A remote sensing image railway line ground object classification method fusing morphological entropy

Abstract

This application presents a method for classifying land features along railway lines using remote sensing imagery based on morphological entropy fusion. The method includes: acquiring Sentinel-1 radar imagery, Sentinel-2 optical imagery, and a fused image of the target area from a GEE platform, and preprocessing the data; generating masks on the Sentinel-1, Sentinel-2, and fused images respectively, and calculating and extracting circumferential morphological entropy, radial morphological entropy, and joint morphological entropy; extracting polarization and texture features from the Sentinel-1 imagery, and spectral and exponential features from the Sentinel-2 imagery, and constructing corresponding composite features on the fused imagery; combining these features with morphological entropy features from corresponding data sources; training a random forest classification model based on each feature set, and using the trained model to classify four typical land features along the railway line: buildings, railways, vegetation, and highways. This method, by fully utilizing multi-source remote sensing information, effectively characterizes the structural complexity and irregularity of land feature outlines through morphological entropy, significantly improving the accuracy and reliability of the classification results.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a method for classifying land features along railway lines using remote sensing images that incorporates morphological entropy. Background Technology

[0002] As a crucial transportation infrastructure, the accurate classification of features along railway lines is of great significance for railway operation safety, environmental monitoring, and planning management. Traditional methods of acquiring information on features along railway lines mainly rely on manual inspections, ground surveys, or limited video surveillance. These methods are not only labor-intensive, inefficient, and time-consuming, but also heavily constrained by weather and terrain conditions, making it difficult to meet the needs of large-scale, continuous, and near-real-time monitoring. In contrast, remote sensing technology offers advantages such as wide coverage, short acquisition cycles, and relatively low cost, providing an effective technical means for classifying features along railway lines. Synthetic Aperture Radar (SAR) data possesses all-weather, day-and-night imaging capabilities, offering unique advantages in target identification in areas with complex structures and variable climates; optical remote sensing data, on the other hand, has strong discriminative power in reflecting the spectral characteristics of features and characterizing differences in feature types. SAR data, represented by Sentinel-1 imagery, and optical remote sensing data, represented by Sentinel-2 imagery, provide a rich information foundation for the detailed classification of features along railway lines.

[0003] However, in the typical scenario of long, narrow railway lines with complex land cover types, existing remote sensing classification methods still have the following prominent problems: For Sentinel-1 imagery, existing methods mostly rely on polarimetric backscattering intensity and its texture features. These features mainly reflect scattering mechanisms and roughness, but are insufficient in depicting the overall geometric shape, boundary undulations, and structural irregularities of land cover outlines, making it difficult to fully utilize the morphological differences of linear targets such as railways and highways and their surrounding land cover. For Sentinel-2 imagery, commonly used multispectral bands and various spectral indices focus on characterizing the spectral differences of land cover, but have limited expressive power in terms of land cover outline shape and structural complexity, especially in distinguishing spectrally similar land cover. For fused imagery, existing methods mostly adopt simple stitching or weighted superposition of radar features and optical features. Although this integrates multi-source information to some extent, it still mainly remains at the spectral, scattering, and texture levels. It is evident that multi-source remote sensing data lacks a deep representation mechanism that can uniformly characterize the common geometric structure of land cover across modalities and quantify the irregularity of outline morphology, and the utilization of the macroscopic morphological features of land cover along railway lines is far from sufficient. Summary of the Invention

[0004] This invention provides a method for classifying land features along railway lines using remote sensing images that incorporates morphological entropy, mainly including:

[0005] S1: Acquire multi-source remote sensing image data of the target railway line area from the GEE platform, including Sentinel-1 image, Sentinel-2 image, and the fused image of the two, and perform preprocessing.

[0006] S2: Construct a mask for the preset bands of the Sentinel-1 image, Sentinel-2 image and the fusion image of the two and extract the contours of the ground features. Calculate the three-channel morphological entropy feature map of the corresponding band based on the contours of the ground features. The three-channel morphological entropy feature map includes three types of morphological entropy features: circumferential morphological entropy, radial morphological entropy and joint morphological entropy.

[0007] S3: Extract the polarization band and texture features of the Sentinel-1 image, the spectral band and spectral index of the Sentinel-2 image, and the polarization band, texture features, spectral band and spectral index of the fused image, and combine them with the corresponding morphological entropy features to generate several feature sets.

[0008] S4: Construct a random forest classification model using a feature set containing morphological entropy;

[0009] S5: Use the trained classification model to classify the features in the area surrounding the railway.

