A karst collapse area identification method for remote sensing images

Abstract

The present application relates to the technical field of image data processing, and proposes a karst collapse region identification method for remote sensing images, which comprises: acquiring a suspected fog region; obtaining an artificial structure index according to the region edge of the suspected fog region and a remote sensing HSV image; obtaining a collapse feature index according to the edge lines in the geographical block around the suspected fog region and the remote sensing HSV image; obtaining an atmospheric light analysis region according to the artificial structure index and the collapse feature index; obtaining a fog concentration index according to the collapse feature index of the atmospheric light analysis region and the remote sensing HSV image; obtaining an estimated value of the global atmospheric light according to the fog concentration index of the atmospheric light analysis region; and obtaining a karst collapse image region according to the estimated value of the global atmospheric light. According to the estimated value of the global atmospheric light, the present application performs fog removal and enhancement on the remote sensing RGB image, thereby improving the accuracy of karst collapse image region segmentation.

Description

Technical Field

[0001] This invention relates to the field of image data processing technology, and specifically to a method for identifying karst collapse areas in remote sensing images. Background Technology

[0002] Karst collapse refers to a geological hazard caused by the gradual formation of cavities in underground spaces due to karst processes. When the overburden covering these cavities can no longer withstand the load from above, the surface suddenly collapses. This phenomenon occurs frequently in karst-developed areas and is often influenced by geological conditions, changes in groundwater levels, and human activities. Karst collapse not only damages the surface environment but can also threaten buildings, transportation infrastructure, and human lives, making it a geological problem that requires focused monitoring and management.

[0003] Currently, remote sensing technology is mainly used for the location and monitoring of karst collapse areas. Specifically, drones are used to acquire remote sensing images of the target area, and machine learning models are used to segment the karst collapse areas in the images to obtain various geological information about the karst collapse areas. However, during the process of acquiring remote sensing images of the target area, because karst collapse areas are mostly located in mountainous areas and usually have abundant water activity, fog often exists in the remote sensing images, affecting the extraction and analysis of image information of karst collapse areas. When using existing dehazing algorithms to process remote sensing images, since remote sensing images do not include the sky area, the atmospheric light estimation points in the dehazing algorithms often fail to reflect the actual degree of fog blurring in the karst collapse area, which may cause color distortion, contrast changes, and other details loss in the collapse area, reducing the accuracy of karst collapse area segmentation. Summary of the Invention

[0004] This invention provides a method for identifying karst collapse areas from remote sensing images, to solve the problem of inaccurate segmentation of existing karst collapse areas. The specific technical solution adopted is as follows: This invention proposes a method for identifying karst collapse areas from remote sensing images, the method comprising the following steps: Acquire remote sensing RGB images, remote sensing HSV images, remote sensing dark channel images, remote sensing grayscale images, and several geographic blocks from remote sensing RGB images; Based on the pixel values ​​of the corresponding pixels in the remote sensing dark channel image of the geographic block, several suspected fog regions are obtained; based on the region edge of the suspected fog region and the H channel value of the corresponding pixels in the remote sensing HSV image of the suspected fog region, the artificial structure index of each suspected fog region is obtained. Obtain the edge lines within the geographic blocks surrounding each suspected fog area. Based on the V channel values ​​of the pixels in the corresponding area of ​​the suspected fog area in the remote sensing HSV image and the edge lines within the geographic blocks surrounding it, obtain the collapse feature index of each suspected fog area. Based on the artificial structure index and collapse feature index of the suspected fog area, the fog analysis weight of each suspected fog area is obtained. Based on the fog analysis weight of the suspected fog area, several atmospheric light analysis areas are obtained. Based on the V channel value and S channel value of the corresponding pixel in the remote sensing HSV image of the atmospheric light analysis area, and the collapse feature index of the atmospheric light analysis area, the fog density index of each atmospheric light analysis area is obtained. Based on the fog density index, remote sensing dark channel image, remote sensing HSV image, and remote sensing grayscale image of the atmospheric light analysis area, the estimated value of global atmospheric light is obtained; based on the estimated value of global atmospheric light, the remote sensing RGB image is defogging, enhanced, and segmented to obtain the karst collapse image area.

