A railway line along the ground feature classification method based on remote sensing image

Abstract

The application discloses a railway along-line ground object classification method based on remote sensing images, which comprises the following steps: firstly, acquiring remote sensing images and completing sample labeling of six types of ground objects, i.e., railways, highways, water areas, vegetation, bare land and buildings; secondly, extracting blue light, green light, red light, near-infrared and short-wave infrared key bands, constructing normalized vegetation index, bare soil index, iron mineral index, highway coupling index, railway coupling index and linear enhancement components and the like, and simultaneously generating three types of physical threshold prior features of water bodies, railways and highways; thirdly, extracting local spatial context features through multi-scale neighborhood mean smoothing, combining normalized global position features and splicing into a high-dimensional feature set; fourthly, training a machine learning classification model based on sample features and completing evaluation; and finally, inputting the remote sensing images into the model to realize full-image classification and output results. Through material feature and linear topological constraint coupling, physical prior threshold constraint and spatial context information fusion, the application solves the problems of railway and highway spectral confusion, line-shaped ground object fragmentation and brokenness and insufficient robustness of complex scene classification, and improves the precision of railway along-line ground object classification.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, and specifically to a method for classifying land features along railway lines based on remote sensing images. Background Technology

[0002] With the rapid development of space remote sensing technology, using satellite imagery for surface monitoring has become an important technical means. However, in the interpretation of remote sensing images, there is a problem of different objects with the same spectrum being classified as ground features.

[0003] This problem is even more pronounced in complex terrain scenarios along railway lines: railway lines not only include man-made objects such as rails, ballast, and roadbeds, but also natural features such as vegetation, water bodies, and bare soil. Furthermore, linear features of different materials, such as railways and highways, have highly similar spectral responses, making them prone to confusion. At the same time, shadowing, occlusion, and mixed pixel effects in complex environments such as mountainous areas and urban-rural fringe areas further increase the uncertainty of spectral features, placing higher demands on the feature discrimination ability and robustness of classification models.

[0004] Existing pixel-level traditional machine learning classification methods, such as Support Vector Machine (SVM) and Random Forest, have the following problems when dealing with complex terrain scenes:

[0005] 1. The spectral characteristics of railway and highway targets are highly misidentifiable. In medium- to high-spatial-resolution remote sensing imagery, railways and highways, as typical linear features, exhibit extremely similar spectral reflectance curves in the conventional visible and near-infrared bands, making them prone to confusion. Existing methods do not construct specific feature operators tailored to the material characteristics of railways, relying solely on general basic indices or band combinations. This fails to effectively capture the unique combined spectral characteristics of railway ballast (bare soil) and metal rails, and makes it difficult to filter out interference objects with only single attributes. Consequently, the feature recognition of railway targets is insufficient, which is one of the core contributing factors to misidentification of railways and highways.

[0006] 2. Insufficient distinguishability of spectral features of land features. Existing methods rely solely on basic vegetation indices or original band combinations, making it difficult to discover deep physical differences between land features. This results in a lack of more reliable information in the model's classification decisions, leading to misclassification of different categories of linear land features and thus affecting the accuracy of land feature classification.

[0007] 3. Fragmentation of Spatial Topology. Traditional machine learning algorithms such as Random Forest essentially rely on independent judgment of each pixel, ignoring the spatial context between pixels. This leads to two problems in classification results: on the one hand, linear features exhibit fragmented, discontinuous characteristics; on the other hand, isolated misclassified points appear within large areas of homogeneous features, thus violating the objective geometric properties of the features.

