A hyperspectral remote sensing image change detection method based on adaptive iterative filtering

Abstract

The application discloses a hyperspectral remote sensing image change detection method based on adaptive iterative filtering, and specifically comprises the following steps: step one, selecting three bands with the largest information quantity from two groups of hyperspectral remote sensing images at different times; step two, writing the selected two groups of bands into the red, green and blue bands of a natural image in the order of the index value from small to large; step three, performing adaptive region growing on the two pictures pixel by pixel, if the difference between the pixel value of the neighborhood and the center pixel is less than the standard deviation of the eight-neighborhood pixels, the pixel is merged into the adaptive region, otherwise, the pixel is not merged; step four, performing region growing on the two filtered pictures in step three again to obtain a change amplitude graph; and step five, calculating the threshold of the obtained change amplitude graph to obtain a final change detection result graph. By adopting the method, the homogenous regions on the picture are more obvious, and the precision of the hyperspectral remote sensing image change detection is improved.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image change detection technology, specifically relating to a hyperspectral remote sensing image change detection method based on adaptive iterative filtering. Background Technology

[0002] Hyperspectral remote sensing imagery boasts high spectral and spatial resolution, encompassing hundreds of bands and possessing rich spectral information, significantly enhancing its value in numerous applications. Change detection using hyperspectral images requires large datasets, and abundant datasets are the most effective way to validate the effectiveness and stability of a change detection method. In recent years, with the rapid development of spaceborne remote sensing technology and the increasing number of satellite launches, the amount of hyperspectral remote sensing data available for research and analysis has been continuously increasing. Hyperspectral remote sensing imagery has been widely applied in various fields such as agricultural development, environmental monitoring, urban planning, and natural disaster assessment. Traditional methods use fixed windows to filter images, leading to blurred boundaries between different ground features and low detection accuracy. Summary of the Invention

[0003] The purpose of this invention is to provide a hyperspectral remote sensing image change detection method based on adaptive iterative filtering, which removes noise from hyperspectral images and makes homogeneous regions in the image more obvious, thereby improving the accuracy of hyperspectral remote sensing image change detection.

[0004] This invention adopts the following technical solution: a method for detecting changes in hyperspectral remote sensing images based on adaptive iterative filtering, the method comprising the following steps:

[0005] Step 1: For two sets of hyperspectral remote sensing images taken by satellite sensors at different times, use a band selection algorithm to select the three bands with the most information.

[0006] Step 2: Write the two groups of bands selected in Step 1 into the red, green, and blue bands of the natural image in ascending order of their index values ​​to form two new images.

[0007] Step 3: Perform adaptive region growth on the two images from Step 2 pixel by pixel. When adding pixels to the adaptive region, the eight neighborhoods of the added pixel must be traversed. If the difference between the pixel value of the neighborhood and the center pixel is less than the standard deviation of the eight neighborhood pixels, then the pixel is merged into the adaptive region; otherwise, it is not merged.

[0008] For each pixel, the mean value of the adaptive region is calculated, and then the mean value is assigned to that pixel. This process is repeated until the entire image is traversed, resulting in two filtered images.

[0009] Step 4: Perform region growing on the two filtered images from Step 3 again, calculate the mean of the adaptive region obtained for each pixel, and then measure the distance between the two means using the transformation vector analysis method. Write this value into a blank image of the same size as the original image. After traversing all images, obtain the transformation amplitude map.

[0010] Step 5: Calculate the threshold for the change amplitude map obtained in Step 4, perform threshold segmentation on the change amplitude map, and obtain the final change detection result map.

[0011] Furthermore, in step three, if the difference between the standard deviation of the adaptive region after adding pixels and the adaptive region without adding pixels is less than 0.1, then the pixel is merged into the current adaptive region; otherwise, it is considered that the pixel has completed region growth, and region growth is performed on the next pixel.

[0012] Furthermore, in step four, the pixel difference within the adaptive region does not exceed 10, and the number of pixels in the adaptive region does not exceed 30.

[0013] Furthermore, in step five, the threshold is calculated using the Otsu method from the variation amplitude graph obtained in step four.