[0005] Furthermore, the specific method for obtaining several suspected fog areas based on the pixel values ​​of pixels within the corresponding region of the geographic block in the remote sensing dark channel image includes: The optimal segmentation threshold of the remote sensing dark channel image is obtained using the Otsu thresholding method. Pixels in the remote sensing dark channel image that are larger than the optimal segmentation threshold are recorded as suspected fog pixels. Pixels at the same position in the remote sensing RGB image corresponding to the suspected fog pixels are recorded as control suspected fog pixels. In the remote sensing RGB image, the geographical blocks containing control suspected fog pixels are recorded as suspected fog blocks. The area formed by several adjacent suspected fog blocks is denoted as the suspected fog area.

[0006] Furthermore, the specific method for obtaining the artificial structure index of each suspected fog region based on the region edge of the suspected fog region and the H channel value of the corresponding pixel in the remote sensing HSV image is as follows: On the edge of any suspected fog area, any edge pixel is recorded as the target edge pixel. The direction from the target edge pixel to the next adjacent edge pixel is recorded as the extension direction of the target edge pixel, along the clockwise direction of the edge of the suspected fog area. The minimum angle between the extension direction of the target edge pixel and the extension direction of the next adjacent region edge pixel is denoted as the degree of extension change of the target edge pixel. The average value of the extension variation of all edge pixels in the suspected fog area is recorded as the edge distortion degree of the suspected fog area; The variance of the H channel values ​​of all pixels in the corresponding region of the suspected fog area in the remote sensing HSV image is recorded as the degree of color disorder in the suspected fog area; The inversely proportional normalized result of the product of the edge distortion degree and the color disorder degree of the suspected fog area is denoted as the artificial structure index of the suspected fog area.

[0007] Furthermore, the specific method for obtaining the edge lines within the geographic blocks surrounding each suspected fog area, and obtaining the collapse feature index of each suspected fog area based on the V channel values ​​of the pixels within the corresponding region in the remote sensing HSV image and the edge lines within the surrounding geographic blocks, includes: The average V channel value of all pixels in the corresponding region of any suspected fog region in the remote sensing HSV image is recorded as the regional brightness of the suspected fog region. For any suspected fog area, the region consisting of all geographical blocks adjacent to the suspected fog area is denoted as the crack analysis region of the suspected fog area; The Canny edge detection algorithm was used to perform edge detection on the crack analysis area of ​​the suspected fog area, and several edge lines were obtained. For any edge line, among all endpoints of the edge line, select the endpoint with the greatest Euclidean distance from the centroid of the suspected fog region, and denote it as the characteristic endpoint of the edge line; denote the shortest length along the edge line between any endpoint of the edge line and the characteristic endpoint as the characteristic distance of the endpoint; denote the maximum value among all characteristic distances of the edge line as the characteristic length of the edge line; denote the projection length of the edge line onto the line connecting the characteristic endpoint of the edge line and the centroid of the suspected fog region as the extension length of the edge line; denote the ratio of the extension length of the edge line to the characteristic length as the texture extension degree of the edge line. Among all the endpoints of the edge line, the endpoint that is farthest from the centroid of the suspected fog region by the Euclidean distance is marked as the endpoint of the edge line; Use the Sobel operator to obtain the gradient values ​​of all pixels on the edge line; Let any pixel on the edge line be the target pixel. Let the absolute value of the difference between the gradient value of the next pixel along the direction from the target pixel to the marked endpoint and the gradient value of the target pixel be the gradient weakening contribution of the target pixel. The average value of the gradient weakening contribution of all pixels on the edge line is denoted as the texture weakening degree of the edge line. Based on the regional brightness of the suspected fog area and the degree of texture extension and texture weakening of the edge lines within the crack analysis area, the collapse characteristic index of each suspected fog area is obtained.

[0008] Furthermore, based on the regional brightness of the suspected fog region and the degree of texture extension and texture weakening of the edge lines within the crack analysis area, the collapse feature index of each suspected fog region is obtained. The specific method for obtaining this index is as follows: In the formula, For the first Collapse characteristic index of a suspected fog area; For the first The number of edge lines within the crack analysis area of ​​a suspected foggy area; For the first The crack analysis area within the suspected fog area is the first... The extent of texture extension of each edge line; For the first The crack analysis area within the suspected fog area is the first... The degree of textural weakening of each edge line; For the first Brightness of the area suspected to be foggy; Represents the sigmoid function; This is a hyperparameter.

