A remote sensing image segmentation method combining quadtree structure and scale estimation

Abstract

The present invention discloses a remote sensing image segmentation method that combines a quadtree structure with scale estimation. The method specifically includes remote sensing image preprocessing, quadtree partitioning, spatial segmentation scale estimation, attribute segmentation scale estimation, fractal network evolution algorithm segmentation, merging segmentation results, and accuracy verification. To address the image over-segmentation, under-segmentation, and mis-segmentation phenomena caused by the scale characteristics of different land types, the present invention uses the FNEA algorithm as a basis, a quadtree structure, and a scale estimation method based on spectral spatial statistics to achieve remote sensing image segmentation that includes multiple land feature features. This effectively improves segmentation accuracy and avoids the problems of linear land feature extraction interruption and classification errors caused by mis-segmentation and under-segmentation. This method is of great significance for remote sensing image classification and information extraction.

Description

Technical Field

[0001] The present invention belongs to the field of remote sensing image segmentation, and in particular relates to a remote sensing image segmentation method combining a quadtree structure and scale estimation. Background Art

[0002] In recent years, my country has seen rapid development in high-resolution Earth observation technology, with the launch of numerous remote sensing satellites capable of acquiring sub-meter and meter-level imagery. These high-resolution images, characterized by distinct spectral characteristics and rich texture structures, enable clear identification of ground features, playing a vital role in areas such as natural resource management and disaster monitoring. Due to the large number of pixels and high spatial resolution in high-resolution remote sensing images, traditional pixel-based classification methods struggle to ensure image integrity. Consequently, object-based image analysis (OBIA) emerged as a leading approach to high-resolution remote sensing image classification.

[0003] Image segmentation is a crucial step in OBIA and determines image classification accuracy. Currently, most OBIAs employ the Fractal Net Evolution Approach (FNEA) for multi-scale image segmentation. In multi-scale segmentation, an inappropriate segmentation scale can lead to over-segmentation or under-segmentation, reducing the accuracy of high-resolution remote sensing image classification. Therefore, determining the optimal segmentation scale is a core aspect of multi-scale segmentation.

[0004] Currently, segmentation scale parameters are mostly determined using supervised evaluation and unsupervised evaluation. Supervised evaluation involves comparing segmentation results with reference data, evaluating the segmentation effect using a similarity metric, and determining the optimal segmentation scale based on the optimal similarity metric. For example, Liu et al. proposed three similarity metrics: potential segmentation error, segmentation ratio, and Euclidean distance, to determine the optimal segmentation scale, based on the overlap between reference data and segmentation results, as well as over-segmentation and under-segmentation. Unsupervised evaluation methods determine the optimal segmentation scale based on the homogeneity and heterogeneity of the segmented patches without external data. The local variance and its rate of change map, proposed by Woodcock and Strahler in 1987, has been widely used for estimating segmentation scales. Ming et al., using the mean-shift algorithm as an example, employed the average local variance to achieve optimal estimation of spatial segmentation parameters without considering the anisotropy of the ground objects. Building on this approach, Ming summarized scale parameters into spatial segmentation scale parameters, attribute segmentation scale parameters, and merging threshold parameters, and proposed a method for determining the optimal segmentation scale using spectral spatial statistics. This method, based on the optimal estimation of spatial segmentation parameters, determines the attribute segmentation scale parameter by counting the image spectral values ​​corresponding to the first peak of the local variance histogram. Ma Yanni analyzed the relationship between the FNEA algorithm and the spectral spatial characteristics of imagery and applied a high-resolution image segmentation scale estimation method based on spectral spatial statistics to the FNEA algorithm. Experiments were conducted in farmland and building experimental areas to validate the scale estimation method.

