An adaptive high-resolution remote sensing image supervised classification method combining spatial-spectral features

Abstract

This invention discloses a supervised classification method for high-resolution remote sensing images based on adaptive spatial-spectral feature integration, comprising the following steps: Step 1: Randomly and manually label multiple pixels of each land cover class from the ground reference ground value as initial training sample labels; record the position coordinates of the initial training samples in the high-resolution remote sensing image; and perform adaptive region growing on each pixel, extracting spatial and spectral features from the grown region of each pixel; Step 2: Use the initial training sample labels and spatial-spectral feature map from Step 1 as training data, input them into an SVM classifier for classification, and obtain an intermediate classification map; Step 3: Use the intermediate classification map from Step 2 as the prior probability of a Markov random field, combine it with the high-resolution remote sensing image for segmentation, and obtain the result map. This supervised classification method has good stability and robustness, effectively reducing the manual and time costs of labeling samples.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image classification technology, specifically relating to a high-resolution remote sensing image supervised classification method that combines adaptive spatial-spectral features. Background Technology

[0002] In recent years, with the rapid development of satellite and aerial remote sensing technologies, the temporal and spatial resolution of imagery has been significantly improved. Through the analysis and processing of remote sensing and aerial images, land cover information can be obtained quickly and effectively, providing data support for human scientific understanding of the Earth. This is of great significance for environmental monitoring, cadastral surveys, precision agriculture, urban planning, government macro-control, and policy formulation. However, while the improved image resolution has indeed enhanced the ability to acquire information about land features and provided richer details, the overlapping spectra of different land features, along with numerous instances of "different spectra for the same object" and "the same object with different spectra," increases intra-class variance and decreases inter-class variance. This results in significant salt-and-pepper noise and numerous misclassifications when traditional classification methods are applied, greatly reducing the reliability and accuracy of the classification results. Summary of the Invention

[0003] The purpose of this invention is to provide a high-resolution remote sensing image supervised classification method with adaptive spatial-spectral feature combination, which has good stability and robustness, and effectively reduces the manual and time costs of labeling samples.

[0004] This invention employs the following technical solution: a supervised classification method for high-resolution remote sensing images based on adaptive spatial-spectral feature joint methods, comprising the following steps:

[0005] Step 1: Randomly and manually label multiple pixels of each land cover type from the ground reference ground value as initial training sample labels; and record the location coordinates of the initial training samples in the high-resolution remote sensing image.

[0006] Adaptive region growing is performed on each pixel, spatial and spectral features are extracted from the grown region of each pixel, and the spatial and spectral features are combined as the spatial-spectral joint features of the pixel.

[0007] Step 2: Use the initial training sample labels and spatial-spectral feature maps from Step 1 as training data, input them into the SVM classifier for classification, and obtain intermediate classification maps;

[0008] Step 3: Use the intermediate classification image from Step 2 as the prior probability of the Markov random field, and combine it with the high-resolution remote sensing image for segmentation to obtain the result image.

[0009] Furthermore, in step one, the number of pixels for each type of land feature is less than 50.

[0010] Further, in step one, region growing is performed on each pixel of each type of land cover. The growing rule is to traverse the eight neighboring pixels of the center pixel. If the difference between the neighboring pixel value and the center pixel value is less than T1, the pixel is merged into the growing region; otherwise, the pixel is not merged. If the number of pixels in the growing region is greater than T2, the region growing ends. Here, T1 is the minimum inter-class variance, i.e. T2 is represented as a scale factor in feature extraction, where: N is the number of land cover categories, L... i For the i-th type of land feature pixel, Std i Let m be the standard deviation of the pixels of the i-th type of land cover, and m be the standard deviation of the L-th type i Number of pixels.

[0011] Furthermore, spatial and spectral features are extracted from the growth region of each pixel, where the formula for calculating spatial features is: The formula for calculating spectral characteristics is: Among them W i，j The calculation formula is: i,j are the coordinates of the growth point where the region is grown from this point, Region represents the region where the region is grown from pixels i and j, and m,n are the coordinates of the region where the region is grown from the growth point i and j.

[0012] The beneficial effects of this invention are: (1) With a small number of manually labeled samples, the classification accuracy is still very good, and the difference between the overall accuracy and the average accuracy obtained by classifying 10 training samples per class and 120 training samples per class is within 3% and 6% respectively, showing good stability and robustness, effectively reducing the manual and time costs of labeling samples. (2) When the number of training samples gradually increases, the accuracy of this method always remains within 5%, indicating that this method is suitable for a small number of samples and has good stability. (3) Only two parameters, T1 and T2, limit the growth of the region in the process, resulting in a high degree of automation. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating the method for implementing the present invention.