[0008] To address the shortcomings of existing technologies, this invention provides a method for classifying features along railway lines based on remote sensing imagery. The aim is to solve the technical challenges of confusing railways and highways and fragmenting spatial topology in low-to-medium resolution imagery by constructing enhanced features encompassing macroscopic indices, physical materials, metal oxides, and multi-scale spatial information, combined with a random forest algorithm. Summary of the Invention

[0009] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is a method for classifying land features along railway lines based on remote sensing imagery, the method comprising the following steps:

[0010] S1: Acquire remote sensing images and perform manual labeling of land cover categories, including railways, highways, water bodies, vegetation, bare land, and buildings, to obtain a sample dataset;

[0011] S2: Extract key bands from remote sensing images and perform band operations to construct enhancement features and physical threshold prior features. The key bands include: Blue, Green, Red, Near-Infrared (NIR), and Short-Wave Infrared (SWIR). The enhancement features specifically include: Normalized Difference Vegetation Index (NDVI), Normalized Difference Building Index (NDBI), Modified Normalized Difference Water Index (MNDWI), Bare Soil Index (BSI), Ferric Mineral Index, Road Coupling Index (REI), Linear Enhancement Component (Line_norm), and Railway Coupling Index (RFI). The physical threshold prior features include three categories: water threshold prior features (water_priori), railway threshold prior features (rail_priori), and road threshold prior features (road_priori).

[0012] S3: Perform neighborhood mean smoothing on the enhanced features to extract the local spatial context features of the pixels; extract the two-dimensional spatial coordinates of each pixel in the image as the global position features of the pixel; stitch together the basic band features, enhanced features, physical threshold prior features, local spatial context features and global position features of the remote sensing image to obtain a high-dimensional feature set for each pixel in the remote sensing image.

[0013] S4: Transform the sample data into geographic coordinates to the corresponding cell row and column index, extract the corresponding sample feature vector from the high-dimensional feature set, match the sample feature vector with its corresponding land cover category label, and divide it into a training set and a validation set; train a machine learning classification model using the training set, evaluate the trained machine learning classification model using the validation set, and save the trained model.

[0014] S5: Input remote sensing images into a trained machine learning classification model for prediction and output the land cover classification results.

[0015] The railway coupling index (RFI) in S2 is constructed by coupling the bare soil index and the iron mineral index, and its calculation formula is as follows:

[0016]

[0017] in, The bare soil index, The iron oxide mineral index, For linear enhancement components.

[0018] The Bare Soil Index (BSI) captures the differences in reflectance of features along the railway line in the shortwave infrared band, and its calculation formula is as follows:

[0019]

[0020] in, It is in the shortwave infrared band. It is in the red light band. It is in the near-infrared band. It is in the blue light band. It is a very small constant;

[0021] The Ferric index captures the physical characteristics of metal oxidation in and around the railway tracks, and its calculation formula is as follows:

[0022]

[0023] in, It is in the red light band. It is in the blue light band. It is a very small constant.

[0024] The linear enhancement component (Line_norm) is coupled with the material features as a spatial topological constraint, and its calculation formula is as follows:

[0025]

[0026] in, Indicates a specific spatial scale Below, the linear response function values ​​constructed using the eigenvalues ​​of the Hessian matrix are... This indicates that the maximum value is taken from the multi-scale response result, which is used to cover linear targets of different widths. , These represent the maximum and minimum values ​​of the linear response intensity of the entire remote sensing image, respectively. To prevent extremely small constants with a denominator of zero.

[0027] Furthermore, the enhanced features in step S2 are features constructed for different physical materials, and the physical threshold prior features are used to perform binary feature representation of the enhanced features by setting a preset physical threshold.

[0028] Furthermore, step S2 includes constructing prior features for water body thresholds, the calculation formula of which is as follows:

[0029]

[0030] in, This is an improved normalized water index. The prior feature is assigned a value of 1 when a pixel meets the above threshold conditions, and 0 otherwise.

[0031] Furthermore, step S2 includes constructing railway threshold prior features, the calculation formula of which is as follows:

[0032]

[0033] in, For railway coupling index, This is the Normalized Difference Vegetation Index. When a pixel simultaneously satisfies the extreme value of railway material characteristics and the condition of low vegetation cover, this prior feature is assigned a value of 1; otherwise, it is assigned a value of 0.