[0005] Overall, supervised evaluation methods require extensive experimentation to determine optimal segmentation parameters, resulting in limited generalizability. While unsupervised evaluation methods can quantitatively estimate the segmentation scale before segmentation, their applicability to images composed of a large number of diverse ground objects remains to be verified experimentally. Different ground object characteristics require different segmentation scales. Too small a segmentation scale will result in oversegmentation. Too large a segmentation scale will result in undersegmentation, resulting in the appearance of mixed objects. Numerous researchers have proposed solutions to this problem. For example, Yang Haiping proposed a high-resolution image segmentation algorithm based on multi-layer optimal scales. Myint, to improve object-oriented classification accuracy, incorporated land cover information, selected segmentation scales for different land classes, and ultimately merged these into a multi-scale segmentation result. Li Qin et al., using an improved local variance method, established a local variance change map that varies with segmentation scale. They then selected appropriate segmentation scales based on the differences in ground object characteristics, establishing multi-layer ground object feature segmentation rules to achieve high-resolution remote sensing image segmentation. Addressing the difficulty of achieving both local and global optimality in segmentation results, Hong Liang et al. proposed a new multi-scale segmentation optimization algorithm for high-resolution remote sensing images that balances both global and local optima.

[0006] In summary, previous studies have used multiple segmentation experiments to estimate the scale characteristics of different land types. Therefore, based on the FNEA algorithm, this paper proposes a remote sensing image segmentation method that combines a quadtree structure with scale estimation to achieve adaptive segmentation of remote sensing images of different land types. Summary of the Invention

[0007] In view of the above technical problems, the purpose of the present invention is to overcome the shortcomings of the existing technology and propose a remote sensing image segmentation method that combines quadtree structure and scale estimation. First, the remote sensing image is quadtree segmented based on the quadtree structure to realize image block partitioning; then, the scale estimation method of spectral space statistics is used to estimate the secondary segmentation scale of each region; finally, the fractal network evolution algorithm (FNEA) is used to realize the adaptive multi-scale segmentation of remote sensing images. Specifically, it includes remote sensing image preprocessing, quadtree partitioning, spatial segmentation scale estimation, attribute segmentation scale estimation, fractal network evolution algorithm segmentation, merging segmentation results, and accuracy verification. The advantage of this method is that it does not require segmentation to specify segmentation parameters and does not require completing multiple segmentation experiments at different segmentation scales to achieve adaptive segmentation of remote sensing images.

[0008] To achieve the above functions, the present invention provides a remote sensing image segmentation method combining a quadtree structure with scale estimation, comprising the following steps:

[0009] S1: Remote sensing image preprocessing: mainly includes radiometric calibration, atmospheric correction, orthorectification, image registration, and image cropping;

[0010] S2: Quadtree partitioning: The quadtree structure is a tree-like data structure. The quadtree segmentation algorithm is an adaptive segmentation algorithm based on the quadtree structure to perform uniform detection and segmentation on the image. This paper uses the quadtree algorithm to partition the image and demarcate the area for subsequent adaptive segmentation.

[0011] S3: Spatial segmentation scale estimation: Spatial segmentation scale estimation is to use the spatial scale estimation method to estimate the spatial segmentation parameter d in remote sensing image segmentation. s The estimation process mainly includes building a window, calculating the local variance of the window, calculating the average local variance of the image, forming a local variance curve, and determining the segmentation scale according to the curve. This step is mainly for each segmentation area in step S2, and the spatial segmentation parameter d of each segmentation area is calculated. s Make estimates;

[0012] S4: Attribute segmentation scale estimation: Attribute segmentation scale estimation is the process of estimating the attribute segmentation parameters in remote sensing image segmentation using the attribute scale estimation method. This process mainly includes establishing a 2D s +1 scale estimation window, calculate the local variance of the window, statistically calculate the histogram of the local variance map, and determine the attribute segmentation scale according to the histogram. This step mainly estimates the attribute segmentation parameters of each segmentation area in step S2;

[0013] S5: Fractal Network Evolution Algorithm Segmentation: The attribute segmentation scale obtained in step S4 is the segmentation scale in the FNEA algorithm. The fractal network evolution algorithm is used to perform secondary segmentation on the remote sensing images of each region obtained by the quadtree structure segmentation in S2 at the segmentation scale determined in S4.