[0009] Furthermore, the method for obtaining the fog analysis weight for each suspected fog region based on the artificial structure index and collapse characteristic index of the suspected fog region, and obtaining several atmospheric light analysis regions based on the fog analysis weights of the suspected fog regions, includes the following specific methods: For any suspected fog area, the product of the inversely proportional normalized result of the artificial structure index of the suspected fog area and the collapse characteristic index is recorded as the fog analysis weight of the suspected fog area. The sequence of fog analysis weights for all suspected fog areas, arranged in ascending order, is denoted as the analysis weight sequence. The difference between each fog analysis weight in the analysis weight sequence and its preceding neighbor is recorded as the degree of increase of that fog analysis weight. In the analysis weight sequence, the fog analysis weight with the largest increase and all subsequent fog analysis weights corresponding to the suspected fog regions are respectively recorded as atmospheric light analysis regions.

[0010] Furthermore, based on the V channel and S channel values ​​of the pixels within the corresponding region of the atmospheric light analysis region in the remote sensing HSV image, and the collapse characteristic index of the atmospheric light analysis region, the fog density index of each atmospheric light analysis region is obtained. The specific method for obtaining this index is as follows: In the formula, For the first Haze density index of one atmospheric light analysis region; For the first Collapse characteristic index of an atmospheric light analysis region; For the first The number of pixels within an atmospheric light analysis region; For the first The first atmospheric light analysis region The V channel value of each pixel at the corresponding position in the remote sensing HSV image; For the first The first atmospheric light analysis region The S-channel value of a pixel at the corresponding position in the remote sensing HSV image; This is the weight normalization function; This is a hyperparameter.

[0011] Furthermore, the method for obtaining the estimated value of global atmospheric light based on the fog density index, remote sensing dark channel image, remote sensing HSV image, and remote sensing grayscale image of the atmospheric light analysis area includes the following specific methods: The atmospheric light analysis region with the highest fog density index is designated as the atmospheric light determining region; the dark channel value of the corresponding pixel in the remote sensing dark channel image for any pixel within the atmospheric light determining region is designated as the analysis index of that pixel; the region with the highest analysis index within the atmospheric light determining region is designated as the atmospheric light determining region. The pixels representing the proportion are denoted as atmospheric light analysis pixels; among them, This is the preset screening ratio; For any atmospheric light analysis pixel, the product of the inversely proportional normalized result of the S channel value of the corresponding pixel in the remote sensing HSV image and the V channel value is recorded as the analysis degree of the atmospheric light analysis pixel. The grayscale value of the corresponding pixel in the remote sensing grayscale image of the atmospheric light analysis pixel with the highest degree of analysis is used as the estimated value of global atmospheric light.

[0012] Furthermore, the specific method for dehazing, enhancing, and segmenting the remotely sensed RGB image based on the estimated global atmospheric light to obtain the karst collapse image region includes: Based on the estimated global atmospheric light, the remote sensing RGB image is enhanced by using a dark channel dehazing algorithm to obtain an enhanced remote sensing image. The enhanced remote sensing image is then input into a semantic segmentation neural network to obtain several karst collapse image regions.

[0013] Furthermore, the specific method for obtaining several geographic blocks in the remotely sensed RGB image is as follows: The remote sensing RGB image is segmented using a superpixel segmentation algorithm to obtain several image sub-regions in the remote sensing RGB image; each image sub-region is recorded as a geographic block.