[0034] Furthermore, step S2 includes constructing prior features for highway thresholds, the calculation formula of which is as follows:

[0035]

[0036] in, For the normalized building index, Normalized Difference Vegetation Index (NDVI) This is the highway enhancement index. A prior feature is assigned a value of 1 when a pixel simultaneously satisfies building features, non-vegetation features, and the spectral response of asphalt material; otherwise, it is assigned a value of 0.

[0037] Furthermore, in step S3, the enhanced feature neighborhood mean smoothing process uses sliding filter windows of different sizes to calculate the mean based on the physical spatial distribution characteristics of different features; specifically, a 3×3 pixel sliding window is used for mean smoothing of the Normalized Difference Vegetation Index (NDVI), and a 5×5 pixel sliding window is used for mean smoothing of the near-infrared band.

[0038] Furthermore, in step S3, extracting global position features involves normalizing the horizontal and vertical coordinates of each pixel using the total width and height of the image, thereby normalizing the spatial position information to the [0,1] interval.

[0039] Furthermore, the machine learning classification model in step S4 is a random forest model.

[0040] The beneficial effects of adopting the above technical solution are as follows:

[0041] 1. Constructing a Railway Coupling Index to Solve the Spectral Confusion Problem Between Railways and Highways. Addressing the technical challenge of extreme spectral confusion between railways and highways in conventional spectral characteristics, this solution constructs a Railway Coupling Index (RFI). This index nonlinearly superimposes the bare soil properties (BSI) representing ballast and gravel with the metallic properties (Ferric) of the rails, and introduces a linear enhancement component (Line_norm) as a spatial topological constraint to form features specific to railway targets. This filters out interference objects with only a single attribute, improving the ability to distinguish between linear features such as railways and highways.

[0042] 2. Constructing multi-source indices and physical prior features to improve the robustness of classification in complex railway scenarios. This scheme constructs a road enhancement index (REI) to capture the combined scattering characteristics of asphalt pavement in the short-wave infrared and blue light bands, amplifying the spectral response of artificially paved roads, and distinguishing smooth road surfaces from rough railway ballast in the feature space. Simultaneously, it introduces three types of physical threshold prior features: water bodies, railways, and roads. Pre-set physical criteria are used to establish decision boundaries for the classification model, guiding the model to make predictions based on physical boundaries in complex background environments, thus enhancing the robustness of the classification results.

[0043] 3. Improve the spatial continuity of land features in railway line classification results. This solution addresses the shortcomings of traditional machine learning algorithms that focus only on pixel-by-pixel classification and ignore spatial relationships. By integrating local spatial features with normalized prior location information, the model can make judgments based on the surrounding environment. Neighborhood mean smoothing reduces fluctuations caused by spectral differences within a certain land feature class, thus reducing classification noise; normalized coordinate features provide the model with basic spatial location information. Combining these two aspects of information ensures that the classification results are consistent with the geometric attributes of the spatial distribution, improving the accuracy of land feature classification results. Attached Figure Description

[0044] Figure 1 This is a flowchart illustrating the overall process of a railway feature classification method based on remote sensing imagery, according to an embodiment of the present invention.

[0045] Figure 2 This is an RGB remote sensing image land cover distribution map according to an embodiment of the present invention.

[0046] Figure 3 This is a map showing the land cover classification results of an embodiment of the present invention.

[0047] Figure 4 Reference remote sensing image for embodiments of the present invention Detailed Implementation

[0048] 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 a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0049] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0050] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0051] A flowchart of a railway feature classification method based on remote sensing imagery is shown below. Figure 1 As shown, it includes the following steps:

[0052] S1: Acquire remote sensing images and perform manual labeling of land cover categories, including railways, highways, water bodies, vegetation, bare land, and buildings, to obtain a sample dataset;

[0053] First, acquire remote sensing imagery of the study area, such as Sentinel-2 data. Then, manually label representative sample points using professional geographic information system software like QGIS or ArcGIS. The sample point categories include railways, highways, vegetation, water bodies, buildings, and bare land. The labeling results are saved as a sample dataset containing geographic coordinates and feature category labels.