[0014] S6: Merge segmentation results: merge all the region segmentation results in step S5 to complete the image segmentation;

[0015] S7: Accuracy verification: Randomly generate a certain number of accuracy verification points in the segmentation area, obtain the image segmentation block attributes of the accuracy verification points according to the segmentation block, visually determine whether the accuracy judgment point belongs to the current segmentation block or the adjacent segmentation block, and calculate the confusion matrix and overall segmentation accuracy based on the accuracy verification points to evaluate the image segmentation results.

[0016] Furthermore, step S1 remote sensing image preprocessing mainly includes the following steps:

[0017] (1) Radiometric calibration is the process of converting the brightness grayscale value of an image into absolute radiometric brightness, which is used to calculate the spectral reflectance or spectral radiometric brightness of the ground object. This process is called radiometric calibration.

[0018] (2) Atmospheric correction is the process of using an atmospheric correction model to eliminate the radiation error caused by atmospheric influence and invert the true surface reflectivity of the ground object;

[0019] (3) Orthorectification refers to the use of control data and digital elevation models (DEMs) to correct the image for tilt and projection errors and resample the image into an orthophoto;

[0020] (4) Image registration refers to the process of specifying the spatial position and coordinate system of remote sensing images using ground control points or other control data;

[0021] (5) Image cropping is the process of cropping remote sensing images using the research scope to extract the image of the research area from the remote sensing image.

[0022] Furthermore, step S2 of quadtree partitioning mainly includes the following steps:

[0023] (1) Specifying a predetermined threshold for the quadtree structure;

[0024] (2) The image is divided into four sub-blocks. The image is represented as a tree node of a quadtree. The four sub-blocks are leaf nodes of the quadtree. The variance of the sub-block represented by each leaf node is calculated.

[0025] (3) Determine the relationship between the variance value and the predetermined threshold;

[0026] (4) If the variance value is less than the predetermined threshold, the mean of all grayscale values ​​of the sub-block is calculated as the grayscale value of the sub-block, and the block is segmented;

[0027] (5) If the variance is greater than the predetermined threshold, the sub-block is divided into four more sub-blocks. This process is called pruning. Continue to determine whether the variance of the four sub-blocks is less than the predetermined threshold. If not, continue pruning until the variance of the sub-blocks is less than the set threshold.

[0028] Furthermore, step S3 of spatial segmentation scale estimation mainly includes the following steps:

[0029] (1) Determine a 3×3 window and calculate the local variance value within the window;

[0030] The local variance calculation formula is as follows:

[0031]

[0032] Where n is the window size, i and j are the row and column numbers of the local window, f(i, j) is the gray value of the i-th row and j-th column, f is the average grayscale value of the window, LV(i, j) is the local variance value corresponding to the center pixel of the window, and reflects the discrete degree of the grayscale value of the window pixel;

[0033] (2) Calculate the local variance value in each window of the entire image pixel by pixel and obtain the average local variance;

[0034] The formula for calculating the average local variance is:

[0035]

[0036] Where m is the size of the image in row direction, n is the size of the image in column direction, and LV is the average local variance of a spectral component of the image.

[0037] (3) Gradually increase the calculation window and repeat steps (1) and (2) to calculate the average local variance of each window to form the average local variance curve ALV (Average Local Variogram);

[0038] (4) When the change rate threshold of the average local variance is less than 0.01 for the first time, stop increasing the window size, and the window size is the optimal spatial segmentation scale d s .

[0039] Furthermore, step S4 of attribute segmentation scale estimation mainly includes the following steps:

[0040] (1) Establish a 2ds+1 size window to calculate the local variance pixel by pixel and form a local variance map;

[0041] (2) Statistical histogram of local variance map;

[0042] (3) The square root of the local variance corresponding to the first peak of the local variance histogram is the optimal attribute segmentation scale, that is, the segmentation scale parameter.