[0014] The beneficial effects of this invention are as follows: In remote sensing dark channel images, the brighter parts are usually composed of artificial structure areas or areas of strong fog on the ground. The dark channel pixel values ​​of these artificial structure areas cannot reflect global atmospheric light parameters. This invention obtains the artificial structure index of each suspected fog area by using the area edge of the suspected fog area and the H-channel values ​​of the pixels in the corresponding area of ​​the suspected fog area in the remote sensing HSV image. This determines the probability that each suspected fog area is an artificial structure, eliminating the influence of artificial structures on the estimation of global atmospheric light. When dehazing remote sensing RGB images, the main purpose is to remove fog from karst collapse areas, thereby more accurately obtaining the relevant geological features of karst collapse areas. This invention uses the V-channel values ​​of the pixels in the corresponding area of ​​the suspected fog area in the remote sensing HSV image and the surrounding geographical blocks... By analyzing the edge lines, a collapse feature index is obtained for each suspected fog area to determine the likelihood of karst collapse in each area. When evaluating global atmospheric light in remote-sensed RGB images, it is necessary to select the part with the densest fog in the karst collapse area. This invention obtains the atmospheric light analysis area through the artificial structure index and collapse feature index of the suspected fog area, and then obtains the fog density index based on the remote-sensed HSV image and the collapse feature index of the atmospheric light analysis area to evaluate the fog density of the atmospheric light analysis area. Remote-sensed RGB images contain multiple regions. During the dehazing process of remote-sensed RGB images, since the remote-sensed RGB images do not include the sky, the fog density index of the atmospheric light analysis area is combined with the remote-sensed HSV image, remote-sensed dark channel image, and remote-sensed grayscale image to obtain an estimate of the global atmospheric light. Thus, this invention improves the accuracy of karst collapse image region segmentation by using accurate estimates of global atmospheric light to dehaze remote-sensed RGB images. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic flowchart of a method for identifying karst collapse areas from remote sensing images, provided in one embodiment of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 The diagram illustrates a flowchart of a method for identifying karst collapse areas in remote sensing images, provided by an embodiment of the present invention. The method includes the following steps: Step S001: Obtain remote sensing RGB image, remote sensing HSV image, remote sensing dark channel image, remote sensing grayscale image, and several geographic blocks in the remote sensing RGB image.

[0019] It should be noted that the purpose of this embodiment is to enhance the remote sensing image by combining the characteristics of the karst collapse area in the remote sensing image, thereby reducing the impact of fog on the segmentation of the karst collapse area. Therefore, the remote sensing image is acquired first.

[0020] Specifically, the area where karst collapse monitoring is required is designated as the target area; an unmanned aerial vehicle equipped with a high-resolution remote sensing sensor is used to collect initial remote sensing RGB images of the target area; Gaussian filtering is applied to the initial remote sensing RGB image to obtain the remote sensing RGB image; Convert the remote sensing RGB image to the HSV channel to obtain the remote sensing HSV image; Acquire the dark channel image of the remote sensing RGB image; The remote sensing RGB image is converted to grayscale to obtain a remote sensing grayscale image; the methods for Gaussian filtering and dark channel image acquisition are well-known techniques, and the specific methods will not be described here.

[0021] It should be noted that karst collapses are usually located in mountainous and hilly areas with steep slopes and abundant groundwater activity. These areas may include karst collapse areas, vegetation-covered areas, water distribution areas, and human activity areas. Different types of areas have different texture and color characteristics. In order to improve the accuracy of subsequent remote sensing RGB image analysis, the remote sensing RGB image needs to be divided into several geographic blocks.

[0022] Specifically, the superpixel segmentation algorithm is used to segment the remote sensing RGB image to obtain several image sub-regions in the remote sensing RGB image; each image sub-region is recorded as a geographic block; the superpixel segmentation algorithm is a well-known technology, and the specific method will not be described here; It should be noted that individual geographic blocks have similar texture and color characteristics, so individual geographic blocks usually represent the same type of area.

[0023] Step S002: Based on the pixel values ​​of the corresponding pixels in the remote sensing dark channel image of the geographic block, obtain several suspected fog regions; based on the region edges of the suspected fog regions and the H channel values ​​of the corresponding pixels in the remote sensing HSV image, obtain the artificial structure index of each suspected fog region.

[0024] It should be noted that remote sensing RGB images contain multiple regions. During the dehazing process of remote sensing RGB images, since remote sensing RGB images do not include the sky, it is difficult to select the region with the densest fog using traditional global atmospheric light selection methods, resulting in unsatisfactory dehazing effects.

[0025] It's important to further clarify the prior knowledge behind the dark channel dehazing algorithm: in natural RGB color images, non-sky regions typically have one channel with a very low pixel value. In practical applications, this prior knowledge can be extended to mean that only whitish objects in the image have relatively large pixel values ​​in all three channels. Remote sensing RGB images primarily consist of hills and mountains, typically excluding the sky. Therefore, in remote sensing dark channel images, brighter areas are usually composed of artificial structures on the ground or areas of strong fog, and the dark channel pixel values ​​of these artificial structures cannot reflect global atmospheric light parameters. To improve the accuracy of global atmospheric light analysis, it is necessary to evaluate the artificial structure index of different local regions in remote sensing RGB images using remote sensing dark channel images.