[0054] S2: Extract key bands from remote sensing images and perform band operations to construct enhancement features and physical threshold prior features. The key bands include: Blue, Green, Red, Near-Infrared (NIR), and Short-Wave Infrared (SWIR). The enhancement features specifically include: Normalized Difference Vegetation Index (NDVI), Normalized Difference Building Index (NDBI), Modified Normalized Difference Water Index (MNDWI), Bare Soil Index (BSI), Ferric Mineral Index, Road Coupling Index (REI), Linear Enhancement Component (Line_norm), and Railway Coupling Index (RFI). The physical threshold prior features include three categories: water threshold prior features (water_priori), railway threshold prior features (rail_priori), and road threshold prior features (road_priori).

[0055] The Normalized Difference Vegetation Index (NDVI) is primarily used to distinguish vegetation cover along railway lines. By extracting vegetation cover areas with high confidence, it can be used to eliminate vegetation obstruction of railway lines and assist models in detecting the conditions along railway lines. Its calculation formula is as follows:

[0056]

[0057] in, Indicates near-infrared reflectivity. Indicates the reflectivity in the red light band. This represents a very small constant.

[0058] The Normalized Building Index (NDBI) is used to identify buildings along railway lines. By enhancing the characteristics of artificial impermeable surfaces, it distinguishes point-like building clusters from linear highways and railways. The calculation formula is as follows:

[0059]

[0060] in, Indicates the reflectivity in the shortwave infrared band. Indicates near-infrared reflectivity. This represents a very small constant.

[0061] The Improved Normalized Water Index (MNDWI) is used to extract water bodies. By introducing the shortwave infrared (SWIR) band, it identifies and suppresses shadows cast by low solar altitude angles in situations such as high-rise buildings, preventing shadowed areas from being misidentified as water bodies. Its calculation formula is as follows:

[0062]

[0063] in, Indicates the reflectivity in the green light band. Indicates the reflectivity in the shortwave infrared band. This represents a very small constant.

[0064] The Basement Index (BSI) is used as a physical characteristic to distinguish railways from highways. Railway tracks contain gravel, and their physical properties are similar to exposed rock. The BSI index, by integrating the reflectance characteristics of short-wave infrared (SWIR) and red light, can sensitively capture the signals of these bare soils. In imagery, the BSI indices of railway gravel and highway asphalt pavement differ. The calculation formula is as follows:

[0065]

[0066] in, Indicates the reflectivity in the shortwave infrared band. Indicates the reflectivity in the red light band. Indicates near-infrared reflectivity. Indicates the reflectivity of the blue light band. This represents a very small constant.

[0067] The Ferric Oxide Index is used to assess the light reflection of railway tracks and metal structures. It leverages the strong absorption of iron oxides in the blue light band and the high reflectivity in the red light band to amplify the response intensity of metallic rails. Given the unique characteristic of metal steel rails in railways, the Ferric Oxide Index is a crucial basis for distinguishing railways from other linear features. Its calculation formula is as follows:

[0068]

[0069] in, Indicates the reflectivity in the red light band. Indicates the reflectivity of the blue light band. This represents a very small constant.

[0070] The Railway Coupling Index (RFI) is used to comprehensively characterize the physical material and spatial geometric features of railway targets. By nonlinearly coupling the Bare Soil Subbase Index (BSI), which characterizes the properties of ballast and gravel, the Ferric Mineral Index, which characterizes the metallic properties of rails, and the Linear Norm component, which characterizes the linear topology, an exclusive identification operator for railway targets is formed. This operator can effectively filter out interfering features with only a single attribute. Its calculation formula is as follows:

[0071]

[0072] in, For bare soil foundation index, The iron oxide mineral index, For linear enhancement components.

[0073] The Highway Enhancement Index (REI) is used to enhance the spectral response characteristics of artificially paved pavements. It utilizes the combined scattering characteristics of asphalt pavements in the shortwave infrared (SWIR) and blue light bands, and combines this with a ratio calculation in the green light band. This effectively decouples the smooth highway pavement from the rough railway ballast within the characteristic space. The calculation formula is as follows:

[0074]

[0075] in, For shortwave infrared reflectivity, For blue light band reflectivity, For green light band reflectivity, It is a very small constant.