[0043] Furthermore, step S5 of the fractal network evolution algorithm mainly includes the following steps:

[0044] (1) Merge adjacent pixels into homogeneous objects based on spectral information;

[0045] (2) Calculate the object heterogeneity considering shape and spectral characteristics, compare the scale segmentation parameters in step S4 and gradually iteratively merge the objects to complete the image segmentation;

[0046] The heterogeneity criterion of the object is determined by the spectral heterogeneity and shape heterogeneity of the object, and the calculation formula is:

[0047] sp=w color ·Δh color +w shape ·Δh shape (3)

[0048] Where: Δh color and Δh shape Respectively represent the heterogeneity of color and shape during image segmentation, w color With w shape They represent the weights of “color” and “shape” respectively during image segmentation, where w color +w shape =1;

[0049] (3) If the heterogeneity of the color and shape of the two objects is less than the set threshold, they can be merged; if the size is greater than the set threshold, they cannot be merged. This feature loop is repeated, and the decision of whether to merge is made based on the heterogeneity until it is no longer possible to merge.

[0050] Furthermore, step S6 of merging the segmentation results mainly includes the following steps:

[0051] (1) Determine a certain area in the partition S2 as the reference area;

[0052] (2) Copy the segmentation results of other areas except the benchmark area to the benchmark area to complete the merging of image segmentation results.

[0053] Furthermore, step S7 accuracy verification mainly includes the following steps:

[0054] (1) Randomly generate a certain number of accuracy verification points within the segmentation area, and obtain the image segmentation block attributes of the accuracy verification points according to their spatial positions;

[0055] (2) Visually judge the accuracy of the judgment point to determine whether it belongs to the current segmentation block or the adjacent segmentation block. The confusion matrix and the overall segmentation accuracy are calculated based on the accuracy verification points to evaluate the image segmentation results.

[0056] Beneficial effects: Compared with existing technologies, a remote sensing image segmentation method combining quadtree structure and scale estimation has the following advantages:

[0057] (1) The present invention first performs quadtree segmentation on remote sensing images based on a quadtree structure to achieve image block partitioning. Afterwards, the scale estimation method of spectral space statistics is used to estimate the secondary segmentation scale of each region. Finally, the FNEA algorithm is used to segment the remote sensing images of each region under the segmentation scale threshold of each region, and the segmentation results are merged to complete the image segmentation. The present invention can segment remote sensing images at an adaptive scale, effectively avoiding the problems of linear feature extraction interruption and classification errors caused by mis-segmentation and under-segmentation.

[0058] (2) The present invention only needs to specify a predetermined threshold value when partitioning the quadtree blocks, so as to adaptively segment the remote sensing image and realize the segmentation of images of different scales for different land types. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] The description of the contents of the present invention will become more apparent and easier to understand when taken in conjunction with the following drawings, in which:

[0060] Figure 1 This is a flow chart of a remote sensing image segmentation method combining a quadtree structure and scale estimation according to the present invention;

[0061] Figure 2 Schematic diagram of a quadtree segmentation algorithm in a specific embodiment of the present invention;

[0062] Figure 3 This is a schematic diagram of the image segmentation process of the fractal network evolution algorithm in a specific embodiment of the present invention. DETAILED DESCRIPTION

[0063] according to Figure 1 The steps shown are used to describe in detail a remote sensing image segmentation method combining a quadtree structure and scale estimation according to the present invention.

[0064] Step 1: Remote sensing image preprocessing. This includes the following specific steps:

[0065] (1) Radiometric calibration is the process of converting the brightness grayscale value of an image into absolute radiometric brightness, which is used to calculate the spectral reflectance or spectral radiometric brightness of the ground object. This process is called radiometric calibration.

[0066] (2) Atmospheric correction is the process of using an atmospheric correction model to eliminate the radiation error caused by atmospheric influence and invert the true surface reflectivity of the ground object;

[0067] (3) Orthorectification refers to the use of control data and digital elevation models (DEMs) to correct the image for tilt and projection errors and resample the image into an orthophoto;

[0068] (4) Image registration refers to the process of specifying the spatial position and coordinate system of remote sensing images using ground control points or other control data;

[0069] (5) Image cropping is the process of cropping remote sensing images using the research scope to extract the image of the research area from the remote sensing image.