[0010] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0011] To address the aforementioned problems, this invention proposes a method for classifying land features along railway lines using remote sensing imagery that integrates morphological entropy. This method transforms the angle and radius variations of land feature outlines into probability distributions and measures their morphological irregularities from an entropy perspective. Specifically, firstly, on Sentinel-1, Sentinel-2, and their fused images, land feature masks and outlines are extracted through adaptive thresholding and morphological operations. The outline coordinates are transformed from Cartesian to polar coordinates to construct angle and radius sequences. Subsequently, the angle and radius increments of adjacent outline points are calculated, and the corresponding probability distributions are obtained using kernel density estimation. Based on collision entropy, circumferential morphological entropy, radial morphological entropy, and joint morphological entropy are defined to quantitatively characterize the irregularities and structural complexity of land feature outlines. Furthermore, the aforementioned morphological entropy features are integrated with the polarization and texture features of Sentinel-1, the spectral and exponential features of Sentinel-2, and the composite features of the fused imagery. Based on this, a random forest model is used to classify four typical land features along the railway line: buildings, railways, vegetation, and highways. By fusing morphological entropy features into multi-source remote sensing features, this invention effectively enhances the model's ability to describe the complexity of ground structures and its sensitivity to differences in complex ground morphology, thereby achieving superior classification performance in complex scenarios along railway lines. Specifically, this includes:

[0012] 1. This invention, based on the traditional polarization and texture features of Sentinel-1 imagery, introduces circumferential morphological entropy, radial morphological entropy, and joint morphological entropy, providing new discriminative dimensions from the perspective of the geometric shape and structural irregularity of ground features. By quantitatively characterizing the macroscopic morphology of ground features in Sentinel-1 imagery, it effectively compensates for the insufficient discriminative power of relying solely on scattering intensity and texture features, thereby improving the discriminative power of Sentinel-1 image-driven classification.

[0013] 2. In addition to the traditional features of multispectral bands and spectral indices in Sentinel-2 imagery, this invention also introduces circumferential morphological entropy, radial morphological entropy, and joint morphological entropy to quantify the overall irregularity and structural complexity of ground features. This macroscopic morphological characterization can provide additional distinguishability among complex ground features along railway lines with similar spectral characteristics, improving the classification ambiguity problem when relying solely on spectral and index features.

[0014] 3. To address the problem that existing multi-source fusion methods often rely on simple splicing or weighted superposition of radar and optical features without fully utilizing geometric structural information, this invention extracts morphological entropy features from the fused Sentinel-1 and Sentinel-2 images and combines them with polarization, texture, spectral, and exponential features to improve the overall classification accuracy and adaptability to complex backgrounds under fused data conditions.

[0015] 4. This invention, based on probability distribution and collision entropy, uniformly maps the angle and radius variations of ground feature outlines from different remote sensing data sources into morphological entropy features, thereby quantifying the irregularity of ground feature outlines and deeply mining the inherent geometric structure information of ground features along railway lines. By embedding this statistical shape metric into the multi-source feature space, not only is the classification performance of a single data source improved, but more importantly, the morphological consistency verification from multiple perspectives enhances the model's ability to identify the essential features of ground features and the reliability of the classification results. (See attached figures.)

[0016] Figure 1 This is a flowchart of a method for classifying land features along railway lines using remote sensing images based on morphological entropy, according to the present invention.

[0017] Figure 2 A false-color composite image of Sentinel-1 imagery;

[0018] Figure 3 A true-color composite image of Sentinel-2 images;

[0019] Figure 4 This is an evaluation map of the distinguishability of different land cover categories for each spectral band of Sentinel-2 image;

[0020] Figure 5 Detailed information map of the seven bands B1, B2, B4, B7, B8, B8A, and B11 of Sentinel-2 imagery;

[0021] Figure 6 Fisher score histograms of spectral indices for different land cover categories in Sentinel-2 imagery;

[0022] Figure 7 Fisher score histograms of texture features for different land cover categories in Sentinel-1 imagery;

[0023] Figure 8 Feature distribution maps of radial morphological entropy, circumferential morphological entropy, and joint morphological entropy extracted based on VV bands for different land cover categories;

[0024] Figure 9 This describes a classification map of fused images using a trained random forest model. Detailed implementation method.

[0025] 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.

[0026] like Figure 1 This embodiment of a method for classifying land features along railway lines using remote sensing images based on morphological entropy fusion may specifically include:

[0027] S1: Acquire multi-source remote sensing image data of the target railway line area from the GEE platform, including Sentinel-1 image, Sentinel-2 image, and the fused image of the two, and perform preprocessing.

[0028] S2: Construct a mask for the preset bands of the Sentinel-1 image, Sentinel-2 image and the fusion image of the two and extract the contours of the ground features. Calculate the three-channel morphological entropy feature map of the corresponding band based on the contours of the ground features. The three-channel morphological entropy feature map includes three types of morphological entropy features: circumferential morphological entropy, radial morphological entropy and joint morphological entropy.

[0029] S3: Extract the polarization band and texture features of the Sentinel-1 image, the spectral band and spectral index of the Sentinel-2 image, and the polarization band, texture features, spectral band and spectral index of the fused image, and combine them with the corresponding morphological entropy features to generate several feature sets.

[0030] S4: Construct a random forest classification model using a feature set containing morphological entropy;

[0031] S5: Use the trained classification model to classify the features in the area surrounding the railway.

