects not being detected and treated in a timely manner, seriously affecting traffic safety and municipal operational efficiency. At the same time, with the popularization of sensing devices such as drones, lidar, and cameras, road data collection has become more efficient and automated. However, how to effectively utilize massive amounts of image and spatial data to achieve high-precision, intelligent road condition assessment and maintenance decisions remains a technical bottleneck for the industry.

[0003] In recent years, deep learning has made groundbreaking progress in computer vision tasks such as image recognition, object detection, and semantic segmentation. Its powerful feature extraction and pattern recognition capabilities have provided new solutions for the automatic identification and classification of road surface defects. However, deep learning models typically focus only on image information and lack an understanding of geospatial attributes, making it difficult to achieve global optimization decisions based on spatial location. Geographic Information Systems (GIS), as a powerful spatial data management and analysis tool, can efficiently process multi-source data related to road maintenance, such as spatial distribution, topography, traffic flow, and historical maintenance records, providing technical support for realizing full lifecycle management of roads.

[0004] Therefore, integrating deep learning with GIS technology to build a comprehensive decision-making system with intelligent image recognition and spatial information processing capabilities has become an important direction for improving the intelligence level of road maintenance. By using deep learning models to identify and classify road surface defects in images, and combining this with a GIS platform to comprehensively analyze factors such as the spatial distribution characteristics, affected areas, and traffic priorities of these defects, more scientific, efficient, and refined road maintenance decisions can be achieved. This integrated technical approach not only improves the automation and accuracy of road condition detection but also helps optimize resource allocation, reduce maintenance costs, and provides solid data support and technical assurance for smart city construction. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a method and system for intelligent decision-making in road maintenance based on deep learning and GIS.

[0006] To achieve the above objectives, the present invention employs the following technical solution: This invention provides an intelligent decision-making method for road maintenance based on deep learning and GIS, comprising the following steps: S1. Collect road surface images and corresponding road geographic data, perform manual annotation, and construct the GIS image dataset GISVision; S2. Preprocess the GISVision dataset to obtain road image data. and road geographic data It is divided into training set, validation set and test set; S3. Road image data The image features are input into the image feature extractor to obtain the image features. ; road geographic data The input is fed into the GIS feature encoder to obtain the GIS coded features. The image feature extractor includes convolutional layers, pooling layers, and fully connected layers; the GIS feature encoder includes a fully connected layer and an output layer, with the output layer undergoing L2 normalization. Utilizing the image feature extractor and GIS feature encoder to achieve simultaneous multimodal data processing improves the comprehensiveness and accuracy of disease identification, effectively avoiding the misjudgment and missed judgment problems that may arise from traditional methods relying solely on image data.

[0007] S4. Image Features and GIS coding features After processing by the geographic alignment image encoder, joint features are obtained. Image features and joint features The sums are then input into the disease morphology modeler to obtain the disease modeling features. ; S5. Modeling features of diseases The data is input into the road risk level scorer to obtain the predicted risk level. ; Modeling features of diseases and GIS coding features The sums are then input into the maintenance visualization generator to obtain the predicted maintenance visualization image. ; S6. Constructing a total loss function to measure risk level Maintenance visualization images The differences between manual annotation and training were analyzed in steps S3-S5. The Adam optimizer was used to iterate the parameters through backpropagation, and finally the trained model was obtained. S7. After the road images and geographic data to be predicted are preprocessed in step S2, they are input into the trained model to obtain the predicted risk level and maintenance visualization image. S8. Input the predicted risk level and maintenance visualization image of each road segment into the GIS road maintenance scientific decision-making system to obtain the maintenance prediction of each road segment on the map.

[0008] Furthermore, step S2 specifically includes: S21. Unify and normalize the image sizes in the GISVision dataset, scale the RGB three-channel pixel values ​​to the [0,1] range, and use the CLAHE method to process dark areas to obtain road image data. ; S22. Normalize the road geographic data. Numerical fields are scaled to the [0,1] range using min-max scaling, while categorical fields maintain one-hot encoding. Pair each image data with its corresponding road geographic data using a unique index to form structured road geographic data. .

