CN100390566C - A Method of Detecting Surface Change Based on Remote Sensing Image Processing - Google Patents

Abstract

The present invention relates to a surface change detection method based on remote sensing image treatment, which comprises the following steps: step 1, a standard image and an image to be detected are respectively separated to get separated result images, and each separated region in the separated result images has the characteristic value of a regional gray scale and the characteristic value of a regional area; step 2, the change ranges of the regional gray scale and the regional area which are corresponding to the separated result image are detected point-by-point to judge if change happens in the regional gray scale of a corresponding point and change does not happen in the regional area, the change is false change, namely, the change does not happen in actuality, if the change happens in the regional area of the corresponding point and the change does not happen in the regional gray scale, the change is false change, and if the change happens in the gray scale and the change happens in the area, small change quantity is adopted as measure value which is used for determining corresponding surface change. The adoption of the method can gain a more correct detection result to the surface change by remote sensing images.

Description

A Method of Detecting Surface Change Based on Remote Sensing Image Processing

technical field

The invention relates to image processing technology and its application, in particular to a method for detecting changes in the ground surface based on remote sensing image processing. By using this method, the authenticity of the changed area can be distinguished more conveniently, so that a more accurate detection result of the surface change can be obtained through the remote sensing image.

Background technique

With the further development of sensor technology, aviation and aerospace platform technology, and data communication technology, the information provided by remote sensing will increase at an unimaginable rate. Theories and technologies such as how to accurately and quickly use remote sensing data to find changes, and how to automatically extract changing areas from images are of great significance to the future development of remote sensing technology. At present, the commonly used change detection methods are mainly divided into the following two categories: one is the change detection method based on gray scale; the other is the change detection method based on segmentation, that is, the change detection is performed according to the image segmentation results before and after the change.

The grayscale-based change detection method subtracts the grayscale values of corresponding pixels in different temporal images, and the resulting image represents the change of the two temporal images. The characteristic of this method is that the operation is simple and direct, and it can detect all the changing areas on the image. However, there are many reasons for image grayscale changes, such as different seasons, changes in the sun's altitude angle, etc., so there will be many false change areas in the detection results of the change detection method based solely on grayscale.

The change detection method based on segmentation refers to segmenting each image separately, and then judging the changed area according to the difference of the corresponding pixel category. An important advancement of segmentation-based change detection is that it can overcome the inconvenience caused by differences in sensor properties, resolution, etc. of multi-temporal images, and no data normalization process is required because the two images are segmented separately. There are already many methods for image segmentation, such as clustering, edge detection, and region growth. These methods mainly divide the image into different regions according to the grayscale of the image, while the Watersheds algorithm not only considers grayscale features in the segmentation process, but also considers features such as Neighborhood features of elements. Therefore, the Watersheds algorithm has attracted more and more attention from experts.

The Watersheds algorithm regards the image as an undulating mountain model, where the gray value of each pixel represents the altitude of the point. In such terrain, there are not only low-lying valleys (locally small areas), but also high-rise ridges (watersheds), and steep or gentle hillsides (water storage basins) between the ridges and valley bottoms. To simulate flooding is to puncture a small hole at the bottom of each valley, and then slowly immerse the whole model in the water. As the water level rises, the water surface slowly expands up the hillside, and when it reaches the ridge, it will overflow, and it is here Build dikes, and so on until the entire model is submerged. All dikes then become watersheds separating the various storage basins.

The Watersheds algorithm implementation method includes the following two steps:

(1) Pixel sorting: scan the image, and establish the index of the pixel gray value from low to high, so as to directly access the pixels of the same gray value;

(2) Simulate water immersion: Simulate water immersion starting from the lowest value h0 of the gray value of the pixel, assuming that the water storage basins to which the pixels with a gray level less than or equal to h have been identified, then processing the gray level h+1 pixels, the pixels adjacent to the marked water storage basins in this gray level are sent to a first-in-first-out (FIFO) queue, and then starting from these pixels, according to the geodesic distance, the marked water storage basins are The basin is extended to h+1 gray level. If there are unmarked pixels in h+1 gray level, it will be regarded as a newly emerging local minimum area and given a new area label (label). Finally, in the simulated water immersion results, the pixels with the same label belong to the same water storage basin, and the pixels with the same distance from different water storage basins are marked as watershed points.

The important feature of the Watersheds algorithm is to combine the gray level information of the pixel with the neighborhood information. As the gray level increases, the neighbor pixels of the previous pixel are analyzed one by one until all the pixels are analyzed. For the eight-neighborhood case, there are 8 neighbor points that affect the segmentation result of the current point, which are easily affected by noise and cause over-segmentation, and merging over-segmented regions will consume a lot of time and memory. Therefore, when segmenting large-scale remote sensing images, the wide application of this algorithm is largely limited.