[0032] Specifically, S1 includes:

[0033] (1) Determination of the railway line area. Taking the railway line in Hengshui City, Hebei Province as the research object, the target area was delineated within a buffer zone of about 700 m on both sides of the railway, totaling 76.8982 square kilometers. Sample data of different types of land features were collected in the target area, including four categories: buildings, railways, vegetation and highways, with sample numbers of 180, 110, 80 and 170, respectively.

[0034] (2) Acquisition and preprocessing of multi-source remote sensing image data. Sentinel-1 images from April 1 to April 15, 2025, including VV and VH bands, IW mode, including ascending and descending orbits, were selected and filtered. The Sentinel-1 imaging principle causes speckle noise in the data, which manifests as granular texture on the image. This noise is not the actual information of the ground features. Therefore, filtering the image can eliminate this interference and restore the true features of the ground features. Figure 2 This is a false-color composite image of Sentinel-1 imagery. Sentinel-2 imagery from April 1st to April 15th, 2025, with cloud cover less than 20%, was selected across all bands. Cloud masking was performed using the SCL band, and all images were uniformly resampled to 10-meter resolution. Clouds and cloud shadows in Sentinel-2 imagery alter surface reflectance; for example, thick clouds can increase the spectral reflectance of underlying vegetation, while cloud shadows can decrease it. This can severely interfere with vegetation index calculations and land cover classification. Even after selecting images with less than 20% cloud cover, localized thin clouds or cloud shadows may still exist. Precise masking using the SCL band ensures that every pixel in the image represents true surface information, preventing errors in analysis results due to cloud cover. Figure 3 It is a true-color composite image of Sentinel-2 images; finally, the processed Sentinel-1 and Sentinel-2 images are fused to generate a fused image.

[0035] (3) Divide the training set and the test set. The spatial partitioning method is adopted, and the samples are divided into training set and test set based on the geographical location of the samples. This avoids adjacent samples appearing in both the training set and the test set at the same time, so as to more realistically evaluate the performance of the model in unknown areas. The ratio of training set to test set is 7:3.

[0036] Specifically, S2 includes:

[0037] (1) Masks are generated independently for each preset band of Sentinel-1, Sentinel-2 and fused image respectively. The contours of ground features are extracted within the mask area. Based on the centroid coordinates of the contours, the contour coordinates are converted from Cartesian coordinates to polar coordinates to obtain angle and radius sequences.

[0038] First, the bands (B1, B2, B3, B4, B5, B6, B7, B8, B8A, B9, B11, B12) of the Sentinel-2 imagery are filtered. By loading training samples with the Sentinel-2 imagery, the effective pixel values ​​of each band are extracted. Fisher scores (between-class variance / within-class variance), separating indices (minimum between-class distance / maximum within-class standard deviation), mean range, and standard deviation are calculated to objectively measure the band's ability to distinguish between different classes. Figure 4 As shown, the band-category relationship is visually presented using mean line charts, error bar charts, and column charts. The bands with the strongest ability to distinguish the four types of land cover were selected, and finally, seven bands, B1, B2, B4, B7, B8, B8A, and B11, were chosen. Figure 5 This is a detailed display of the seven selected bands.

[0039] For Sentinel-1 images, VV and VH bands were selected. The threshold range was determined based on the 10% and 90% quantiles of the effective pixels in each band within the training area. The original backscatter values ​​were linearly normalized and combined with the Otsu thresholding method to determine the binary segmentation threshold, generating an initial binary mask. Noise and small holes were removed through opening and dilation operations, and finally two independent masks were obtained.

[0040] For Sentinel-2 images, seven bands, B1, B2, B4, B7, B8, B8A and B11, were selected. The 25th and 75th percentiles were statistically determined in the training area to determine the range of pixel values. After normalizing the effective pixels, the Otsu threshold was used for adaptive segmentation. Then, the candidate masks were closed to enhance the connectivity of the target area and suppress isolated noise, and finally seven independent masks were obtained.

[0041] For the fused image, the band experimental configurations of Sentinel-1 and Sentinel-2 were used. Binary masks were generated independently for each band of the fused image using the same adaptive threshold segmentation method as described above. All band masks were uniformly adjusted to the same spatial size, and dilation operation was performed on each mask to enhance connectivity, resulting in 9 independent masks.

[0042] Furthermore, the outline of the ground features is extracted within the mask area, its centroid coordinates are calculated, and the Cartesian coordinates of the outline pixels are converted to polar coordinates relative to the centroid to obtain the angle sequence and the corresponding radius sequence.

[0043] The angle sequence is as follows:

[0044]

[0045] This represents the direction angle from the centroid of the feature to the point on the outline, where Pointing directly to the right, the angle increases counterclockwise. The total number of contour points;

[0046] The radius sequence is as follows:

[0047]

[0048] This represents the Euclidean distance from the centroid of the feature to the point on the contour. This represents the total number of contour points. The calculation formula is:

[0049]

[0050] in Here are the coordinates of the contour points. These are the coordinates of the centroid.

[0051] (2) Calculate the adjacent increments of the angle sequence and the radius sequence, and construct the probability distribution of the angle increment, the probability distribution of the radius increment, and the two-dimensional joint increment probability distribution by kernel density estimation.

