Railway and accompanying road intelligent identification method based on remote sensing image

Abstract

The application discloses a railway and accompanying highway intelligent identification method based on remote sensing images, relates to remote sensing image processing technology, and aims at solving the problems of lack of labeled samples and difficulty in automatic identification in the prior art. The application selects high-resolution remote sensing images and samples the images to image patches with relatively high computing power. Based on the idea of contrast learning, the model is pre-trained by using a plurality of unlabeled samples, and then the model is optimized by using a small amount of supervised samples. The road network search method is used to replace grid traversal for calculation. According to the prediction result, a morphological filtering algorithm is used to filter noise signals caused by singular points, and a center point splicing algorithm is used to combine the patches into a whole image. The semi-supervised image classification method used in the application can pre-train a large number of remote sensing image sample features, and a model with good classification effect can be obtained by using only a small amount of labeled samples.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, specifically to an intelligent identification method for railways and their accompanying highways based on remote sensing images. Background Technology

[0002] Since the founding of the People's Republic of China, the rapid development of China's railways has been closely linked to the country's growing strength. In 1952, the Chengdu-Chongqing Railway, the first railway in the new China, began operation, marking the beginning of China's railway construction. After the Third Plenary Session of the Eleventh Central Committee of the Communist Party of China, the construction strategy of China's main railway lines ushered in a new era of economic transformation. Today, my country's railway operating mileage has exceeded 140,000 kilometers, with high-speed railway mileage approaching 40,000 kilometers, ranking first in the world. According to the jointly issued "Medium and Long-Term Railway Network Plan," my country has already built a "four vertical and four horizontal" railway network, and the "eight vertical and eight horizontal" high-speed railway network is being steadily promoted. However, while high-speed railways are developing rapidly, they also place higher demands on railway construction and maintenance. High-speed railways not only require regular inspections and maintenance, but also emergency repairs in the face of disasters such as earthquakes, landslides, mudslides, and mudslides. The construction and maintenance of railways mainly rely on the accompanying highways to provide personnel and materials. Before the widespread use of remote sensing imagery, railway maintenance and emergency repairs were hampered by the fact that problems could only be detected when trains or patrol personnel arrived on site, delaying repairs and even leading to dangerous situations where trains couldn't react in time. Furthermore, the lack of image-based location tracking made it difficult for maintenance personnel to directly report the specific locations requiring repair and maintenance. With the widespread availability of high-resolution remote sensing imagery, after a disaster, staff can promptly use it to observe the condition of high-speed railways and adjacent highways in the affected area, pinpointing sections requiring maintenance and repair, effectively saving time and reducing risks. Summary of the Invention

[0003] In order to solve the problems existing in the prior art, the present invention proposes to provide a method for intelligent identification of railways and their accompanying highways based on remote sensing images, which aims to solve the problems of scarce labeled samples and difficulty in automated identification in the prior art.

[0004] A method for intelligent identification of railways and their accompanying highways based on remote sensing imagery includes the following steps:

[0005] Step S1: Select high-resolution remote sensing images, manually construct accurate semantic annotations for key targets, and sample them into image blocks with relatively high computing power. Key targets include stations.

[0006] Step S2: Pre-train the model using unlabeled samples, then optimize the model using supervised samples; and finally, iterate to obtain a model with excellent classification performance by replacing the classifier and loss function.

[0007] Step S3: Use the road network search method to perform prediction calculations on the remote sensing image from step S1;

[0008] Step S4: Based on the prediction results of step S3, a morphological filtering algorithm is used to filter out noise signals caused by singularities, and a center point stitching algorithm is used to synthesize the segments into a complete image.

[0009] Preferably, step S1 includes the following steps:

[0010] Step S11: Based on the high-resolution satellite remote sensing image, select a buffer zone of 200-400m to clip the railway line vector;

[0011] Step S12: Use QGIS tools to vectorize the buffer zone images of stations and residential areas;

[0012] Step S13: Construct remote sensing image slices of the road network containing railways and their accompanying highways, and store the positive sample slices containing railways and their accompanying highways and the negative sample slices containing other information separately; wherein the positive sample slices include labeled remote sensing image blocks and unlabeled remote sensing image blocks.