[0009] Furthermore, the geographic alignment image encoder includes an image feature linear mapping layer, a GIS feature linear mapping layer, a location encoder, and a fusion projection layer; the image feature linear mapping layer, the GIS feature linear mapping layer, and the fusion projection layer are all fully connected layers, and the ReLU activation function is used for all of them. Image features Image linear mapping features are obtained after the image feature linear mapping layer. GIS coding features GIS linear mapping features are obtained after the GIS feature linear mapping layer. ; The image linear mapping feature and GIS linear mapping features The input is fed into the position encoder to obtain the fixed position code. The position encoder uses a sinusoidal position encoding formula to encode fixed positions. Linear mapping features of the image respectively and GIS linear mapping features By adding element by element, the image location code is obtained. and GIS location coding ; Image position encoding and GIS location coding The features are summed along the channel dimension and then fed into the fusion projection layer to obtain the joint features. .

[0010] Furthermore, by extracting deeper features of diseases through the residual module, the ability to judge different disease types and severity is enhanced, ensuring the accuracy of risk level assessment. The disease morphology modeler includes a normalization layer, a residual morphology modeling module, and an output layer; the residual morphology modeling module includes a linear layer, a Dropout layer, and a normalization layer; the output layer is a linear projection layer. The image features and joint features After element-wise addition, the data is processed by L2 normalization through a normalization layer to obtain normalized features. Normalization characteristics After residual morphology module construction and output layer processing, the disease modeling features are obtained. , , Representing disease modeling features The i-th dimension feature.

[0011] Furthermore, a risk level scoring mechanism is designed to quantify road damage risks for scientific early warning and decision-making. The system automatically assesses the risk level of road sections and formulates maintenance plans, improving the objectivity and consistency of decision-making. The aforementioned road damage modeling features... The data is input into a road risk level scorer for prediction, yielding the predicted risk level probability. Select the probability of all risk levels The category with the highest probability is used as the predicted risk level. The formula for the scoring function is expressed as follows: in, Representing disease modeling features The i-th dimension feature; exp represents an exponential function with base e; | | indicates to Take the absolute value; β represents the balance constant; γ represents the gain coefficient; λ represents the regularization term constant; This indicates the probability of predicting a risk level of c; C represents the number of categories.

[0012] Furthermore, a deconvolutional network is used to transform complex prediction results into intuitive images, helping maintenance personnel to efficiently execute tasks and remotely monitor and schedule operations. The maintenance visualization generator includes a feature fusion and normalization layer, a feature expansion mapping module, and a visualization image reconstruction module; the feature expansion mapping module includes a fully connected layer and a reconstruction mapping layer; the reconstruction mapping layer reconstructs the features; the visualization image reconstruction module is a deconvolutional layer. Disease modeling features and GIS coding features After element-wise addition, the input is fed into a feature fusion normalization layer for normalization processing, resulting in fused and normalized features. ; Integration and normalization features The input is fed into the feature expansion mapping module and the visualization image reconstruction module to obtain the predicted maintenance visualization image. .

[0013] Furthermore, the total loss function is the sum of the image loss function and the risk level loss function; the formula for calculating the image loss function is: , in, Visualization of maintenance The value of the i-th pixel; This represents the label of the i-th pixel in a manually labeled image; The constant represents the risk level loss function, which is calculated as follows: , Where N represents the number of samples; Indicates risk level The value, This indicates the true risk level.

[0014] This invention also provides an intelligent decision-making system for road maintenance based on deep learning and GIS, including... Data acquisition module: used to collect road surface images and corresponding road geographic data, perform manual annotation, and construct the GIS image dataset GISVision; Data preprocessing module: used to preprocess the GISVision dataset to obtain road image data and road geographic data, and divide them into training set, validation set and test set; Feature extraction module: Used to input road image data into the image feature extractor to obtain image features; input road geographic data into the GIS feature encoder to obtain GIS encoded features; Feature alignment module: This module processes image features and GIS encoded features through a geographic alignment image encoder to obtain joint features; the image features and joint features are then added together and input into the lesion morphology modeler to obtain lesion modeling features. Risk level and maintenance visualization image prediction module: used to input the road disease modeling features into the road risk level scorer to obtain the predicted risk level; and to add the disease modeling features and GIS coding features and input them into the maintenance visualization generator to obtain the predicted maintenance visualization image. Model optimization module: Constructs a total loss function to measure the risk level, the difference between maintenance visualization images and manual annotations, trains the model, and uses the Adam optimizer to iterate the parameters through backpropagation to finally obtain the trained model; Detection module: This module is used to preprocess the road images and geographic data to be predicted, and then input them into the trained model to obtain the predicted risk level and maintenance visualization image. Maintenance prediction output module: This module is used to input the predicted risk level and maintenance visualization image of each road segment into the GIS road maintenance scientific decision-making system to obtain the maintenance prediction status of each road segment on the map.