[0052] Calculate the angle and radius increments of adjacent contour points based on the angle and radius sequences:

[0053]

[0054]

[0055] in, It is the first The polar angle of each contour point relative to the centroid. It is the first The polar angle of the next point adjacent to each contour point relative to the centroid. It is the angular increment between adjacent contour points; It is the first The distance from each contour point to the centroid. It is the first The distance from the centroid to the next adjacent point of each contour point. It is the radius increment between adjacent contour points. It is the index of adjacent point pairs. It is the total number of contour points, which ultimately constitute the angle increment sequence. and radius increment sequence ;

[0056] Construct an evaluation grid based on angle and radius increments:

[0057]

[0058]

[0059]

[0060]

[0061]

[0062]

[0063] in It is the range of angle increments. It is the angular increment between adjacent contour points. It is the maximum value among all angle increments. It is the minimum value among all angle increments. It is the range of radius increments. It is the radius increment between adjacent contour points. It is the maximum value among all radius increments. It is the minimum value among all radius increments; It is the angle increment fill amount. It is the radius increment fill amount. It is a scaling factor that adds extra padding outside the sample range to ensure that the evaluation grid for kernel density estimation can cover the tail region of the sample distribution; It is an angle increment evaluation grid, used to evaluate the one-dimensional equally spaced grid point sequence for angle increment kernel density estimation. It is a radius increment evaluation grid, used to evaluate a one-dimensional, equally spaced grid point sequence for radius increment kernel density estimation. It is a function that generates equally spaced points. It is the number of grid points;

[0064] Calculate the density values ​​at grid points using Gaussian kernel density estimation:

[0065]

[0066]

[0067] in It is the kernel density estimation function of the angle increment. It is the independent variable on the angle increment axis. It is the radius increment kernel density estimation function. It is the independent variable on the radius increment axis. The length of the incremental sequence is denoted by , and the total number of contour points minus one is denoted by . For the bandwidth parameter in the direction of angle increment, This refers to the bandwidth parameter in the radius increment direction. It is the angular increment between adjacent contour points. It is the radius increment between adjacent contour points. The Gaussian kernel function has the following form:

[0068]

[0069] in Indicates the Gaussian kernel at a point The value at that location, It is a random variable that follows a standard normal distribution. It is the normalization constant of the distribution, ensuring that the integral of the entire probability density function over the entire real number field equals 1. It is the index part;

[0070] Convert density values ​​to discrete probability values:

[0071]

[0072]

[0073] in Indicates the first Incremental grid points at each angle Discrete probability values ​​at point , It is the angle increment kernel density estimation function at the grid points The value at that location, It is the angle increment grid spacing. It is the number of grid points for angle increment. An approximation is made by numerical integration over the entire angle increment axis, summing the density of all grid points by the spacing. Indicates the first Radius increment grid points Discrete probability values ​​at point , It is the radius increment kernel density estimation function at the grid points The value at that location, It is the radius increment grid spacing. It is the number of grid points in the radius increment. An approximation is made by numerical integration over the entire radius increment axis, summing the density of all grid points by the spacing;

[0074] Construct a two-dimensional evaluation grid based on the angle and radius increments:

[0075]

[0076]

[0077]

[0078] in It is an angle increment evaluation grid. It is the angular increment between adjacent contour points. It is the maximum value among all angle increments. It is the minimum value among all angle increments. It is the angle increment fill amount; It is a radius increment evaluation grid. It is the radius increment between adjacent contour points. It is the maximum value among all radius increments. It is the minimum of all radius increments. It is the radius increment fill amount; It is a function that generates equally spaced points. It is the number of points in each dimension of the evaluation grid used for two-dimensional joint kernel density estimation; through The two-dimensional mesh generated by the function It is an angle-incremental two-dimensional grid, and each row is... , It is a radius-incremental two-dimensional grid, and each column is... ;

[0079] Calculate the joint density value at grid points using Gaussian kernel density estimation:

[0080]

[0081] in It is the joint kernel density estimation function for angle increment and radius increment. It is the independent variable on the angle increment axis. It is the independent variable on the radius increment axis. The length of the increment sequence. For the bandwidth parameter in the direction of angle increment, This refers to the bandwidth parameter in the radius increment direction. It is the angular increment between adjacent contour points. It is the radius increment between adjacent contour points. The Gaussian kernel function;

[0082] Convert density values ​​to discrete joint probability values:

[0083]

[0084] in Indicates at grid points Discrete joint probability value at , It is the joint kernel density estimation function of angle increment and radius increment at grid points The value at that location, It is the angle increment grid spacing. It is the radius increment grid spacing. It is the number of grid points for angle increment. It is the number of grid points in the radius increment. It is an approximation of the entire two-dimensional grid by numerical integration, multiplying the density of each small grid by its area and then summing them all together.

[0085] (3) Calculate the circumferential morphological entropy, radial morphological entropy and joint morphological entropy based on the collision entropy, and normalize the entropy values ​​to the range of 0 to 1.