[0076] The Linear Enhancement Component (Line_norm) is used to enhance the linear topological continuity in imagery, addressing the challenge of linear feature breaks. By performing multi-scale Hessian matrix eigenvalue analysis on the red band, an anisotropic response function reflecting the linear structure is constructed. The maximum response value at each scale is then linearly normalized to extract the linear consistency component with values ​​ranging from [0,1]. Its calculation formula is as follows:

[0077]

[0078] in, This represents the maximum linear response intensity across multiple scales. and These represent the maximum and minimum values ​​of the full graph response, respectively. It is a very small constant.

[0079] The water prior feature (water_priori) is used to initially locate water areas and suppress building shadow interference using preset physical criteria. When the improved normalized water index (MNDWI) of a pixel is greater than the preset threshold of 0.1, the pixel is considered to possess significant water physical properties, and the feature is assigned a value of 1; otherwise, it is assigned a value of 0. Its calculation formula is as follows:

[0080]

[0081] in, This is an improved normalized water index. The prior feature is assigned a value of 1 when a pixel meets the above threshold conditions, and 0 otherwise.

[0082] The rail prior feature (rail_priori) is used to identify potential rail areas through the dual constraints of physical material and vegetation cover. A pixel is considered to possess potential rail distribution characteristics when its Rail Coupling Index (RFI) is in a high range and its Normalized Difference Vegetation Index (NDVI) is in a low range; otherwise, it is assigned a value of 0. The calculation formula is as follows:

[0083]

[0084] in, For railway coupling index, This is the Normalized Difference Vegetation Index. When a pixel simultaneously satisfies the extreme value of railway material characteristics and the condition of low vegetation cover, this prior feature is assigned a value of 1; otherwise, it is assigned a value of 0.

[0085] The road prior feature (road_priori) is used to provide hard boundaries for road extraction through multi-dimensional physical attribute constraints. A pixel is considered a potential road area and is assigned a value of 1 when it simultaneously satisfies the characteristics of impermeable surface (high NDBI), non-vegetation (low NDVI), and asphalt material spectral response (high REI); otherwise, it is assigned a value of 0. Its calculation formula is as follows:

[0086]

[0087] in, For the normalized building index, Normalized Difference Vegetation Index (NDVI) This is the highway enhancement index. A prior feature is assigned a value of 1 when a pixel simultaneously satisfies building features, non-vegetation features, and the spectral response of asphalt material; otherwise, it is assigned a value of 0.

[0088] S3: Perform neighborhood mean smoothing on the enhanced features to extract the local spatial context features of the pixels; extract the two-dimensional spatial coordinates of each pixel in the image as the global position features of the pixel; stitch together the basic bands of the remote sensing image, the enhanced features, the physical threshold prior features, the local spatial context features, and the global position features to construct a high-dimensional feature set for each pixel in the remote sensing image.

[0089] After obtaining the basic bands and enhanced features in step S2, spatial context features are extracted using sliding window convolution techniques to construct the NDVI layer. The mean-sliding filter window is used to construct the near-infrared band. The sliding filter window performs smooth filtering with a step size of 1 on the corresponding layer, integrating the features around each pixel into the current pixel to achieve low-pass filtering and enhance the continuity of linear features in the feature space.

[0090] Extract the two-dimensional spatial coordinates of each pixel in the image as the global position feature of the pixel. Construct two two-dimensional matrices with the exact same size as the image to record the original x-coordinate of each pixel. with the vertical axis Next, based on the total width of the image... and total height As the denominator, perform linear normalization to map the pixel-level row and column numbers to... The interval. Its calculation formula is as follows:

[0091]

[0092]

[0093] in, The normalized x-axis, The ordinate is the normalized ordinate.