[0015] The advantages of this invention are: This invention deeply integrates deep learning and GIS technologies to construct a GISVision dataset containing road surface images and geographic information. It utilizes an image feature extractor and a GIS feature encoder to achieve simultaneous multimodal data processing, improving the comprehensiveness and accuracy of road damage identification and effectively avoiding the misjudgments and omissions that may arise from traditional methods relying solely on image data. Simultaneously, it extracts deep-level features of road damage through a residual module, enhancing the ability to judge different types and severity of damage and ensuring the accuracy of risk level assessment. Furthermore, a risk level scoring mechanism is designed to quantify road damage risks for scientific early warning and decision-making. The system automatically assesses the risk level of road sections and formulates maintenance plans, improving the objectivity and consistency of decision-making. The maintenance visualization image generation function transforms complex prediction results into intuitive images through a deconvolutional network, assisting maintenance personnel in efficiently executing tasks and remotely monitoring and scheduling. Finally, the GIS map visualization of decision results enables location-based maintenance decision-making, mapping road section risk levels and maintenance needs to a GIS map, optimizing resource allocation and maintenance plans, and improving overall management efficiency. This provides an intelligent and automated solution for road maintenance management and has broad application prospects. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0017] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is the original image of the present invention; Figure 3 These are maintenance visualization images for the present invention; Figure 4 The road segment decision results obtained by the decision system of this invention. Detailed Implementation

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

[0019] Example 1 This embodiment provides a road maintenance intelligent decision-making method based on deep learning and GIS, such as... Figure 1 As shown, the specific steps include: S1. Collect road surface images and corresponding road geographic data, and manually annotate them to construct the GIS image dataset GISVision.

[0020] S2. Preprocess the GISVision dataset to obtain road image data. and road geographic data It is divided into training set, validation set and test set.

[0021] Specifically, in step S21, the image size in the GISVision dataset is unified to 512×512 pixels and normalized. The RGB three-channel pixel values ​​are scaled to the [0,1] range, and the CLAHE method is used to process the dark areas to obtain road image data. ; S22. Normalize the road geographic data. Numerical fields are scaled to the [0,1] range using min-max scaling, while categorical fields maintain one-hot encoding. Pair each image data with its corresponding road geographic data using a unique index to form structured road geographic data. .

[0022] S3. Road image data The image features are input into the image feature extractor to obtain the image features. ; road geographic data The input is fed into the GIS feature encoder to obtain the GIS coded features. .

[0023] Specifically, the image feature extractor includes convolutional layer 1, pooling layer 1, convolutional layer 2, pooling layer 2, convolutional layer 3, pooling layer 3, convolutional layer 4, pooling layer 4, and a fully connected layer; the convolutional layers 1, 2, 3, and 4 all have a kernel size of 3×3, a stride of 1, padding of 1, and use the ReLU activation function, with the number of kernels being 32, 64, 128, and 256 respectively; the pooling layers 1, 2, 3, and 4 all have a kernel size of 2×2, a stride of 2, and use max pooling; the fully connected layer has an output size of 2048 and uses the ReLU activation function. Road image data The data is sequentially input into convolutional layer 1, pooling layer 1, convolutional layer 2, pooling layer 2, convolutional layer 3, pooling layer 3, convolutional layer 4, pooling layer 4, and a fully connected layer to obtain image features. .