[0013] Preferably, step S2 includes the following steps:

[0014] Step S21: Based on contrastive learning, using InfoNCE as the loss function, and taking individual discrimination as the proxy task, a self-supervised feature learning model is constructed. The positive and negative sample blocks obtained in Step 1 are used to pre-train the model parameters; the formula for the InfoNCE loss function is as follows:

[0015]

[0016] In the formula L InfoNCE This is the loss function that measures contrastive learning, where q is the original sample value, and k is the loss function. + q is an approximation obtained after transformations such as rotation, cropping, and color difference adjustment; k- represents other samples; and τ is the temperature used to limit L. InfoNCE The smoothness.

[0017] Step S22: Using the model parameters obtained from the self-supervised training in S21, replace the model's classifier and loss function, and fine-tune the semantic segmentation network using labeled samples. The loss function used during fine-tuning includes:

[0018]

[0019] L ||ε％,ε|| =|ε%-ε| (3)

[0020] L(ε%,ε)=Lms-ssim (ε%,ε)+L ||ε％,ε|| (4)

[0021] In the formula, M represents the scaling factor, [ε%, ε] represent the real image and the predicted image, respectively, and [μ] represents the scaling factor. ε％ ,μ ε ] represent the mean values ​​of the actual image and the predicted image, respectively, [σ] ε％ ,σ ε ] represent the variances of the actual image and the predicted image, respectively, σ ε％ε [β] represents the covariance between the real image and the predicted image. m ,γ m [c1, c2] are two constants representing the importance of the mean and variance. [c1, c2] are two very small constant factors to prevent the denominator from being zero. ||ε％,ε|| It is the L1 norm difference between the real image and the predicted image. ms-ssim (ε%,ε) represents the structural similarity gap between the real image and the predicted image. L(ε%,ε) represents the final loss function used for fine-tuning during supervised training.

[0022] Preferably, step S3 includes the following steps:

[0023] Step S31: Initialize the prediction using the model trained in S2 based on the buffer constructed in step S11;

[0024] Step S32: When a railway appears in the prediction results, establish a root node; then perform a regional grid search according to the direction of the railway. When a nearby accompanying road is found, establish a child node, and further traverse the entire accompanying road through the regional grid search. When the search of the entire road is completed, return to the child node and continue to make predictions along the direction of the railway to complete the identification and extraction of the railway and accompanying road in the entire region.

[0025] Preferably, step S4 includes the following steps:

[0026] Step S41: Based on the morphological closing operation method, morphological filtering is performed using the convolution operator to reduce noise contamination of the results;

[0027] Step S42: Considering that the closer the image slice is to the center, the higher the accuracy, the center point connection algorithm is used to stitch adjacent slices together;

[0028] Step S43: Use breakpoint detection and breakpoint matching to reconnect disconnected wires.

[0029] The beneficial effects of this invention include:

[0030] 1. The semi-supervised image classification method adopted in this scheme can be pre-trained using features from a large number of remote sensing image samples, requiring only a small number of labeled samples to obtain a model with good classification performance. This model can effectively perform calculations on high-resolution remote sensing images and can identify and extract vector information of railways and their accompanying highways.

[0031] 2. A region search-based railway and its accompanying highway identification scheme was constructed. This scheme transforms the model from a region traversal method to a region search method by establishing a root node-child node structure, which effectively reduces computing power consumption and significantly accelerates the prediction speed.

[0032] 3. Morphological filtering effectively removes noise points from images. A center-point connection algorithm is used to stitch adjacent slices together, significantly improving the average pixel accuracy of the prediction results. Furthermore, breakpoint detection and matching are employed to connect broken lines, preventing gaps between lines. Attached Figure Description

[0033] Figure 1 This is a flowchart of an intelligent identification method for railways and their accompanying highways based on remote sensing imagery, as described in Example 1.

[0034] Figure 2 Example 1 illustrates the vectorization of line features.

[0035] Figure 3 This demonstrates the semantic segmentation and self-supervised training model structure.

[0036] Figure 4 Example 1 involves the prediction results obtained after pre-training the individual discrimination and comparison learning parameters of the model and fine-tuning the supervised model.

[0037] Figure 5 The result of splicing adjacent blocks using the center point connection algorithm in Example 1 is shown.