[0086] The collision entropy The formula is as follows:

[0087]

[0088] in Describes the first probability distribution in the probability distribution. A probability value, The sum of squared probabilities is used to measure the concentration of a distribution. Transform it into an entropy value;

[0089]

[0090]

[0091]

[0092] in It is the circumferential morphological entropy. It is the collision entropy of the probability distribution of angle increment. It is an angle increment probability distribution. It is the maximum possible entropy of the angle increment. It is the number of grid points for angle increment. Indicates the first Incremental grid points at each angle Discrete probability values ​​at point , It is the collision entropy of the probability distribution of angle increment; It is radial morphological entropy. It is the collision entropy of the radius increment probability distribution. It is a probability distribution of radius increment. It is the maximum possible entropy of the radius increment. It is the number of grid points in the radius increment. Indicates the first Radius increment grid points Discrete probability values ​​at point , It is the collision entropy of the radius increment probability distribution; It is the joint morphological entropy. It is the collision entropy of the joint probability distribution. It is a joint probability distribution. It is the maximum possible entropy of the joint distribution. Indicates at grid points Discrete joint probability value at , It is the collision entropy of the joint probability distribution.

[0093] (4) Assign the three types of morphological entropy corresponding to each contour to all pixel positions within the contour area to form a three-channel morphological entropy feature map with the same size as the original image.

[0094] Figure 8 Taking the VV band as an example, the morphological entropy of four types of land features—buildings, railways, vegetation, and highways—is presented. Vegetation exhibits the highest morphological entropy (average approximately 0.91-0.93), with a concentrated distribution, indicating complex and stable morphology. Buildings have the lowest morphological entropy (average approximately 0.73-0.74), with a dispersed distribution and some exhibiting a bimodal pattern, indicating relatively simple but highly variable morphology. Railways and highways fall between the two (average approximately 0.87-0.92), with railways slightly higher than highways, and both tending towards high entropy values. Overall, higher morphological entropy values ​​indicate more complex land feature morphology and more irregular outlines; lower morphological entropy values ​​indicate simpler land feature morphology and more regular outlines. This provides a basis for land feature classification that incorporates morphological entropy characteristics.

[0095] Specifically, S3 includes:

[0096] First, the correlation between derived spectral indices of four types of land features (buildings, railways, vegetation, and highways) around railways and specified bands (B1, B2, B3, B4, B7, B8, B8A, and B11) of Sentinel-2 imagery was analyzed. The core approach involved extracting the effective pixel values ​​of the bands from the Sentinel-2 imagery of the training samples and calculating 19 related spectral indices: NDVI, SAVI, EVI, GNDVI, RVI, DVI, NDRE, CI_RedEdge, NDVI8A, NDVI2, S2REP, NDBI, BSI, BRBA, IronOxide, NDWI, MNDWI, NDSI, and NDVI_RedEdge. Fisher scores (between-class variance / within-class variance), separation indices (minimum between-class distance / maximum within-class standard deviation), mean range, and standard deviation were used to evaluate the distinguishing ability of the spectral indices for the four types of land features. The analysis results were presented through multi-dimensional visualization charts (mean line graph, error bar chart, and discrimination histogram). Figure 6The Fisher score histograms for each spectral index were used to select the spectral indices with the highest differentiation among the four land cover types. Ultimately, five spectral indices were selected: NDVI (Normalized Difference Vegetation Index), SAVI (Soil-Adjusted Vegetation Index), NDVI2 (Normalized Difference Vegetation Index based on the blue band), NDRE (Normalized Red Edge Index), and GNDVI (Growth Normalized Difference Vegetation Index).

[0097] Next, multi-class texture features of the Sentinel-1 images of the training samples were calculated, including GLCM (Gray-Level Co-occurrence Matrix) texture features, local statistical features, and gradient features. The ability of texture features to distinguish four types of land cover was evaluated using Fisher scores (between-class variance / within-class variance), separating indices (minimum between-class distance / maximum within-class standard deviation), mean range, and standard deviation. The analysis results were presented in multi-dimensional visualization charts (mean line chart, error bar chart, and histogram). Figure 7 The Fisher score histograms of various texture features were used to select the texture features with the highest distinguishability among the four types of land cover. Finally, eight texture features were selected: VH local mean, VH gradient magnitude mean, VH gray-level co-occurrence matrix homogeneity, VV local mean, VV local median, VV gray-level co-occurrence matrix homogeneity, energy, and dissimilarity.

[0098] As shown in Table 1, the feature extraction of the three sets of remote sensing images is as follows: For each sample, the VH mean, VH local mean, VH gradient magnitude mean, VH gray-level co-occurrence matrix homogeneity, VV mean, VV local mean, VV local median, VV gray-level co-occurrence matrix homogeneity, energy, and dissimilarity are extracted from the Sentinel-1 image; the B1, B2, B4, B7, B8, B8A, and B11 means, as well as NDVI, SAVI, NDVI2, NDRE, and GNDVI are extracted from the Sentinel-2 image; the feature types and calculation methods of the above Sentinel-1 and Sentinel-2 images are reused on the fused image.