[0070] Step 2: Quadtree partitioning. This includes the following specific steps:

[0071] (1) Specifying a predetermined threshold for the quadtree structure;

[0072] (2) The image is divided into four sub-blocks. The image is represented as a tree node of a quadtree. The four sub-blocks are leaf nodes of the quadtree. The variance of the sub-block represented by each leaf node is calculated.

[0073] (3) Determine the relationship between the variance value and the predetermined threshold;

[0074] (4) If the variance value is less than the predetermined threshold, the mean of all grayscale values ​​of the sub-block is calculated as the grayscale value of the sub-block, and the block is segmented;

[0075] (5) If the variance is greater than the predetermined threshold, the sub-block is divided into four more sub-blocks. This process is called pruning. Continue to determine whether the variance of the four sub-blocks is less than the predetermined threshold. If not, continue pruning until the variance of the sub-blocks is less than the set threshold.

[0076] Step 3: Estimation of spatial segmentation scale. This includes the following specific steps:

[0077] (1) Determine a 3×3 window and calculate the local variance value within the window;

[0078] The local variance calculation formula is as follows:

[0079]

[0080] Where n is the window size, i and j are the row and column numbers of the local window, f(i, j) is the gray value of the i-th row and j-th column, f is the average grayscale value of the window, LV(i, j) is the local variance value corresponding to the center pixel of the window, and reflects the discrete degree of the grayscale value of the window pixel;

[0081] (2) Calculate the local variance value in each window of the entire image pixel by pixel and obtain the average local variance;

[0082] The formula for calculating the average local variance is:

[0083]

[0084] Where m is the size of the image in row direction, n is the size of the image in column direction, and LV is the average local variance of a spectral component of the image.

[0085] (3) Gradually increase the calculation window and repeat steps (1) and (2) to calculate the average local variance of each window to form the average local variance curve ALV (Average Local Variogram);

[0086] (4) When the change rate threshold of the average local variance is less than 0.01 for the first time, stop increasing the window size, and the window size is the optimal spatial segmentation scale d s .

[0087] Step 4: Attribute segmentation scale estimation. This includes the following specific steps:

[0088] (1) Establish a 2ds+1 size window to calculate the local variance pixel by pixel and form a local variance map;

[0089] (2) Statistical histogram of local variance map;

[0090] (3) The square root of the local variance corresponding to the first peak of the local variance histogram is the optimal attribute segmentation scale, that is, the segmentation scale parameter.

[0091] Step 5: Fractal network evolution algorithm. This includes the following specific steps:

[0092] (1) Merge adjacent pixels into homogeneous objects based on spectral information;

[0093] (2) Calculate the object heterogeneity considering shape and spectral characteristics, compare the scale segmentation parameters in step S4 and gradually iteratively merge the objects to complete the image segmentation;

[0094] The heterogeneity criterion of the object is determined by the spectral heterogeneity and shape heterogeneity of the object, and the calculation formula is:

[0095] sp=w color ·Δh color+w shape ·Δh shape (6)

[0096] Where: Δh color and Δh shape Respectively represent the heterogeneity of color and shape during image segmentation, w color With w shape They represent the weights of “color” and “shape” respectively during image segmentation, where w color +w shape =1;

[0097] (3) If the heterogeneity of the color and shape of the two objects is less than the set threshold, they can be merged; if the size is greater than the set threshold, they cannot be merged. This feature loop is repeated, and the decision of whether to merge is made based on the heterogeneity until it is no longer possible to merge.

[0098] Step 6: Merge the segmentation results. This includes the following specific steps:

[0099] (1) Determine a certain area in the partition S2 as the reference area;

[0100] (2) Copy the segmentation results of other areas except the benchmark area to the benchmark area to complete the merging of image segmentation results.