[0038] Figure 6 This is a comparison diagram before and after road morphology filtering and breakpoint reconnection involved in Example 1. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0040] Example 1

[0041] The following is in conjunction with the appendix Figure 1 -Appendix Figure 6 Specific embodiments of the present invention will be described in detail;

[0042] like Figure 1 As shown, a method for intelligent identification of railways and their accompanying highways based on remote sensing imagery includes the following steps:

[0043] Step S1: Select high-resolution remote sensing images, manually construct precise semantic annotations for key targets such as stations, and sample them into image blocks with appropriate computing power. This involves choosing one of three image sizes—256×256, 512×512, or 1024×1024—based on the computer's graphics card memory. 512×512 is typically used. (Refer to...) Figure 2 In this dataset, red represents railway tracks, purple represents highways, and green represents side roads. While establishing the vector representation, the width of the roads is measured to create a buffer zone, generating an accurate labeled sample dataset of railways and their accompanying highways. This process includes the following steps:

[0044] Step S11: Based on the high-resolution satellite remote sensing image, select a suitable railway line vector for buffer clipping. To avoid errors caused by the railway line vector, the initial clipping range is set at 200m.

[0045] Step S12: Use QGIS tools to vectorize key sections of the buffer image, mainly using complex sections such as stations and residential areas, thereby improving the robustness of the model to negative samples;

[0046] Step S13: Construct a road network slice containing railways and their accompanying highways, and store the positive sample slice containing railways and their accompanying highways and the negative sample slice containing other information separately.

[0047] Step S2: Based on the idea of ​​contrastive learning, the model is pre-trained using a majority of unlabeled samples, and then optimized using a small number of supervised samples. By replacing the classifier and loss function, a model with excellent classification performance can be obtained with only a few iterations. (See reference...) Figure 3 The specific steps are as follows:

[0048] Step S21: Based on contrastive learning, using InfoNCE as the loss function, a self-supervised feature learning model is constructed with individual discrimination as the proxy task, and the model parameters are pre-trained. The InfoNCE loss function formula is as follows:

[0049]

[0050] In the formula L InfoNCE This is the loss function that measures contrastive learning, where q is the original sample value, and k is the loss function. + It is an approximation of q after transformations such as rotation, cropping, and color difference, k - For other samples, τ is the temperature used to limit L. InfoNCE The smoothness.

[0051] Step S22: Using the model parameters obtained from the unsupervised training in S21, replace the model's classifier and loss function, and perform fine-tuning training using labeled samples. The loss function used during fine-tuning consists of the following functions:

[0052]

[0053] L ||ε％,ε|| =|ε%-ε| (3)

[0054] L(ε%,ε)=L ms-ssim (ε%,ε)+L ||ε％,ε|| (4)

[0055] In the formula, M represents the scaling factor, [ε%, ε] represent the real image and the predicted image, respectively, and [μ] represents the scaling factor. ε％ ,μ ε ] represent the mean values ​​of the actual image and the predicted image, respectively, [σ] ε％ ,σ ε ] represent the variances of the actual image and the predicted image, respectively, σ ε％ε [β] represents the covariance between the real image and the predicted image. m ,γ m [c1, c2] are two constants representing the importance of the mean and variance. [c1, c2] are two very small constant factors to prevent the denominator from being zero. ||ε％,ε|| It is the L1 norm difference between the real image and the predicted image. ms-ssim (ε%,ε) represents the structural similarity gap between the real image and the predicted image. L(ε%,ε) represents the final loss function used for fine-tuning during supervised training.

[0056] Step S3: Considering that only a small portion of the high-resolution remote sensing image contains railways and their accompanying highways, and the rest can be regarded as background, a road network search method is used instead of grid traversal for calculation; for example... Figure 4 As shown, Original Img is the original high-resolution remote sensing image, Train Label is a manually created label, and Pre is the result predicted by the model; the specific steps of step S3 are as follows:

[0057] Step S31: Initialize the prediction using the model trained in S22 based on the range initially determined in S11;

[0058] Step S32: When a railway appears in the prediction results, a root node is established. Then, a regional grid search is performed according to the direction of the railway. When a nearby accompanying road is found, a child node is established, and the entire accompanying road is traversed through the regional grid search. When the search of the entire road is completed, the process returns to the child node and continues to make predictions along the direction of the railway to complete the identification and extraction of the railway and accompanying road in the entire region.

[0059] Step S4: Based on the prediction results, a morphological filtering algorithm is used to filter out noise signals caused by singularities, and a center point stitching algorithm is used to combine the sliced ​​blocks into a complete image; the specific steps of step S4 are as follows:

[0060] Step S41: Based on the morphological closing operation method, a 10×10 operator is used for morphological filtering to reduce noise contamination of the results;

[0061] Step S42: Considering that the closer the image slice is to the center, the higher the accuracy, the center point connection algorithm is used to stitch adjacent slices together; refer to Figure 5 The dotted red box in the left image represents the most accurate recognition range for a single slice. Since the predicted image is built using a 50% overlap slice method, using the central region as the prediction result for image stitching can obtain the most accurate image stitching result.