Contents of the invention

Aiming at the defects or deficiencies in the prior art, the present invention provides a method for detecting changes in the ground surface based on remote sensing image processing. With this method, the authenticity of the changed area can be distinguished more conveniently, so that the remote sensing image can be used to obtain the correctness. The more accurate detection results of surface changes, or in other words, make the detection results more in line with objective reality.

The general technical concept of the present invention is to obtain the feature value of each region through image segmentation, including two feature values of the area of the region and the gray level of the region, and perform change detection according to the changes of these two quantities, so as to obtain the corresponding value through the remote sensing image. More accurate detection results of surface changes. For example, image segmentation is performed on the two images for change detection to obtain the segmentation result image; the difference method is used to detect the change of the gray level and area of each point of the result image; if the gray level of the corresponding point changes , but the area of the region does not change, then the change is a pseudo-change (actually no change), on the contrary, if the area of the corresponding point changes, but the gray level does not change, it is also a pseudo-change. If both grayscale and area have changed, take the smaller amount as the measurement value to determine the surface change.

Technical scheme of the present invention is as follows:

The method for detecting surface changes based on remote sensing image processing is characterized in that it includes the following steps: Step 1, image segmentation is performed on the reference image and the image to be detected respectively to obtain a segmentation result image, and each segmentation area in the result image has Regional grayscale eigenvalues and regional area eigenvalues; step 2, detect the range of change of the regional grayscale and regional area of the corresponding segmentation result image point by point and make the following judgments: if the regional grayscale of the corresponding point changes, but the area If the area does not change, the change is a pseudo-change, that is, there is no actual change; on the contrary, if the area of the corresponding point changes, but the gray level of the area does not change, it is also a pseudo-change; if both the gray level and the area have changed, Then the small amount of change is used as the measurement value to determine the corresponding surface change.

The method for processing remote sensing images used for detecting surface changes is characterized in that it includes performing image segmentation on the remote sensing images, the image segmentation includes initial segmentation and region merging, performing regional identification on each segmented region formed by the initial segmentation, and For each merged region formed by region merging, the gray value of the region and the characteristic value of the region area are determined.

The area identification includes geometric features.

The geometric feature refers to the area of the region.

The region identification also includes region position, number of neighbors, gray mean value of the region, and neighborhood marks.

The initial segmentation includes the following steps: step A, performing wavelet transformation on the image, and calculating the gradient of the lowest resolution image; step B, performing initial segmentation on the gradient result map using watershed transformation to form several initial segmentation regions, and for each region Assign a mark.

The region merging is carried out according to the similarity of the gray mean value of the region, if the square difference of the gray mean value of two adjacent regions is less than a certain threshold, then merge, and recalculate the feature value of the new region; when any two similar When the gray similarity of the adjacent regions is greater than a certain threshold, the merging is completed; the low-resolution image formed after the region merging is subjected to inverse wavelet transform to reconstruct the original resolution image.

The wavelet transform in the step A adopts the Haar wavelet, the gradient operator in the gradient calculation in the step A adopts the Robert operator, and the step B includes: scanning the gradient result image, and the pixel increases according to the gradient value The index value of each level of gradient is established sequentially; starting from the lowest gradient value, classification is performed according to the gradient value and the neighborhood label of the pixel.

The classification according to the gradient value and the neighborhood labeling of the pixels refers to: starting from the lowest value _h0 of the gray value of the pixel to simulate flooding, assuming that the water storage basins to which the pixels with a gray level less than or equal to h belong have been identified Then, when processing pixels of h+1 gray level, the pixels adjacent to the marked water storage basin in this gray level are sent to a first-in first-out (FIFO) queue, and then start from these pixels, According to the geodesic distance, the marked water storage basin is extended to h+1 gray level. If there are unmarked pixels in the h+1 gray level, it will be used as a new local minimum area and assigned to a new area. label; until all pixels are traversed.

The reconstruction of the original resolution image includes the following steps: step a, performing inverse wavelet transform on image M _L to obtain M _L-1 ; step b, performing Robert gradient operation on image M _L-1 and then performing watershed transform, and converting the inner region Set it to 0, set the boundary to 255, and get the binary image B _L-1 ; step c, refine the boundary and area of M L _-1 according to the boundary line of B _L-1 , that is: combine B _L-1 and M _L-1 corresponds point by point, and counts the number of points in each area of B _L-1 mapped to the area of M _L-1 ; counts the mark value of the area in M _L-1 with the most mapping points; assigns the mark value to The corresponding area of B _L-1 ; step d, repeat the above three processes until L=0.