[0094] High-dimensional feature sets are assembled using tensor concatenation. In the channel dimension, the original spectral reflectance bands, enhanced features, neighborhood smoothing local spatial context features, and global location features are sequentially stacked. Multidimensional matrix overlay is then used to integrate them into a single high-dimensional feature set.

[0095] S4: Transform the sample data into geographic coordinates to the corresponding cell row and column index, extract the corresponding sample feature vector from the high-dimensional feature set, match the sample feature vector with its corresponding land cover category label, and divide it into a training set and a validation set; train a machine learning classification model using the training set, evaluate the trained machine learning classification model using the validation set, and save the trained model.

[0096] The geographic reference coordinates of the sample dataset in step S1 are projected and transformed to the coordinate system of the image, and the corresponding pixel row and column indices are calculated. Based on the indices, the corresponding sample feature vectors are extracted from the high-dimensional feature set generated in step S3, and the feature vectors are mapped to the land cover category labels. All samples are randomly divided into training and validation sets in a 4:1 ratio, where the training set is used to optimize model parameters and the validation set is used to verify the classification effect.

[0097] During the classifier training process, the random forest algorithm was selected, and its hyperparameters were configured. The number of decision trees was set to 100 to 500 independent decision trees, utilizing the ensemble voting results of multiple trees to improve the robustness of classification. The maximum number of features for node splitting was set to the square root of the total number of features. Each tree randomly selected some features for competition during node splitting, increasing the discriminative power between trees and reducing the risk of overfitting. The maximum depth of the decision trees was not limited, allowing the model to fully explore the nonlinear relationships in high-dimensional space. A relatively small leaf node sample threshold of 2 was set to ensure the ability to capture features of slender linear ground features. The Gini coefficient was used as the node splitting evaluation criterion, and the optimal split point was selected by calculating the purity gain after feature splitting. The performance of the random forest model was evaluated using a validation set, and the trained random forest classification model was saved.

[0098] S5: Input remote sensing images into a trained machine learning classification model for prediction and output the land cover classification results.

[0099] The remote sensing image to be predicted is input into a trained random forest classification model. The model performs feature mapping and classification on all image pixels, calculates the conditional probability of each pixel vector belonging to each land cover category, and determines the category corresponding to the maximum probability as the final classification label. The distribution of land covers in the RGB remote sensing image is shown below. Figure 2 As shown, the model prediction results are as follows: Figure 3 As shown, red represents railways, orange represents highways, blue represents water bodies, green represents vegetation, brown represents bare land, and yellow represents buildings. Figure 4 It serves as a reference for other higher-resolution remote sensing images to clearly display various land features.

[0100] The predicted pixel label matrix is ​​correlated with the affine transformation parameters and spatial reference frame (CRS) of the original remote sensing image. The latitude and longitude or projected coordinate information of the image is recovered, and the classification results are finally exported as a data format of spatial geographic attributes.