[0026] Specifically, the optimal segmentation threshold of the remote sensing dark channel image is obtained using the Otsu thresholding method. Pixels in the remote sensing dark channel image that are larger than the optimal segmentation threshold are recorded as suspected fog pixels. Pixels at the same position in the remote sensing RGB image corresponding to the suspected fog pixels are recorded as control suspected fog pixels. In the remote sensing RGB image, geographical blocks containing control suspected fog pixels are recorded as suspected fog blocks. The Otsu thresholding method is a well-known technique, and its specific method will not be described here.

[0027] It should be noted that the suspected foggy area initially exhibits foggy characteristics. To reduce interference from artificial structure regions during subsequent atmospheric light estimation, it is necessary to analyze whether the suspected foggy area possesses artificial structure features. In remote sensing RGB images, artificial structure regions affecting global atmospheric light analysis are typically buildings with light-colored surfaces, such as those in mining areas. In remote sensing RGB images, the areas containing these buildings usually have clear geometric boundaries compared to natural areas, are composed of artificial materials, and are typically relatively uniform in color; therefore, they are used to calculate the artificial structure index of the region.

[0028] Specifically, the area formed by several adjacent suspected fog blocks is denoted as the suspected fog area.

[0029] It should be noted that the suspected fog area may be formed by fog itself or it may be an artificial structure. Artificial structures have relatively regular geometric boundaries and a single color, so the judgment can be made based on the regularity of the extension of the boundary pixels of the suspected fog area and the uniformity of the color of the internal pixels of the suspected fog area.

[0030] Specifically, on the edge of any suspected fog area, any edge pixel on the edge of the area is recorded as the target edge pixel. Along the clockwise direction of the edge of the suspected fog area, the direction from the target edge pixel to the next adjacent edge pixel is recorded as the extension direction of the target edge pixel. The minimum angle between the extension direction of the target edge pixel and the extension direction of the next adjacent region edge pixel is denoted as the degree of extension change of the target edge pixel. The average value of the extension variation of all edge pixels in the suspected fog area is recorded as the edge distortion degree of the suspected fog area; The variance of the H channel values ​​of all pixels in the corresponding region of the suspected fog area in the remote sensing HSV image is recorded as the degree of color disorder in the suspected fog area; The inversely proportional normalized result of the product of the edge distortion degree and the color disorder degree of the suspected fog area is denoted as the artificial structure index of the suspected fog area.

[0031] Step S003: Obtain the edge lines within the geographic blocks surrounding each suspected fog area. Based on the V channel values ​​of the pixels in the corresponding area of ​​the suspected fog area in the remote sensing HSV image and the edge lines within the geographic blocks surrounding it, obtain the collapse feature index of each suspected fog area.

[0032] It should be noted that when dehazing remote sensing RGB images, the main purpose is to remove the fog from karst collapse areas, thereby more accurately obtaining the relevant geological features of karst collapse areas. The fog density in mountainous areas is related to the actual terrain and vegetation distribution, and the estimation of global atmospheric light has a significant impact on the dehazing effect. Therefore, global atmospheric light should be obtained as much as possible based on the fog characteristics of the karst collapse area and its surrounding areas. Thus, it is necessary to determine the probability of karst collapse in suspected fog areas.

[0033] It should be further noted that karst collapses mostly develop in depressions and valleys of karst mountains where groundwater changes rapidly. Karst collapse areas usually receive less sunlight. Compared with other suspected fog areas, karst collapse areas are shaded colored areas with relatively low brightness.

[0034] Specifically, the average V channel value of all pixels within the corresponding region of any suspected fog area in the remote sensing HSV image is recorded as the regional brightness of the suspected fog area.

[0035] It should be noted that during karst collapse, fracture zones typically form around the collapsed area, while areas resembling fog that are not part of the karst collapse zone do not exhibit these fracture zone characteristics. In remote sensing RGB images, fracture zones appear as textures extending outwards from the collapsed area, thus serving as a basis for determining the extent of texture extension.