[0099] For each band in step S2, extract the morphological entropy feature value corresponding to each sample position from the three-channel morphological entropy feature map. This includes circumferential morphological entropy, radial morphological entropy, and joint morphological entropy. Calculate the successful extraction rate of each morphological feature in the sample and delete low-quality feature columns below a preset threshold (default 70%). For the remaining morphological features, independently calculate the mean morphological entropy within the training region according to the band dimension, and use this statistic to fill missing values ​​in the training and test sets. For non-morphological features, fill with the mean of the training set and perform range cropping to eliminate extreme outliers. As shown in Table 2, the three sets of features extracted above are combined with their corresponding morphological entropy features.

[0100] A hierarchical K-fold cross-validation method with 5 folds is used for hyperparameter search. The F1 score is used as the evaluation metric for hyperparameter selection. The average F1 score of each candidate n_estimators value is calculated through 5-fold cross-validation, and the n_estimators value with the highest F1 score is selected as the optimal hyperparameter. max_features is the integer part of the square root of the total number of features. A random forest classifier is created using the optimal n_estimators value and the calculated max_features value, and then the model is trained using a feature set containing morphological entropy. Predictions are performed on the test set, and evaluation metrics, including accuracy, F1 score, and Kappa coefficient, are calculated.

[0101] Table 1. Three sets of original features extracted from Sentinel-1, Sentinel-2, and fused image data.

[0102]

[0103] Table 2 shows the three feature sets of Sentinel-1, Sentinel-2, and the final fused image data.

[0104]

[0105] Specifically, S4 includes:

[0106] The experimental results are shown in Table 3. For fused images, adding joint morphological entropy improved the accuracy of the test set by 1.91 percentage points, the F1 score by 1.99 percentage points, and the Kappa score by 2.66 percentage points. It is the most significant morphological feature in the fused data experiment. Adding radial or circumferential morphological entropy alone can improve performance, but the effect is weaker than that of joint entropy. For Sentinel-1 images, adding radial or circumferential morphological entropy can improve performance, with an accuracy increase of about 1.28 percentage points and an F1 score increase of about 1.16 percentage points. Adding joint entropy only brings a slight improvement. For Sentinel-2 images, circumferential morphological entropy improves the accuracy by 1.27 percentage points, the F1 score by 1.31 percentage points, and the Kappa score by 1.81 percentage points. Experiments show that morphological entropy provides a new discriminative dimension from the perspective of the geometric complexity of ground feature outlines. This macroscopic morphological representation effectively complements the microscopic features of spectrum and texture, enriching the discriminative information in the feature space. In addition, the effectiveness of morphological entropy in different remote sensing data sources has been verified. More importantly, through multi-view morphological consistency representation, the model's ability to identify the essential geometric structure of ground features and the reliability of classification results have been enhanced.

[0107] As shown in Table 4, railway category identification has been fundamentally improved, with producer accuracy increasing significantly from 75.76% to 84.85%, and the misclassification error decreasing by 9.09 percentage points. This indicates that a large number of railway samples that were originally missed (misclassified as other categories) were correctly identified after the addition of joint morphological entropy. The identification results for highway category are more reliable, with user accuracy increasing from 84.00% to 89.36%, and the misclassification error decreasing by 5.36 percentage points. This shows that among the samples predicted as highways by the model, the proportion that are actually highways is higher. Joint morphological entropy enhances the model's grasp of highway contour features and reduces the situation of misclassifying other land features as highways. Building and vegetation categories maintain high accuracy. The spectral or textural features of these two categories are already very unique and easy to distinguish. Therefore, the producer accuracy and user accuracy have both reached 100%, and the addition of morphological entropy has no further impact, indicating that their features are sufficient.

[0108] Table 3. Experiment and evaluation indicators.

[0109]

[0110] Table 4 shows the evaluation results for producer accuracy / % and user accuracy / % for each category.

[0111]

[0112] Specifically, S5 includes:

[0113] Load the pre-trained random forest model and automatically parse the required multi-source features from the target image according to the model's feature requirements; divide the entire image into 512×512 blocks, extract feature vectors from all pixels in each block and stack them into a feature matrix; use the model to perform pixel-by-pixel prediction to obtain class labels, and finally output class raster and color rendered image. Figure 9 The results are from classifying fused images using the model trained in Experiments 3-4.