[0024] The GIS feature encoder includes a fully connected layer 1, a fully connected layer 2, a fully connected layer 3, a fully connected layer 4, and an output layer; the output dimensions of the fully connected layers 1, 2, 3, and 4 are 128, 256, 512, and 1024, respectively, and the ReLU activation function is used for all of them; The road geographic data The data is processed sequentially through fully connected layers 1, 2, 3, and 4, resulting in the output of fully connected layer 4. The output of fully connected layer 4 then undergoes L2 normalization in the output layer to obtain the GIS coded features. .

[0025] S4. Image Features and GIS coding features After processing by the geographic alignment image encoder, joint features are obtained. Image features and joint features The sums are then input into the disease morphology modeler to obtain the disease modeling features. .

[0026] Specifically, the geographic alignment image encoder includes an image feature linear mapping layer, a GIS feature linear mapping layer, a location encoder, and a fusion projection layer; the image feature linear mapping layer and the GIS feature linear mapping layer are both fully connected layers with an output dimension of 512 and both use the ReLU activation function; the fusion projection layer is a fully connected layer with an input dimension of 512 and an output dimension of 2048, and both use the ReLU activation function. Image features Image linear mapping features are obtained after the image feature linear mapping layer. GIS coding features GIS linear mapping features are obtained after the GIS feature linear mapping layer. ; The image linear mapping feature and GIS linear mapping features The input is fed into the position encoder to obtain the fixed position code. The position encoder uses a sinusoidal position encoding formula, which is expressed as follows: , in, Indicating the position code of the first Line number Values ​​in each dimension; Indicates the first Line number Values ​​in each dimension; The total dimension for location encoding is set to 512; It is a position index; fixed position encoding. For a two-row matrix, when =0 indicates a linear mapping feature in the image. Used to calculate each element in the first row, when =1 indicates a linear mapping feature in GIS. Used to calculate each element in the second row; i is the dimension index, with a value range of i∈{0,1,...,255}; Encoding fixed positions. Linear mapping features of the image respectively and GIS linear mapping features By adding element by element, the image location code is obtained. and GIS location coding ; Image position encoding and GIS location coding The features are summed along the channel dimension and then fed into the fusion projection layer to obtain the joint features. .

[0027] The disease morphology modeler includes a normalization layer 1, a residual morphology modeling module 1, a residual morphology modeling module 2, and an output layer. The residual morphology modeling module 1 includes a linear layer 1, a dropout layer 1, a linear layer 2, and a normalization layer 2. The residual morphology modeling module 2 has the same structure as the residual morphology modeling module 1. The output dimensions of the normalization layer 1 and normalization layer 2 are 2048. The input dimension of the linear layer 1 is 2048, and the output dimension is 1024. The dropout rate of the dropout layer 1 is 0.1. The input dimension of the linear layer 2 is 1024, and the output dimension is 2048. The output layer is a linear projection layer with an input dimension of 2048 and an output dimension of 1024, and the activation function is ReLU. The image features and joint features After element-wise addition, the data is subjected to L2 normalization through normalization layer 1 to obtain normalized features. Normalization characteristics The inputs are processed in residual morphology building module 1, residual morphology building module 2, and the output layer to obtain the disease modeling features. , , Representing disease modeling features The i-th dimension feature.

[0028] S5. Modeling features of diseases The data is input into the road risk level scorer to obtain the predicted risk level. ; Modeling features of diseases and GIS coding features The sums are then input into the maintenance visualization generator to obtain the predicted maintenance visualization image. .

[0029] Specifically, the disease modeling features The data is input into a road risk level scorer for prediction, yielding the predicted risk level probability. Select the probability of all risk levels The category with the highest probability is used as the predicted risk level. The formula for the scoring function is expressed as follows: in, Representing disease modeling features The i-th dimension feature; exp represents an exponential function with base e; | | indicates to Take the absolute value; β represents the balance constant; γ represents the gain coefficient; λ represents the regularization term constant; This represents the probability of predicting a risk level of c; C represents the number of categories.