[0014] The beneficial effects of this invention are: (1) Adaptive iterative filtering is used to process the image after band selection. After processing, the homogeneous areas in the image are more obvious, and the boundaries between different land features are clearer. It is possible to better calculate the changed and unchanged areas, which is also the main reason for the improved accuracy. (2) There is no need for a manual sample labeling program, which reduces the cost of labeling samples. (3) This method only limits the difference between the standard deviation of the adaptive region and the adaptive region without added pixels, and the difference is 0.1, which has a high degree of automation. Attached Figure Description

[0015] Figure 1 Example data illustration: 1a and 1b are hyperspectral remote sensing images of a farmland taken in Yancheng, Jiangsu Province in 2006 and 2007, respectively; 1c is the ground reference ground value.

[0016] Figure 2 Visual comparison results for different methods: 2a shows the results under the TDRD method, 2b shows the results under the GETNET method, 2c shows the results under the PTCD method, and 2d shows the results under the method of this invention. Detailed Implementation

[0017] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0018] This invention discloses a method for detecting changes in hyperspectral remote sensing images based on adaptive iterative filtering. The method includes the following steps:

[0019] Step 1: For two sets of hyperspectral remote sensing images taken by satellite sensors at different times, use a band selection algorithm to select the three bands with the most information.

[0020] Step 2: Write the two groups of bands selected in Step 1 into the red, green, and blue bands of the natural image in ascending order of their index values ​​to form two new images.

[0021] Step 3: Perform adaptive region growth on the two images from Step 2 pixel by pixel. When adding pixels to the adaptive region, the eight neighboring regions of the added pixel must be traversed. If the difference between the pixel value of the neighboring region and the center pixel is less than the standard deviation of the eight neighboring pixels, then the pixel is merged into the adaptive region; otherwise, it is not merged.

[0022] For each pixel, the mean value of the adaptive region is calculated, and then the mean value is assigned to that pixel. This process is repeated until the entire image is traversed, resulting in two filtered images.

[0023] Adaptive iterative filtering was used to process the band-selected images. After processing, homogeneous areas in the images became more obvious, the boundaries between different ground features became clearer, and the changed and unchanged areas could be calculated more effectively.

[0024] Step 4: Perform region growing on the two filtered images from Step 3 again, calculate the mean of the adaptive region obtained for each pixel, and then measure the distance between the two means using the transformation vector analysis method. Write this value into a blank image of the same size as the original image. After traversing the entire image, obtain the change amplitude map. According to the comparative experiment, the pixel difference within the adaptive region does not exceed 10, and the number of pixels in the adaptive region does not exceed 30.

[0025] Step 5: Calculate the threshold of the change amplitude map obtained in Step 4, perform threshold segmentation on the change amplitude map, and obtain the final change detection result map.

[0026] In step three above, if the difference between the standard deviation of the adaptive region after adding pixels and the adaptive region without adding pixels is less than 0.1, then the pixel is merged into the current adaptive region; otherwise, it is considered that the pixel has completed region growth, and region growth is performed on the next pixel.

[0027] In step four above, the adaptive region obtained for each pixel is measured using the change vector analysis method to obtain the corresponding value.

[0028] In step five above, the threshold is calculated using the Otsu method based on the change amplitude graph obtained in step four.

[0029] To better verify the hyperspectral remote sensing image change detection method based on adaptive iterative filtering in this invention, the following experiment was conducted:

[0030] The dataset used in this embodiment is as follows: The "Farmland" dataset consists of a pair of hyperspectral remote sensing images acquired by the hyperspectral sensor on the Earth Observer 1 satellite at two different time periods. The study area is wetland agricultural land, exhibiting typical "change" characteristics. The ground reference ground value is as follows: Figure 1 As shown.

[0031] For two sets of hyperspectral remote sensing images captured by satellite sensors at different times, a band selection algorithm was used to select the three bands with the most information to reduce the dimensionality of the hyperspectral remote sensing images.

[0032] The two selected bands are written into the red, green, and blue bands of the natural image in ascending order of their index values ​​to form two new images.

[0033] The two new images are adaptively grown pixel by pixel. When adding pixels to the adaptive region, the eight neighborhoods of the added pixel are traversed. If the difference between the pixel value of the neighborhood and the center pixel is less than the standard deviation of the eight neighborhood pixels, the pixel is merged into the adaptive region; otherwise, it is not merged.