[0114] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for classifying land features along railway lines from remote sensing images using morphological entropy fusion, characterized in that, The method includes: S1: Acquire multi-source remote sensing image data of the target railway line area from the GEE platform, including Sentinel-1 image, Sentinel-2 image, and the fused image of the two, and perform preprocessing. S2: Construct a mask for the preset bands of the Sentinel-1 image, Sentinel-2 image and the fusion image of the two and extract the contours of the ground features. Calculate the three-channel morphological entropy feature map of the corresponding band based on the contours of the ground features. The three-channel morphological entropy feature map includes three types of morphological entropy features: circumferential morphological entropy, radial morphological entropy and joint morphological entropy. The calculation of the three-channel morphological entropy feature map of the corresponding band based on the ground feature outline specifically includes: Extract the outline of ground features within the mask area, and convert the outline coordinates from the Cartesian coordinate system to the polar coordinate system based on the centroid coordinates of the outline to obtain the angle sequence and radius sequence. Calculate the adjacent increments of the angle sequence and the radius sequence, and construct the probability distribution of the angle increment, the probability distribution of the radius increment, and the two-dimensional joint increment probability distribution by kernel density estimation; Based on the collision entropy, the circumferential morphological entropy, radial morphological entropy, and combined morphological entropy are calculated respectively, and the entropy values ​​are normalized to the range of 0 to 1. The collision entropy The formula is as follows: in Describes the first probability distribution in the probability distribution. A probability value, The sum of squared probabilities is used to measure the concentration of a distribution. Transform it into an entropy value; The three types of morphological entropy corresponding to each contour are assigned to all pixel positions within the contour area to form a three-channel morphological entropy feature map with the same size as the original image. The circumferential morphological entropy, radial morphological entropy, and combined morphological entropy are calculated based on the collision entropy, and the entropy values ​​are normalized to the range of 0 to 1, specifically as follows: in It is the circumferential morphological entropy. It is the collision entropy of the probability distribution of angle increment. It is an angle increment probability distribution. It is the maximum possible entropy of the angle increment. It is the number of grid points for angle increment. Indicates the first Incremental grid points at each angle Discrete probability values ​​at point , It is the collision entropy of the probability distribution of angle increment; It is radial morphological entropy. It is the collision entropy of the radius increment probability distribution. It is a probability distribution of radius increment. It is the maximum possible entropy of the radius increment. It is the number of grid points in the radius increment. Indicates the first Radius increment grid points Discrete probability values ​​at point , It is the collision entropy of the radius increment probability distribution; It is the joint morphological entropy. It is the collision entropy of the joint probability distribution. It is a joint probability distribution. It is the maximum possible entropy of the joint distribution. Indicates at grid points Discrete joint probability value at , It is the collision entropy of the joint probability distribution; S3: Extract the polarization band and texture features of the Sentinel-1 image, the spectral band and spectral index of the Sentinel-2 image, and the polarization band, texture features, spectral band and spectral index of the fused image, and combine them with the corresponding morphological entropy features to generate several feature sets. S4: Construct a random forest classification model using a feature set containing morphological entropy; S5: Use the trained classification model to classify the features in the area surrounding the railway.

2. The method according to claim 1, characterized in that, S1 specifically includes: Delineate the target area along the railway line, and collect sample data of four typical land features—buildings, railways, vegetation, and highways—within the target area; Based on the sample data, Sentinel-1 images containing VV and VH bands are selected and filtered to obtain processed Sentinel-1 data; Sentinel-2 images containing all bands are selected and declouding masking is performed to obtain processed Sentinel-2 data. The processed Sentinel-1 and Sentinel-2 data are fused to generate a fused image.

3. The method according to claim 1, characterized in that, The corresponding bands include: The Sentinel-1 imagery selects the VV and VH bands. The bands of the Sentinel-2 image include B1, B2, B4, B7, B8, B8A, and B11; The fused image was constructed using the Sentinel-1 and Sentinel-2 band experimental configurations.