[0030] Disease modeling features and GIS coding features The sums are then input into the maintenance visualization generator to obtain the predicted maintenance visualization image. Specifically: The maintenance visualization generator includes a feature fusion normalization layer, a feature expansion mapping module, and a visualization image reconstruction module; the feature expansion mapping module includes a fully connected layer 5 and a reconstruction mapping layer; the fully connected layer 5 has an input dimension of 1024, an output dimension of 4096, and an activation function of ReLU; the reconstruction mapping layer reconstructs the 4096-dimensional features into a three-dimensional tensor with 64 channels and a spatial resolution of 8×8; The visualization image reconstruction module includes deconvolution layer 1, deconvolution layer 2, and deconvolution layer 3; each of the three deconvolution layers uses a 4×4 convolution kernel with a stride of 2 and padding of 1, with 64, 32, and 16 input channels, 32, 16, and 1 output channel, respectively, and uses the ReLU activation function. Disease modeling features and GIS coding features After element-wise addition, the input is fed into a feature fusion and normalization layer for normalization, resulting in fused and normalized features with a dimension of 1024. ; Integration and normalization features The input is fed into the feature expansion mapping module and the visualization image reconstruction module to obtain the predicted maintenance visualization image. .

[0031] S6. Constructing a total loss function to measure risk level Maintenance visualization images The differences between manual annotation and training were analyzed in steps S3-S5. The Adam optimizer was used to iterate the parameters through backpropagation, and finally the trained model was obtained. Specifically, the total loss function is the sum of the image loss function and the risk level loss function; the formula for calculating the image loss function is: , in, Visualization of maintenance The value of the i-th pixel; This represents the label of the i-th pixel in a manually labeled image; The constant represents the risk level loss function, which is calculated as follows: , Where N represents the number of samples; Indicates risk level The value, This indicates the true risk level.

[0032] S7. After preprocessing in step S2, the road images and geographic data to be predicted are input into the trained model to obtain the predicted risk level and maintenance visualization image.

[0033] S8. Input the predicted risk level and maintenance visualization image of each road segment into the GIS road maintenance scientific decision-making system to obtain the maintenance prediction of each road segment on the map.

[0034] Example 2 To demonstrate the effectiveness of this invention in image recognition and maintenance prediction, a comparison between the original image and the maintenance visualization image can be shown.

[0035] like Figure 2 The image shown is a raw, uncollected image, which typically includes road surface defects, but may not be easily identifiable by directly identifying the specific type and severity of the defects; for example... Figure 3 As shown, the maintenance visualization images generated by the method of this invention demonstrate the severity of road surface defects and maintenance requirements.

[0036] Figure 4 This paper showcases a GIS-based road maintenance decision-making system, visually displaying the maintenance status and priorities of different road sections. In the diagram, green-marked sections indicate sections in good condition or already maintained, while red-marked sections indicate areas with more serious defects or damage requiring immediate maintenance. Through the system's intelligent analysis, red arrows indicate sections requiring priority, consistent with the intelligent maintenance decision-making process based on risk level in this invention. Furthermore, the diagram displays specific information for each road section, including section number and start and end points. Based on this information, the system analyzes and predicts the maintenance needs of each section. The system provides corresponding maintenance suggestions for each section based on defect type, risk level, and other data, such as repair methods and required materials. This predictive result helps maintenance personnel promptly identify which road sections require urgent attention and what maintenance measures to take.

[0037] In summary, this map visualizes maintenance decision-making results through a GIS platform, demonstrating how to intelligently assess the maintenance needs of each road segment and optimize the allocation of maintenance resources. It aligns closely with the pavement distress identification, risk assessment, and maintenance decision-making processes in this invention, improving the efficiency and accuracy of pavement maintenance and ensuring timely and effective maintenance measures.