[0101] Step 7: Accuracy verification. This includes the following specific steps:

[0102] (1) Randomly generate a certain number of accuracy verification points within the segmentation area, and obtain the image segmentation block attributes of the accuracy verification points according to their spatial positions;

[0103] (2) Visually judge the accuracy of the judgment point to determine whether it belongs to the current segmentation block or the adjacent segmentation block. The confusion matrix and the overall segmentation accuracy are calculated based on the accuracy verification points to evaluate the image segmentation results.

[0104] The present invention proposes a remote sensing image segmentation method combining a quadtree structure and scale estimation. Aiming at the problem of over-segmentation and under-segmentation of images caused by the scale characteristics of different land types, a remote sensing image segmentation method combining a quadtree structure and scale estimation is proposed to achieve adaptive segmentation of remote sensing images of different land types.

[0105] The above description is only the best embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A remote sensing image segmentation method combining a quadtree structure and scale estimation, characterized in that: The steps include: S1: Remote sensing image preprocessing: mainly includes radiometric calibration, atmospheric correction, orthorectification, image registration, and image cropping; S2: Quadtree partitioning: The quadtree structure is a tree-like data structure. The quadtree segmentation algorithm is an adaptive segmentation algorithm based on the quadtree structure to perform uniform detection and segmentation on the image. This paper uses the quadtree algorithm to partition the image and demarcate the area for subsequent adaptive segmentation. S3: Spatial segmentation scale estimation: Spatial segmentation scale estimation is to use the spatial scale estimation method to estimate the spatial segmentation parameter d in remote sensing image segmentation. s The estimation process mainly includes building a window, calculating the local variance of the window, calculating the average local variance of the image, forming a local variance curve, and determining the segmentation scale according to the curve. This step is mainly for each segmentation area in step S2, and the spatial segmentation parameter d of each segmentation area is calculated. s Make estimates; S4: Attribute segmentation scale estimation: Attribute segmentation scale estimation is the process of estimating the attribute segmentation parameters in remote sensing image segmentation using the attribute scale estimation method. This process mainly includes establishing a 2D s +1 scale estimation window, calculate the local variance of the window, statistically calculate the histogram of the local variance map, and determine the attribute segmentation scale according to the histogram. This step mainly estimates the attribute segmentation parameters of each segmentation area in step S2; S5: Fractal Network Evolution Algorithm Segmentation: The attribute segmentation scale obtained in step S4 is the segmentation scale in the FNEA algorithm. The fractal network evolution algorithm is used to perform secondary segmentation on the remote sensing images of each region obtained by the quadtree structure segmentation in S2 at the segmentation scale determined in S4. S6: Merge segmentation results: merge all the region segmentation results in step S5 to complete the image segmentation; S7: Accuracy verification: Randomly generate a certain number of accuracy verification points in the segmentation area, obtain the image segmentation block attributes of the accuracy verification points according to the segmentation block, visually determine whether the accuracy judgment point belongs to the current segmentation block or the adjacent segmentation block, and calculate the confusion matrix and overall segmentation accuracy based on the accuracy verification points to evaluate the image segmentation results.

2. The remote sensing image segmentation method combining quadtree structure and scale estimation according to claim 1, characterized in that: The step S1 comprises the following steps: (1) Radiometric calibration is the process of converting the brightness grayscale value of an image into absolute radiometric brightness, which is used to calculate the spectral reflectance or spectral radiometric brightness of the ground object. This process is called radiometric calibration. (2) Atmospheric correction is the process of using an atmospheric correction model to eliminate the radiation error caused by atmospheric influence and invert the true surface reflectivity of the ground object; (3) Orthorectification refers to the use of control data and digital elevation models (DEMs) to correct the image for tilt and projection errors and resample the image into an orthophoto; (4) Image registration refers to the process of specifying the spatial position and coordinate system of remote sensing images using ground control points or other control data; (5) Image cropping is the process of cropping remote sensing images using the research scope to extract the image of the research area from the remote sensing image.