[0014] Figure 2 This is a schematic diagram of data from an example embodiment;

[0015] 2a is a high-resolution aerial remote sensing image;

[0016] 2b is the ground reference truth map.

[0017] Figure 3 The diagram shows the performance of the method in this invention and seven comparative methods under the SVM classifier.

[0018] Figure 4This diagram illustrates the impact of different sample sizes on accuracy in high-resolution aerial remote sensing imagery.

[0019] 4a shows the impact of different training sample sizes on overall accuracy;

[0020] Figure 4b shows the impact of different training sample sizes on average accuracy. Detailed Implementation

[0021] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0022] Using a set of widely used high-resolution remote sensing images as test data,

[0023] Step 1: Randomly and manually label 50 pixels of each land cover type from the ground reference ground value as initial training sample labels; and record the location coordinates of the initial training samples in the high-resolution remote sensing image.

[0024] Adaptive region growing is performed on each pixel of the high-resolution remote sensing image, and spatial and spectral features are extracted from the grown region of each pixel. The spatial and spectral features are combined as the spatial-spectral joint features of the pixel to obtain a spatial-spectral feature map with the same size as the high-resolution remote sensing image.

[0025] Step 2: Use the initial training sample labels and spatial-spectral feature maps from Step 1 as training data, input them into the SVM classifier for classification, and obtain intermediate classification maps;

[0026] Step 3: Use the intermediate classification image from Step 2 as the prior probability of the Markov random field, and combine it with the high-resolution remote sensing image for segmentation to obtain the result image.

[0027] To better understand the technical solution of the present invention, the following description, in conjunction with the accompanying drawings and embodiments, uses high-resolution aerial remote sensing imagery as an example for verification, following the flowchart. Figure 1 The detailed process of this embodiment is as follows:

[0028] The high-resolution remote sensing image used has a spatial resolution of 0.32 meters per pixel, and its false-color image is as follows: Figure 2 As shown in Figure 2a, the ground reference true value is as follows: Figure 2 As shown in Figure 2b.

[0029] Step 1: Randomly select 50 samples of each type of land cover from the ground reference ground value as the initial training sample labels, and record the position coordinates of the initial training samples in the high-resolution remote sensing image.

[0030] Adaptive region growing is performed on each pixel, spatial and spectral features are extracted from the grown region of each pixel, and the spatial and spectral features are combined as the spatial-spectral joint features of the pixel.

[0031] For each pixel of each land cover class in the high-resolution remote sensing image, region growing is performed. The growing rule is to traverse the eight neighboring pixels of the center pixel. If the difference between the neighboring pixel value and the center pixel value is less than T1, the pixel is merged into the growing region; otherwise, the pixel is not merged. If the number of pixels in the growing region is greater than T2, the region growing ends. Here, T1 is the minimum inter-class variance, i.e. T2 acts as a scale factor in feature extraction. Experimental results show that after feature normalization, different T2 values ​​have little impact on feature extraction, indicating that this feature can be considered a scale-invariant operator. Where N is the number of land cover categories, L... i For the i-th type of land feature pixel, STd i Let m be the standard deviation of the pixels of the i-th type of land cover. i Number of pixels.

[0032] Spatial and spectral features are extracted for the growth region of each pixel, where the formula for calculating spatial features is: The formula for calculating spectral characteristics is: Among them W i，j The calculation formula is: i,j are the coordinates of the growth point where the region is grown from this point, Region represents the region where the region is grown from pixels i,j, and m,n are the coordinates of the region where the region is grown from the growth point i,j.

[0033] Step 2: Concatenate the initial training sample labels from Step 1 with the spatial and spectral features as the joint spatial-spectral features and input them into the SVM classifier to obtain the intermediate classification map.

[0034] Step 3: Use the intermediate classification image from Step 2 as the prior probability of the Markov random field, and segment it with the original high-resolution remote sensing image to obtain the result image. Color the result image to obtain the classification image, as shown below. Figure 3 As shown in Figure 3h.

[0035] in, Figure 3 The document presents comparative visualizations of different methods, with 3a-3g representing the results of existing methods. Since T1 and T2 in the method of this invention are affected by the size of the high-resolution remote sensing image and the number of land cover categories, the method of this invention can be considered parameter-free. The specific parameters in the comparison method are as follows:

[0036] ·Zhang, L: T1=40, T2=100, D=20.