[0038] Example 3 This embodiment provides a road maintenance intelligent decision-making system based on deep learning and GIS, including... Data acquisition module: used to collect road surface images and corresponding road geographic data, perform manual annotation, and construct the GIS image dataset GISVision; Data preprocessing module: used to preprocess the GISVision dataset to obtain road image data and road geographic data, and divide them into training set, validation set and test set; Feature extraction module: Used to input road image data into the image feature extractor to obtain image features; input road geographic data into the GIS feature encoder to obtain GIS encoded features; Feature alignment module: This module processes image features and GIS encoded features through a geographic alignment image encoder to obtain joint features; the image features and joint features are then added together and input into the lesion morphology modeler to obtain lesion modeling features. Risk level and maintenance visualization image prediction module: used to input the road disease modeling features into the road risk level scorer to obtain the predicted risk level; and to add the disease modeling features and GIS coding features and input them into the maintenance visualization generator to obtain the predicted maintenance visualization image. Model optimization module: Constructs a total loss function to measure the risk level, the difference between maintenance visualization images and manual annotations, trains the model, and uses the Adam optimizer to iterate the parameters through backpropagation to finally obtain the trained model; Detection module: This module is used to preprocess the road images and geographic data to be predicted, and then input them into the trained model to obtain the predicted risk level and maintenance visualization image. Maintenance prediction output module: This module is used to input the predicted risk level and maintenance visualization image of each road segment into the GIS road maintenance scientific decision-making system to obtain the maintenance prediction status of each road segment on the map.

[0039] 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 road maintenance intelligent decision-making method based on deep learning and GIS, characterized in that, Includes the following steps: S1. Collect road surface images and corresponding road geographic data, perform manual annotation, and construct the GIS image dataset GISVision; S2. Preprocess the GISVision dataset to obtain road image data. and road geographic data It is divided into training set, validation set and test set; S3. Road image data The image features are input into the image feature extractor to obtain the image features. ; road geographic data The input is fed into the GIS feature encoder to obtain the GIS coded features. ; S4. Image Features and GIS coding features After processing by the geographic alignment image encoder, joint features are obtained. ; Image features and joint features The sums are then input into the disease morphology modeler to obtain the disease modeling features. ; S5. Modeling features of diseases The data is input into the road risk level scorer to obtain the predicted risk level. ; Modeling features of diseases and GIS coding features The sums are then input into the maintenance visualization generator to obtain the predicted maintenance visualization image. ; S6. Constructing a total loss function to measure risk level Maintenance visualization images The differences between manual annotation and training were analyzed in steps S3-S5. The Adam optimizer was used to iterate the parameters through backpropagation, and finally the trained model was obtained. S7. After the road images and geographic data to be predicted are preprocessed in step S2, they are input into the trained model to obtain the predicted risk level and maintenance visualization image. S8. Input the predicted risk level and maintenance visualization image of each road segment into the GIS road maintenance scientific decision-making system to obtain the maintenance prediction of each road segment on the map.

2. The intelligent decision-making method for road maintenance based on deep learning and GIS according to claim 1, characterized in that, Step S2 specifically includes: S21. Unify and normalize the image sizes in the GISVision dataset, scale the RGB three-channel pixel values ​​to the [0,1] range, and use the CLAHE method to process dark areas to obtain road image data. ; S22. Normalize the road geographic data. Numerical fields are scaled to the [0,1] range using min-max scaling, while categorical fields maintain one-hot encoding. Pair each image data with its corresponding road geographic data using a unique index to form structured road geographic data. .

3. The intelligent decision-making method for road maintenance based on deep learning and GIS according to claim 2, characterized in that, The image feature extractor includes convolutional layers, pooling layers, and fully connected layers.

4. The intelligent decision-making method for road maintenance based on deep learning and GIS according to claim 3, characterized in that, The GIS feature encoder includes a fully connected layer and an output layer, and the output layer is subjected to L2 normalization.

5. The intelligent decision-making method for road maintenance based on deep learning and GIS according to claim 4, characterized in that, The geographic alignment image encoder in step S4 includes: The geographic alignment image encoder includes an image feature linear mapping layer, a GIS feature linear mapping layer, a location encoder, and a fusion projection layer; the image feature linear mapping layer, the GIS feature linear mapping layer, and the fusion projection layer are all fully connected layers, and the ReLU activation function is used for all of them; Image features Image linear mapping features are obtained after the image feature linear mapping layer. GIS coding features GIS linear mapping features are obtained after the GIS feature linear mapping layer. ; The image linear mapping feature and GIS linear mapping features The input is fed into the position encoder to obtain the fixed position code. The position encoder uses a sinusoidal position encoding formula to encode fixed positions. Linear mapping features of the image respectively and GIS linear mapping features By adding element by element, the image location code is obtained. and GIS location coding ; Image position encoding and GIS location coding The features are summed along the channel dimension and then fed into the fusion projection layer to obtain the joint features. .