Technical effect of the present invention is as follows:

The purpose of the present invention is to overcome the deficiencies of the prior art, and combine two methods of gray scale detection and image segmentation on the basis of the original method. Combined with wavelet transform in the image segmentation process, two processes of initial segmentation and merging regions are performed on low-resolution images. On the premise of increasing the image segmentation process, it effectively reduces the amount of calculation and operation time, and weakens the algorithm. Sensitivity to Noise. In the process of change detection, the changes of grayscale and regional features are detected simultaneously, so that the false change regions in the detection results are reduced.

The method of the present invention includes two parts of image segmentation and change detection. Image segmentation obtains the feature value of each region, including the area of the region and the gray level of the region, and changes are detected according to the changes of these two quantities.

Compared with the prior art, the present invention has following characteristics:

First: In the segmentation process, the wavelet transform and the Watersheds algorithm are combined to perform segmentation on the lowest resolution image, which reduces the amount of calculation and saves calculation time;

Second: Calculate the gradient before the initial segmentation, so that the segmentation process is actually to extract the boundaries in the gradient image, and can accurately locate the boundaries;

Third: The low-resolution image is projected to the original resolution image by using inverse wavelet transform, which solves the problem of thickening and jagged boundaries caused by directly enlarging the image;

Fourth: When performing change detection, detect changes in grayscale and area to make the detection results more in line with objective reality.

In summary, segmentation and grayscale detection methods are combined for change detection. In the segmentation process, combined with Watershed Transform (Watersheds Transform) and wavelet transform, the initial segmentation and region merging are performed on the low-resolution image, and the original resolution image is projected back to the original resolution image through the wavelet inverse transform to obtain the segmentation result, which reduces the amount of calculation and also Reduced sensitivity to noise. In the process of change detection, the change of regional gray level and region area is detected, and false changes are reduced.

Description of drawings

Fig. 1 is the flow chart of image segmentation in the present invention.

Fig. 2 is a regional structure diagram of the present invention.

Fig. 3 is a flow chart of change detection in the present invention.

Detailed ways

Implementation examples of implementing change detection by using the method of the present invention are shown in FIGS. 1 to 3 , which will now be described in detail with reference to the figures.

The realization process of image segmentation of the present invention is as shown in Figure 1, and its work process is:

1) Haar wavelet transform is performed on the entire image, and the gradient of the lowest resolution image is calculated;

2) Perform initial segmentation with Watersheds transformation on the gradient image to form many initial segmentation regions, and assign a label to each region;

3) After the initial segmentation is completed, the feature values of the statistics area include: area, mean value, position, number of neighbors and neighbor label (label);

4) Merge segmented regions: Merge regions according to the gray similarity of adjacent regions. If the square difference of the gray mean values of adjacent regions i and j is less than a certain threshold value d, merge regions i and j and recalculate Eigenvalues for the new region. When the gray similarity of any two adjacent regions is greater than d, the merging is completed; the value d is related to the image. The larger the threshold is, the more regions are merged, and the number of regions in the segmentation result is smaller; otherwise, the smaller the threshold is, the more regions are in the segmentation result.

5) Regional projection: The above three steps are all performed on low-resolution images. In order to reconstruct the original resolution image, regional projection is required. If the image is directly enlarged to the original scale, there will be stepped jagged edges, and image information will be lost during the resampling process. In order to solve the above problems, we perform wavelet inverse transform on the segmented low-resolution image until the full resolution is reached. The specific process is as follows:

(1) Perform inverse wavelet transform on the image M _L to obtain M _L-1 .

(2) Carry out Robert's gradient calculation on the image M _L-1 and then perform watershed transformation, set the interior of the region to 0, and the boundary to 255 to obtain a binary image B _L-1 .

(3) Refine the boundary and area of M _L-1 according to the boundary line of BL _-1 . The refinement process is as follows:

a. Corresponding BL _-1 and ML _-1 point by point, counting the number of points in the area where each area of BL _-1 is mapped to ML _-1 ;

b. Statistically calculate the label value (label) in the M _L-1 area with the most mapping points;

c. Assign the label value to the corresponding area of B _L-1 .

(4) The result obtained in the above process is I _L-1 of the image pyramid. Repeat the above process until L=0.

The regional structure of the present invention is shown in Figure 2, and it comprises the mark (label) of region and regional characteristic parameter, and regional characteristic parameter has 5: regional area, regional position, regional mean value, neighborhood number and neighborhood label.

The realization process of change detection described in the present invention is as shown in Figure 3, and its work process is:

(1) Segment the two images for change detection respectively to obtain the segmentation result image;

(2) Use the difference method to detect the change of the gray level and area of each point of the result image; if the gray level of the corresponding point changes, but the area of the area does not change, then the change is a false change (actual No change), on the contrary, if the area of the corresponding point changes, but the gray level does not change, it is also a pseudo-change. If both the gray scale and the area have changed, take the smaller amount as the change.