Claims

1. A method for classifying land features along railway lines based on remote sensing imagery, characterized in that, Includes the following steps: S1: Acquire remote sensing images and perform manual labeling of land cover categories, including railways, highways, water bodies, vegetation, bare land, and buildings, to obtain a sample dataset; S2: Extract key bands from remote sensing images and perform band operations to construct enhanced features and physical threshold prior features; The key bands include: Blue, Green, Red, Near-Infrared (NIR), and Short-Wave Infrared (SWIR); the enhancement features specifically include: Normalized Difference Vegetation Index (NDVI), Normalized Difference Building Index (NDBI), Modified Normalized Difference Water Index (MNDWI), Bare Soil Index (BSI), Ferric Mineral Index, Road Coupled Index (REI), Linear Enhancement Component (Line_norm), and Railway Coupled Index (RFI); the physical threshold prior features include three categories: water threshold prior features (water_priori), railway threshold prior features (rail_priori), and road threshold prior features (road_priori). S3: Perform neighborhood mean smoothing on the enhanced features to extract the local spatial context features of the pixels; extract the two-dimensional spatial coordinates of each pixel in the image as the global position features of the pixel; stitch together the basic band features, enhanced features, physical threshold prior features, local spatial context features and global position features of the remote sensing image to obtain a high-dimensional feature set for each pixel in the remote sensing image. S4: Transform the sample data into geographic coordinates to the corresponding cell row and column index, extract the corresponding sample feature vector from the high-dimensional feature set, match the sample feature vector with its corresponding land cover category label, and divide it into a training set and a validation set; train a machine learning classification model using the training set, evaluate the trained machine learning classification model using the validation set, and save the trained model. S5: Input remote sensing images into a trained machine learning classification model for prediction and output the land cover classification results. The railway coupling index (RFI) in S2 is constructed by coupling the bare soil index and the iron mineral index, and its calculation formula is as follows: in, The bare soil index, The iron oxide mineral index, For linear enhancement components. The Bare Soil Index (BSI) captures the differences in reflectance of features along the railway line in the shortwave infrared band, and its calculation formula is as follows: in, It is in the shortwave infrared band. It is in the red light band. It is in the near-infrared band. It is in the blue light band. It is a very small constant; The Ferric index captures the physical characteristics of metal oxidation in and around the railway tracks, and its calculation formula is as follows: in, It is in the red light band. It is in the blue light band. It is a very small constant. The linear enhancement component (Line_norm) is coupled with the material features as a spatial topological constraint, and its calculation formula is as follows: in, Indicates a specific spatial scale Below, the linear response function values ​​constructed using the eigenvalues ​​of the Hessian matrix are... This indicates that the maximum value is taken from the multi-scale response result, which is used to cover linear targets of different widths. , These represent the maximum and minimum values ​​of the linear response intensity of the entire remote sensing image, respectively. To prevent extremely small constants with a denominator of zero.

2. The method for classifying railway features based on remote sensing imagery according to claim 1, characterized in that, The enhanced features in step S2 are features constructed for different physical materials, and the physical threshold prior features are binarized to represent the enhanced features by setting a preset physical threshold.

3. The method for classifying railway features based on remote sensing imagery according to claim 1, characterized in that, Step S2 includes constructing prior features for water body thresholds, the calculation formula of which is as follows: in, This is an improved normalized water index. The prior feature is assigned a value of 1 when a pixel meets the above threshold conditions, and 0 otherwise.

4. The method for classifying railway features based on remote sensing imagery according to claim 1, characterized in that, Step S2 includes constructing railway threshold prior features, the calculation formula of which is as follows: in, For railway coupling index, This is the Normalized Difference Vegetation Index. When a pixel simultaneously satisfies the extreme value of railway material characteristics and the condition of low vegetation cover, this prior feature is assigned a value of 1; otherwise, it is assigned a value of 0.

5. A method for classifying railway features based on remote sensing imagery according to claim 1, characterized in that, Step S2 includes constructing prior features for highway thresholds, the calculation formula of which is as follows: in, For the normalized building index, Normalized Difference Vegetation Index (NDVI) This is the highway enhancement index. A prior feature is assigned a value of 1 when a pixel simultaneously satisfies building features, non-vegetation features, and the spectral response of asphalt material; otherwise, it is assigned a value of 0.

6. The method for classifying railway features based on remote sensing imagery according to claim 1, characterized in that, The enhanced feature neighborhood mean smoothing process in step S3 uses sliding filter windows of different sizes to calculate the mean based on the physical spatial distribution characteristics of different features; specifically, a 3×3 pixel sliding window is used for mean smoothing of the Normalized Difference Vegetation Index (NDVI), and a 5×5 pixel sliding window is used for mean smoothing of the near-infrared band.

7. A method for classifying railway features based on remote sensing imagery according to claim 1, characterized in that, The step S3 involves extracting global location features by normalizing the horizontal and vertical coordinates of each pixel using the total width and height of the image, thereby normalizing the spatial location information to the [0,1] interval.

8. A method for classifying railway features based on remote sensing imagery according to claim 1, characterized in that, The machine learning classification model in step S4 is a random forest model.

Images