[0037] ·Benediktsson, JA: SE is a "disk", size =: "2×2, 4×4, 6×6".

[0038] ·Lv,ZY:SE={'disk','line','square','diamond'}.

[0039] ·Kang,X:δ s =4,δ r =0.2, r=4, ∈=0.01.

[0040] ·Kang,X:δ s =200,δ r =30, integration=3.

[0041] ·Kang,X:δ s =4.0,δ r =0.1, integration=5.

[0042] ·Liu, S: k=7, z, r∈[3,10], δ=0.02, GFF: r1=45, δ1=0.3, r2=7, δ2=10 -6 .

[0043] Compare the obtained result graph with Figure 2 The accuracy of each method was obtained by comparing it with 2b in the table. The accuracy comparison results are shown in Table 1.

[0044] Table 1. Accuracy Comparison Results of Different Methods

[0045]

[0046] As shown in Table 1, the method described in this invention outperforms the other seven methods in terms of overall accuracy, average accuracy, Kappa coefficient, F1-score, and SDUA coefficient. Overall accuracy, calculated as the sum of correctly classified pixels divided by the total number of pixels, achieves an overall accuracy of 99.22%, with 99.22% of pixels correctly classified. Average accuracy, the average of accuracy across all feature categories, reaches 98.50. This demonstrates that the method not only performs well overall but also outperforms the other seven methods in each category. SDUA, the standard deviation of accuracy across different feature categories, measures the distribution of accuracy values ​​and describes the deviation between accuracy categories. The SDUA coefficient of the method described in this invention is 1.37, indicating a small standard deviation and relatively concentrated accuracy.

[0047] To verify the method in this invention, since the overall accuracy and average accuracy of different numbers of training samples do not change much, the following experiment was conducted. Using the method in this invention, a high-resolution remote sensing image was selected, and each type of land cover was randomly and manually labeled on its ground reference ground map. 10-120 pixels were labeled as initial training sample labels for verification. Figure 4 The trends of overall accuracy and average accuracy of the method in this invention under different sample sizes are presented. Figure 4 It can be seen that, when using the method of this invention and classifying using RGB features with the same amount of training samples, the overall accuracy and average accuracy achieved by the method of this invention are approximately 9% and 6% higher, respectively, than those achieved using RGB features. Furthermore, when the training samples are increased from 10 training samples per land cover class to 120 training samples per class, the overall accuracy and average accuracy change by 3% and 6%, respectively. Therefore, it can be concluded that the method of this invention possesses a certain degree of robustness and stability.

Claims

1. A supervised classification method for high-resolution remote sensing images with adaptive spatial-spectral feature integration, characterized in that, The supervised classification method includes the following steps: Step 1: Randomly and manually label multiple pixels of each land cover type from the ground reference ground value as initial training sample labels; and record the location coordinates of the initial training samples in the high-resolution remote sensing image. Adaptive region growing is performed on each pixel, spatial and spectral features are extracted from the grown region of each pixel, and the spatial and spectral features are combined as the spatial-spectral joint features of the pixel. Step 2: Use the initial training sample labels and spatial-spectral feature maps from Step 1 as training data, input them into the SVM classifier for classification, and obtain intermediate classification maps; Step 3: Use the intermediate classification image from Step 2 as the prior probability of the Markov random field, and combine it with the high-resolution remote sensing image for segmentation to obtain the result image; Spatial and spectral features are extracted from the growth region of each pixel, respectively. The formula for calculating spatial features is: The formula for calculating spectral characteristics is: ,in The calculation formula is: , , These are the coordinates of the growth point for region growth based on this pixel. Indicated by , The region where pixels grow. , For , The coordinates of the growth point within the growth region.

2. The adaptive spatial-spectral feature joint high-resolution remote sensing image supervised classification method as described in claim 1, characterized in that, In step one, the number of pixels for each type of land cover is less than 50.

3. A high-resolution remote sensing image supervised classification method based on adaptive spatial-spectral feature joint as described in claim 1 or 2, characterized in that, In step one, region growing is performed on each pixel of each type of land cover. The growing rule is to traverse the eight neighboring pixels of the center pixel. If the difference between the neighboring pixel value and the center pixel value is less than T1, the pixel is merged into the growing region; otherwise, the pixel is not merged. If the number of pixels in the growing region is greater than T2, the region growing ends. Here, T1 is the minimum inter-class variance, i.e. T2 acts as a scaling factor in feature extraction, where: The number of land cover categories, For the i-th type of land feature pixel, For the first Standard deviation of pixels for land-like features.

Images