[0062] Step S43: Reconnect disconnected wires using breakpoint detection and breakpoint matching; refer to Figure 6 The image on the left is the original image, the image in the middle is the result after connecting the center points of the slices, and the image on the right is the image after morphological filtering and reconnection of breakpoints. The noise and broken lines in the image have been effectively repaired.

[0063] The embodiments described above merely illustrate specific implementation methods of this application, and while the descriptions are detailed and specific, they should not be construed as limiting the scope of protection of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the technical solution of this application, and these modifications and improvements all fall within the scope of protection of this application.

Claims

1. A method for intelligent identification of railways and their accompanying highways based on remote sensing imagery, characterized in that, Includes the following steps: Step S1: Select high-resolution remote sensing images, manually construct accurate semantic annotations for key targets, and sample them into image blocks with relatively high computing power. Key targets include stations. Step S2: Pre-train the model using unlabeled samples, then optimize the model using supervised samples; finally, iterate by replacing the classifier and loss function to obtain a model with excellent classification performance. Step S3: Use the road network search method to perform prediction calculations on the remote sensing image from step S1; Step S4: Based on the prediction results of step S3, a morphological filtering algorithm is used to filter out noise signals caused by singularities, and a center point stitching algorithm is used to synthesize the segments into a complete image. Step S1 includes the following steps: Step S11: Based on the high-resolution satellite remote sensing image, select a buffer zone of 200-400m to clip the railway line vector; Step S12: Use QGIS tools to vectorize the buffer zone images of stations and residential areas; Step S13: Construct remote sensing image blocks containing railways and their accompanying highways, and store positive sample blocks containing railways and their accompanying highways and negative sample blocks containing other information separately; wherein positive sample blocks include labeled remote sensing image blocks and unlabeled remote sensing image blocks. Step S3 includes the following steps: Step S31: Initialize the prediction using the model trained in S2 based on the buffer constructed in step S11; Step S32: When a railway appears in the prediction results, establish a root node; then perform a regional grid search according to the direction of the railway. When a nearby accompanying road is found, establish a child node, and further traverse the entire accompanying road through the regional grid search. When the search of the entire road is completed, return to the child node and continue to make predictions along the direction of the railway to complete the identification and extraction of the railway and accompanying road in the entire region.

2. The intelligent identification method for railways and their accompanying highways based on remote sensing imagery according to claim 1, characterized in that, Step S2 includes the following steps: Step S21: Based on contrastive learning, using InfoNCE as the loss function, and taking individual discrimination as the proxy task, a self-supervised feature learning model is constructed. The positive and negative sample blocks obtained in Step 1 are used to pre-train the model parameters; the formula for the InfoNCE loss function is as follows: （1）； In the formula It is the loss function that measures contrastive learning, where These are the original values ​​of the sample. yes The approximate value obtained after rotation, cropping, and color difference transformation. These are other samples. Temperature is used to limit The smoothness of the model; Step S22: Using the model parameters obtained from the self-supervised training in S21, replace the model's classifier and loss function, and fine-tune the semantic segmentation network using labeled samples. The loss function used during fine-tuning includes: （2）； （3）； （4）； In the formula This represents the scaling factor. They represent real images and predicted images, respectively. These represent the mean values ​​of the actual image and the predicted image, respectively. These represent the variances of the actual image and the predicted image, respectively. This represents the covariance between the actual image and the predicted image. These two constants represent the importance of the mean and variance. These are two very small constant factors to prevent the denominator from being zero; It is the difference in the first norm between the real image and the predicted image. It is the difference in structural similarity between the real image and the predicted image; This represents the final loss function used for fine-tuning during supervised training.

3. The intelligent identification method for railways and their accompanying highways based on remote sensing imagery according to claim 1, characterized in that, Step S4 includes the following steps: Step S41: Based on the morphological closing operation method, morphological filtering is performed using the convolution operator to reduce noise contamination of the results; Step S42: Considering that the closer the image slice is to the center, the higher the accuracy, the center point connection algorithm is used to stitch adjacent slices together; Step S43: Use breakpoint detection and breakpoint matching to reconnect disconnected wires.

Images