4. The method according to claim 1, characterized in that, The angle increment probability distribution, radius increment probability distribution, and two-dimensional joint increment probability distribution are constructed through kernel density estimation, and are calculated as follows: Calculate the angle and radius increments of adjacent contour points based on the angle and radius sequences: in, It is the first The polar angle of each contour point relative to the centroid. It is the first The polar angle of the next point adjacent to each contour point relative to the centroid. It is the angular increment between adjacent contour points; It is the first The distance from each outline point to the centroid. It is the first The distance from the centroid to the next adjacent point of each contour point. It is the radius increment between adjacent contour points. It is the index of adjacent point pairs. It is the total number of contour points, which ultimately constitute the angle increment sequence. and radius increment sequence ; Construct an evaluation grid based on angle and radius increments: in It is the range of angle increments. It is the angular increment between adjacent contour points. It is the maximum value among all angle increments. It is the minimum value among all angle increments. It is the range of radius increments. It is the radius increment between adjacent contour points. It is the maximum value among all radius increments. It is the minimum value among all radius increments; It is the angle increment fill amount. It is the radius increment fill amount. It is a scaling factor that adds extra padding outside the sample range to ensure that the evaluation grid for kernel density estimation can cover the tail region of the sample distribution; It is an angle increment evaluation grid, used to evaluate the one-dimensional equally spaced grid point sequence for angle increment kernel density estimation. It is a radius increment evaluation grid, used to evaluate a one-dimensional, equally spaced grid point sequence for radius increment kernel density estimation. It is a function that generates equally spaced points. It is the number of grid points; Calculate the density values ​​at grid points using Gaussian kernel density estimation: in It is the kernel density estimation function of the angle increment. It is the independent variable on the angle increment axis. It is the radius increment kernel density estimation function. It is the independent variable on the radius increment axis. The length of the incremental sequence is denoted by , and the total number of contour points minus one is denoted by . For the bandwidth parameter in the direction of angle increment, This refers to the bandwidth parameter in the radius increment direction. It is the angular increment between adjacent contour points. It is the radius increment between adjacent contour points. The Gaussian kernel function has the following form: in Indicates the Gaussian kernel at a point The value at that location, It is a random variable that follows a standard normal distribution. It is the normalization constant of the distribution, ensuring that the integral of the entire probability density function over the entire real number field equals 1. It is the index part; Convert density values ​​to discrete probability values: in Indicates the first Incremental grid points at each angle Discrete probability values ​​at point , It is the angle increment kernel density estimation function at the grid points The value at that location, It is the angle increment grid spacing. It is the number of grid points for the angle increment; An approximation is made by numerical integration over the entire angle increment axis, while summing the density of all grid points by the spacing; Indicates the first Radius increment grid points Discrete probability values ​​at point , It is the radius increment kernel density estimation function at the grid points The value at that location, It is the radius increment grid spacing. It is the number of grid points in the radius increment; An approximation is made by numerical integration over the entire radius increment axis, summing the density of all grid points by the spacing; Construct a two-dimensional evaluation grid based on the angle and radius increments: in It is an angle increment evaluation grid. It is the angular increment between adjacent contour points. It is the maximum value among all angle increments. It is the minimum value among all angle increments. It is the angle increment fill amount; It is a radius increment evaluation grid. It is the radius increment between adjacent contour points. It is the maximum value among all radius increments. It is the minimum of all radius increments. It is the radius increment fill amount; It is a function that generates equally spaced points. It is the number of points in each dimension of the evaluation grid used for two-dimensional joint kernel density estimation; through The two-dimensional mesh generated by the function It is an angle-incremental two-dimensional grid, and each row is... , It is a radius-incremental two-dimensional grid, and each column is... ; Calculate the joint density value at grid points using Gaussian kernel density estimation: in It is the joint kernel density estimation function for angle increment and radius increment. It is the independent variable on the angle increment axis. It is the independent variable on the radius increment axis. The length of the increment sequence. For the bandwidth parameter in the direction of angle increment, This refers to the bandwidth parameter in the radius increment direction. It is the angular increment between adjacent contour points. It is the radius increment between adjacent contour points. The Gaussian kernel function; Convert density values ​​to discrete joint probability values: in Indicates at grid points Discrete joint probability value at , It is the joint kernel density estimation function of angle increment and radius increment at grid points The value at that location, It is the angle increment grid spacing. It is the radius increment grid spacing. It is the number of grid points for angle increment. It is the number of grid points in the radius increment; It is an approximation of the entire two-dimensional grid by numerical integration, multiplying the density of each small grid by its area and then summing them all together.

5. The method according to claim 1, characterized in that... The feature sets are generated by combining the feature sets with the corresponding morphological entropy features, as follows: For each sample, the mean VH, local mean VH, mean VH gradient magnitude, homogeneity of VH gray-level co-occurrence matrix, mean VV, local mean VV, local median VV, homogeneity of VV gray-level co-occurrence matrix, energy, and dissimilarity are extracted from the Sentinel-1 image; the means B1, B2, B4, B7, B8, B8A, and B11, as well as the exponential features NDVI, SAVI, NDVI2, NDRE, and GNDVI are extracted from the Sentinel-2 image. The feature types and calculation methods of the aforementioned Sentinel-1 and Sentinel-2 images are reused on the fused image; For each sample, circumferential morphological entropy, radial morphological entropy, and joint morphological entropy are extracted from the three-channel morphological entropy feature map corresponding to each band, and the three sets of original features extracted above are combined with their corresponding morphological entropy features respectively. The effective coverage rate of the morphological entropy feature record sample set is recorded, the successful extraction ratio of each morphological entropy feature in the sample is calculated, and low-quality feature columns below the preset threshold are deleted. For the remaining morphological entropy features, the mean morphological entropy of the training region is calculated independently by band dimension, and the missing values ​​in the training set and test set are filled using this statistic. Non-morphological entropy features are filled using the mean of the training set, and the range is clipped to eliminate extreme outliers.

6. The method according to claim 1, characterized in that... S4 specifically refers to: A hierarchical K-fold cross-validation method is used to search for hyperparameters to find the optimal n_estimators, and max_features is the integer part of the square root of the total number of features. A random forest classifier is created using the optimal n_estimators value and the calculated max_features value, and then the model is trained using a feature set containing morphological entropy. Predictions are made on the test set and evaluation metrics are calculated, including accuracy, F1 score, and Kappa coefficient.

7. The method according to claim 1, characterized in that... S5 specifically refers to: Load the pre-trained random forest model and automatically parse the required multi-source features from the target image according to the model's feature requirements; divide the entire image into 512×512 blocks, extract feature vectors from all pixels in each block and stack them into a feature matrix; use the model to perform pixel-by-pixel prediction to obtain class labels, and finally output class raster and color rendered image.

Images