3. The remote sensing image segmentation method combining quadtree structure and scale estimation according to claim 1, characterized in that: The S2 comprises the following steps: (1) Specifying a predetermined threshold for the quadtree structure; (2) The image is divided into four sub-blocks. The image is represented as a tree node of a quadtree. The four sub-blocks are leaf nodes of the quadtree. The variance of the sub-block represented by each leaf node is calculated. (3) Determine the relationship between the variance value and the predetermined threshold; (4) If the variance value is less than the predetermined threshold, the mean of all grayscale values ​​of the sub-block is calculated as the grayscale value of the sub-block, and the block is segmented; (5) If the variance value is greater than the predetermined threshold, the sub-block is divided into 4 sub-blocks. This process is called pruning. Continue to determine whether the variance of the 4 small sub-blocks is less than the predetermined threshold. If not, continue pruning until the variance of the sub-block is less than the set threshold.

4. The remote sensing image segmentation method combining quadtree structure and scale estimation according to claim 1, characterized in that: The S3 includes the following steps: (1) Determine a 3×3 window and calculate the local variance value within the window; The local variance calculation formula is as follows: Where n is the window size, i and j are the row and column numbers of the local window, f(i, j) is the gray value of the i-th row and j-th column, f is the average grayscale value of the window, LV(i, j) is the local variance value corresponding to the center pixel of the window, and reflects the discrete degree of the grayscale value of the window pixel; (2) Calculate the local variance value in each window of the entire image pixel by pixel and obtain the average local variance; The formula for calculating the average local variance is: Where m is the size of the image in row direction, n is the size of the image in column direction, and LV is the average local variance of a spectral component of the image; (3) Gradually increase the calculation window and repeat steps (1) and (2) to calculate the average local variance of each window to form the average local variance curve ALV (Average Local Variogram); (4) When the change rate threshold of the average local variance is less than 0.01 for the first time, stop increasing the window size, and the window size is the optimal spatial segmentation scale d s .

5. The remote sensing image segmentation method combining quadtree structure and scale estimation according to claim 1, characterized in that: The S4 comprises the following steps: (1) Establish a 2ds+1 size window to calculate the local variance pixel by pixel and form a local variance map; (2) Statistical histogram of local variance map; (3) The square root of the local variance corresponding to the first peak of the local variance histogram is the optimal attribute segmentation scale, that is, the segmentation scale parameter.

6. The remote sensing image segmentation method combining quadtree structure and scale estimation according to claim 1, characterized in that: The S5 comprises the following steps: (1) Merge adjacent pixels into homogeneous objects based on spectral information; (2) Calculate the object heterogeneity considering shape and spectral characteristics, compare the scale segmentation parameters in step S4 and gradually iteratively merge the objects to complete the image segmentation; The heterogeneity criterion of the object is determined by the spectral heterogeneity and shape heterogeneity of the object, and the calculation formula is: sp=w color ·Δh color +w shape ·Δh shape (3) Where: Δh color and Δh shape Respectively represent the heterogeneity of color and shape during image segmentation, w color With w shape They represent the weights of "color" and "shape" respectively during image segmentation, where w color +w shape =1; (3) If the heterogeneity of the color and shape of the two objects is less than the set threshold, they can be merged; if the size is greater than the set threshold, they cannot be merged. This feature loop is repeated, and the decision of whether to merge is made based on the heterogeneity until it is no longer possible to merge.

7. The remote sensing image segmentation method combining quadtree structure and scale estimation according to claim 1, characterized in that: The S6 comprises the following steps: (1) Determine a certain area in the partition S2 as the reference area; (2) Copy the segmentation results of other areas except the benchmark area to the benchmark area to complete the merging of image segmentation results.

8. The remote sensing image segmentation method combining quadtree structure and scale estimation according to claim 1, characterized in that: The S7 comprises the following steps: (1) Randomly generate a certain number of accuracy verification points within the segmentation area, and obtain the image segmentation block attributes of the accuracy verification points according to their spatial positions; (2) Visually judge the accuracy of the judgment point to determine whether it belongs to the current segmentation block or the adjacent segmentation block. The confusion matrix and the overall segmentation accuracy are calculated based on the accuracy verification points to evaluate the image segmentation results.

Images