It should be pointed out that the specific embodiments described above can enable those skilled in the art to understand the present invention more comprehensively, but do not limit the present invention in any way. Therefore, although the specification has described the present invention in detail with reference to the accompanying drawings and embodiments, those skilled in the art should understand that the present invention can still be modified or equivalently replaced; and everything does not depart from the spirit and technical essence of the present invention The technical solutions and their improvements shall be covered by the protection scope of the patent of the present invention.

Claims (10)

1. The method for detecting surface changes based on remote sensing image processing is characterized in that it comprises the following steps: Step 1, image segmentation is carried out respectively to reference image and image to be detected, obtains segmentation result image, each segmentation in this result image The region has regional grayscale eigenvalues and regional area eigenvalues; step 2, detect the region grayscale and region area change range of the corresponding segmentation result image point by point and make the following judgments: if the regional grayscale of the corresponding point changes, However, if the area area does not change, the change is a pseudo-change, that is, there is no actual change; on the contrary, if the area area of the corresponding point changes, but the gray level of the area does not change, it is also a pseudo-change; if both the gray level and the area have occurred change, the small change is used as the measurement value to determine the corresponding surface change. 2. The method for processing remote sensing images used to detect surface changes, characterized in that it includes image segmentation of remote sensing images, the image segmentation includes initial segmentation and regional merging, and regional identification is carried out for each segmented area formed by the initial segmentation , to determine the regional gray value and regional area characteristic value for each merged region formed by region merger. 3. The method for processing remote sensing images used for detecting surface changes according to claim 2, characterized in that: the area identification includes geometric features. 4. The method for processing remote sensing images for detecting surface changes according to claim 3, wherein the geometric feature refers to the area of the region. 5. The method for processing remote sensing images for detecting surface changes according to claim 3, characterized in that: the region identification also includes region position, neighborhood number, region gray mean value, and/or neighborhood domain tag. 6. The processing method for remote sensing images used for detecting surface changes according to claim 2, characterized in that: said initial segmentation comprises the following steps: step A, carrying out wavelet transform to images, and calculating the minimum resolution image Gradient; step B, initial segmentation is performed on the gradient result map using watershed transformation to form several initial segmentation regions, and each region is assigned a mark. 7. The method for processing remote sensing images for detecting surface changes according to claim 2, characterized in that: said region merging is performed according to the similarity of the gray mean value of the region, if the gray values of two adjacent regions If the square difference of the degree mean value is less than a certain threshold, then merge and recalculate the feature value of the new region; when the gray similarity of any two adjacent regions is greater than a certain threshold, the merger is completed; Inverse wavelet transform is performed on the low-resolution image to reconstruct the original-resolution image. 8. the processing method for the remote sensing image that is used to detect surface change according to claim 6, is characterized in that: the wavelet transform in the described step A adopts Haar wavelet, the gradient calculation of the gradient calculation in the described step A The Robert operator is adopted, and the step B includes; the gradient result image is scanned, and the pixel establishes the index value of each level of gradient according to the order of growth of the gradient value; starting from the lowest gradient value, according to the gradient value and the neighborhood of the pixel Classification of domain labeling situations. 9. The processing method for remote sensing images used for detecting surface changes according to claim 8, characterized in that: said classification according to the gradient value and neighborhood labeling of pixels refers to: from the pixel gray value The lowest value h ₀ starts simulating water immersion, assuming that the water storage basin to which the pixels of the gray level less than or equal to h have been identified, then when processing the pixels of the gray level h+1, the pixels in this gray level The pixels adjacent to the marked water storage basin are sent into a first-in first-out (FIFO) queue, and then starting from these pixels, according to the geodesic distance, the marked water storage basin is extended to h+1 gray level, if h There are still unmarked pixels in the +1 gray level, which will be used as newly emerging local minimum regions, and given new region labels; until all pixels have been traversed. 10. The processing method for remote sensing images used for detecting surface changes according to claim 7, characterized in that: said reconstructing the original resolution image comprises the following steps: step a, performing inverse wavelet transform on image M _L , to get M _L-1 ; step b, perform Robert gradient operation on the image M _L-1 and then perform watershed transformation, set the interior of the region to 0, and the boundary to 255 to obtain a binary image B _L-1 ; step c, according to The boundary line of BL _-1 is used to refine the boundary and area of ML _-1 , namely: correspond BL _-1 and ML _-1 point by point, and map each area of BL _-1 to _ML-1 The number of points in the area; count the area mark value in M _L-1 with the most mapping points; assign the mark value to the corresponding area of B _L-1 ; step d, repeat the above three processes until L=0 .

Images