6. The intelligent decision-making method for road maintenance based on deep learning and GIS according to claim 5, characterized in that, The disease morphology modeler in step S4 includes: The disease morphology modeler includes a normalization layer, a residual morphology modeling module, and an output layer; the residual morphology modeling module includes a linear layer, a Dropout layer, and a normalization layer; the output layer is a linear projection layer. The image features and joint features After element-wise addition, the data is processed by L2 normalization through a normalization layer to obtain normalized features. Normalization characteristics After residual morphology module construction and output layer processing, the disease modeling features are obtained. , , Representing disease modeling features The i-th dimension feature.

7. The intelligent decision-making method for road maintenance based on deep learning and GIS according to claim 6, characterized in that, In step S5, the disease modeling features are... The data is input into the road risk level scorer to obtain the predicted risk level. : Disease modeling features The data is input into a road risk level scorer for prediction, yielding the predicted risk level probability. Select the probability of all risk levels The category with the highest probability is used as the predicted risk level. The formula for the scoring function is expressed as follows: in, Representing disease modeling features The i-th dimension feature; exp represents an exponential function with base e; | | indicates to Take the absolute value; β represents the balance constant; γ represents the gain coefficient; λ represents the regularization term constant; This indicates the probability of predicting a risk level of c; C represents the number of categories.

8. The intelligent decision-making method for road maintenance based on deep learning and GIS according to claim 7, characterized in that, In step S5, the maintenance visualization generator specifically refers to: The maintenance visualization generator includes a feature fusion normalization layer, a feature expansion mapping module, and a visualization image reconstruction module; the feature expansion mapping module includes a fully connected layer and a reconstruction mapping layer; the reconstruction mapping layer reconstructs the features; the visualization image reconstruction module is a deconvolution layer; Disease modeling features and GIS coding features After element-wise addition, the input is fed into a feature fusion normalization layer for normalization processing, resulting in fused and normalized features. ; Integration and normalization features The input is fed into the feature expansion mapping module and the visualization image reconstruction module to obtain the predicted maintenance visualization image. .

9. The intelligent decision-making method for road maintenance based on deep learning and GIS according to claim 8, characterized in that, Step S6 specifically includes: The total loss function is the sum of the image loss function and the risk level loss function; the formula for calculating the image loss function is: ， in, Visualization of maintenance The value of the i-th pixel; This represents the label of the i-th pixel in a manually labeled image; The constant represents the risk level loss function, which is calculated as follows: ， Where N represents the number of samples; Indicates risk level The value, This indicates the true risk level.

10. A pavement maintenance intelligent decision-making system based on deep learning and GIS, executing the pavement maintenance intelligent decision-making method based on deep learning and GIS as described in claim 1, characterized in that, include: Data acquisition module: used to collect road surface images and corresponding road geographic data, perform manual annotation, and construct the GIS image dataset GISVision; Data preprocessing module: used to preprocess the GISVision dataset to obtain road image data and road geographic data, and divide them into training set, validation set and test set; Feature extraction module: Used to input road image data into the image feature extractor to obtain image features; input road geographic data into the GIS feature encoder to obtain GIS encoded features; Feature alignment module: This module processes image features and GIS encoded features through a geographic alignment image encoder to obtain joint features; the image features and joint features are then added together and input into the lesion morphology modeler to obtain lesion modeling features. Risk level and maintenance visualization image prediction module: used to input the road disease modeling features into the road risk level scorer to obtain the predicted risk level; and to add the disease modeling features and GIS coding features and input them into the maintenance visualization generator to obtain the predicted maintenance visualization image. Model optimization module: Constructs a total loss function to measure the risk level, the difference between maintenance visualization images and manual annotations, trains the model, and uses the Adam optimizer to iterate the parameters through backpropagation to finally obtain the trained model; Detection module: This module is used to preprocess the road images and geographic data to be predicted, and then input them into the trained model to obtain the predicted risk level and maintenance visualization image. Maintenance prediction output module: This module is used to input the predicted risk level and maintenance visualization image of each road segment into the GIS road maintenance scientific decision-making system to obtain the maintenance prediction status of each road segment on the map.

Images