[0034] For each pixel, the mean value of the adaptive region is calculated, and then the mean value is assigned to that pixel. This process is repeated until the entire image is traversed, resulting in two filtered images, which clearly define the boundaries between different features.

[0035] In this step, a hyperparameter of 0.1 is set. If the difference between the standard deviation of the adaptive region after adding pixels and the adaptive region without adding pixels is less than 0.1, the pixel is merged into the current adaptive region. Otherwise, it is considered that the region growth of the pixel has been completed, and the region growth is performed on the next pixel.

[0036] Step 4: Perform region growing on the two filtered images from Step 3 again, calculate the mean of the adaptive region for each pixel, and then measure the distance between the two means using the transformation vector analysis method. Write this value into a blank image of the same size as the original image. After traversing the entire image, obtain the change amplitude map. According to the comparative experiment, the pixel difference within the adaptive region is 10, and the number of pixels in the adaptive region is not 30.

[0037] Step 5: Calculate the threshold of the change amplitude map obtained in Step 4 using the Otsu method, and perform threshold segmentation on the change amplitude map to obtain the final change detection result map.

[0038] Figure 2 Visualized change detection results under different methods are presented, where 2a-2c are results from existing methods, and 2d is the result from the method proposed in this invention. Since the method of this invention includes only two parameters—the pixel difference in the adaptive region and the number of pixels in the adaptive region—it can be considered parameterless. The parameters in the comparison methods are as follows:

[0039] TDRD:PC:0.965.

[0040] GETNET: Sample:0.7.

[0041] PTCD: WindowSize: 3*3.

[0042] Compare the obtained result graph with Figure 1 The accuracy of each method is compared by 1c, and the accuracy comparison is shown in Table 1:

[0043] Table 1. Accuracy Comparison Results of Different Methods

[0044]

[0045] As shown in Table 1, the method described in this invention outperforms the other three methods in terms of overall precision, Kappa coefficient, average precision, recall, and F1-score. Overall precision, the proportion of correctly classified samples out of all samples, is 96.37, indicating that 96.37% of pixels were correctly classified. The Kappa coefficient measures the classification effectiveness, ranging from -1 to 1; a higher value indicates better classification. Average precision is the average precision across different categories; this method achieves an average precision of 95.49. This method outperforms the other three methods in both overall precision and average precision, demonstrating its ability to improve change detection accuracy. Recall indicates the percentage of positive cases correctly predicted; a higher value indicates better classification.

Claims

1. A hyperspectral remote sensing image change detection method based on adaptive iterative filtering, characterized in that, The detection method comprises the following steps: Step one, two groups of hyperspectral remote sensing images at different times taken by satellite sensors are selected by a band selection algorithm to obtain three bands with the largest information amount; Step two, the two groups of bands selected in step one are written into the red, green and blue bands of a natural image in order of their index values from small to large, forming two new images; Step three, the two images in step two are subjected to adaptive region growth pixel by pixel, and when adding a pixel to the adaptive region, the eight neighborhoods of the added pixel are traversed, if the difference between the pixel value of the neighborhood and the center pixel is less than the standard deviation of the eight neighborhoods, the pixel is merged into the adaptive region, otherwise it is not merged; The mean value of the adaptive region obtained for each pixel is calculated, and the mean value is assigned to the pixel, until the entire image is traversed to obtain two filtered images; Step four, the two filtered images in step three are subjected to region growth again, the mean value of the adaptive region obtained for each pixel is calculated, and the two mean values are measured by a change vector analysis method to obtain the distance value between them, which is written into a blank image with the same size as the original image, and the change amplitude map is obtained after traversing the entire image; Step five, the change amplitude map obtained in step four is calculated to obtain a threshold, and the change amplitude map is threshold segmented to obtain the final change detection result map; In step three, if the difference between the standard deviation of the adaptive region after adding the pixel and the adaptive region before adding the pixel is less than 0.1, the pixel is merged into the current adaptive region, otherwise, it is considered that the pixel has completed region growth, and the next pixel is subjected to region growth; In step four, the difference between the pixels in the adaptive region does not exceed 10, and the number of pixels in the adaptive region does not exceed 30.

2. The hyperspectral remote sensing image change detection method based on adaptive iterative filtering according to claim 1, characterized in that, In step five, the change amplitude map obtained in step four is calculated to obtain a threshold by the Otsu method.

Images