[0036] Specifically, for any suspected fog area, the region consisting of all geographical blocks adjacent to the suspected fog area is denoted as the crack analysis region of the suspected fog area; The Canny edge detection algorithm was used to perform edge detection on the crack analysis area of ​​the suspected fog area, resulting in several edge lines. The Canny edge detection algorithm is a well-known technology, and the specific method will not be described here. It should be noted that the edge lines are connected edges, not just edges formed by a single line. Because they are connected, there may be multiple endpoints, that is, there are branching situations on the edge lines. In other words, several connected edge pixels together constitute the edge lines, while unconnected pixels do not participate in the formation. For any edge line, among all endpoints of the edge line, select the endpoint with the greatest Euclidean distance from the centroid of the suspected fog region, and denote it as the characteristic endpoint of the edge line; denote the shortest length along the edge line between any endpoint of the edge line and the characteristic endpoint as the characteristic distance of the endpoint; denote the maximum value among all characteristic distances of the edge line as the characteristic length of the edge line; denote the projection length of the edge line onto the line connecting the characteristic endpoint of the edge line and the centroid of the suspected fog region as the extension length of the edge line; denote the ratio of the extension length of the edge line to the characteristic length as the texture extension degree of the edge line.

[0037] It should be noted that the stress field experienced by surface fracture cracks is very complex, including vertical pressure, shear force, and additional stress caused by water infiltration or load changes. This complex stress distribution can cause multiple stress concentration points to form at the crack tip, thus inducing bifurcation and gradually weakening the original fracture crack. Therefore, the gradient of the pixels on the edge lines of the crack analysis area gradually decreases as the distance from the karst collapse area increases.

[0038] Specifically, among all endpoints of any edge line in the crack analysis region of any suspected fog region, the endpoint with the farthest Euclidean distance from the centroid of the suspected fog region is marked as the endpoint of the edge line. The gradient values ​​of all pixels on the edge line are obtained using the Sobel operator; the method of obtaining the pixel gradient values ​​using the Sobel operator is a well-known technique, and the specific method will not be described here. Let any pixel on the edge line be the target pixel. Let the absolute value of the difference between the gradient value of the next pixel along the edge line to the marked endpoint and the gradient value of the target pixel be the gradient weakening contribution of the target pixel. It should be noted that the gradient weakening contribution of the marked endpoint is recorded as 0. The average value of the gradient weakening contribution of all pixels on the edge line is denoted as the texture weakening degree of the edge line.

[0039] It should be noted that when the brightness of a suspected fog area is low, and the texture extension and texture weakening of the edge lines within the crack analysis area of ​​the suspected fog area are both high, the suspected fog area is more likely to be a karst collapse area. Therefore, the collapse characteristic index of the suspected fog area is calculated.

[0040] Specifically, the first The method for calculating the collapse characteristic index of a suspected fog area is as follows: In the formula, For the first Collapse characteristic index of a suspected fog area; For the first The number of edge lines within the crack analysis area of ​​a suspected foggy area; For the first The crack analysis area within the suspected fog area is the first... The extent of texture extension of each edge line; For the first The crack analysis area within the suspected fog area is the first... The degree of textural weakening of each edge line; For the first Brightness of the area suspected to be foggy; This represents the sigmoid function, which is used for normalization in this embodiment. To prevent hyperparameters with a denominator of 0, this embodiment uses... To narrate.

[0041] Step S004: Based on the artificial structure index and collapse feature index of the suspected fog area, obtain the fog analysis weight of each suspected fog area; based on the fog analysis weight of the suspected fog area, obtain several atmospheric light analysis areas; based on the V channel value and S channel value of the corresponding pixel in the remote sensing HSV image of the atmospheric light analysis area, and the collapse feature index of the atmospheric light analysis area, obtain the fog density index of each atmospheric light analysis area.

[0042] It should be noted that when evaluating global atmospheric light in remote sensing RGB images, it is necessary to select the part of the karst collapse area with the densest fog as much as possible. In high-resolution remote sensing images, each pixel represents a small surface area, so it is possible to capture small changes on the ground, such as small cracks around karst collapses and irregular changes on the ground. However, for foggy images, the imaging features of these subtle changes are easily blocked and blurred by fog. Therefore, it is necessary to obtain the fog density index of the area.

[0043] It should be further explained that, in order to minimize the interference of artificial structures and increase the analysis weight of the collapse area, the lower the artificial structure index and the higher the collapse characteristic index of the suspected fog area, the greater the proportion of the suspected fog area in the fog analysis process should be.

[0044] Specifically, for any suspected fog area, the product of the inversely proportional normalized result of the artificial structure index of the suspected fog area and the collapse characteristic index is recorded as the fog analysis weight of the suspected fog area. The sequence of fog analysis weights for all suspected fog areas, arranged in ascending order, is denoted as the analysis weight sequence. The difference between each fog analysis weight in the analysis weight sequence and its preceding neighbor is recorded as the growth rate of that fog analysis weight. It should be noted that the growth rate of the first fog analysis weight in the analysis weight sequence is 0. In the analysis weight sequence, the fog analysis weight with the largest increase and all subsequent fog analysis weights corresponding to the suspected fog regions are respectively recorded as atmospheric light analysis regions.

[0045] It should be noted that when surface reflected light is affected by fog, it is easy to appear "whitened" in remote sensing images. This is mainly because the presence of fog reduces the color saturation of distant objects. Because fog scatters light, colors become blurred and their vividness is reduced. In the image, the lower the pixel saturation, the denser the fog in the image. In addition, the shadowed areas inside karst collapse areas usually have low brightness. Therefore, the fog density index of each atmospheric light analysis area is obtained accordingly.

[0046] It should be further noted that the higher the collapse characteristic index of the atmospheric light analysis area, the greater the probability that the atmospheric light analysis area is a karst collapse area. When the brightness of a pixel within the atmospheric light analysis area is lower, that pixel is more likely to belong to a karst collapse area, and its weight in fog analysis is greater. Conversely, when the saturation of a pixel within the atmospheric light analysis area is lower, the fog represented by that pixel is denser.

[0047] Specifically, the first The fog density index for each atmospheric light analysis region is calculated as follows: In the formula, For the first Haze density index of one atmospheric light analysis region; For the first Collapse characteristic index of an atmospheric light analysis region; For the first The number of pixels within an atmospheric light analysis region; For the first The first atmospheric light analysis region The V channel value of each pixel at the corresponding position in the remote sensing HSV image; For the first The first atmospheric light analysis region The S-channel value of a pixel at the corresponding position in the remote sensing HSV image; Here is the weight normalization function, and the object being normalized is the weight of the first weight. All pixels within the atmospheric light analysis area ; To prevent hyperparameters with a denominator of 0, this embodiment uses... To narrate.

[0048] Step S005: Based on the fog density index of the atmospheric light analysis area, the remote sensing dark channel image, the remote sensing HSV image, and the remote sensing grayscale image, obtain the estimated value of the global atmospheric light; based on the estimated value of the global atmospheric light, perform defogging enhancement and segmentation on the remote sensing RGB image to obtain the karst collapse image area.

[0049] It should be noted that if the estimated global atmospheric light value is too high, the entire dehazed image will appear darker, and vice versa. The estimated global atmospheric light value indirectly affects the visibility of objects, as inappropriate adjustments may lead to loss of detail in excessively dark or bright environments. By combining karst collapse characteristics to evaluate fog density, fog density indices were obtained for each atmospheric light analysis region in the current image. Since the dehazing operation is mainly performed on karst collapse areas to more accurately obtain the relevant geological features of karst collapse areas, the global atmospheric light estimation point for karst collapse areas should be selected from the atmospheric light analysis region with the highest fog density index.

[0050] Specifically, the atmospheric light analysis region with the highest fog density index is designated as the atmospheric light determining region; the dark channel value of the corresponding pixel in the remote sensing dark channel image for any pixel within the atmospheric light determining region is designated as the analysis index of that pixel; the region with the highest analysis index within the atmospheric light determining region is designated as the atmospheric light determining region. The pixels representing the proportion are denoted as atmospheric light analysis pixels; among them, As a preset screening ratio, this embodiment uses To narrate; For any atmospheric light analysis pixel, the product of the inversely proportional normalized result of the S channel value of the corresponding pixel in the remote sensing HSV image and the V channel value is recorded as the analysis degree of the atmospheric light analysis pixel. The grayscale value of the corresponding pixel in the remote sensing grayscale image is used as the estimated value of global atmospheric light. The remote sensing RGB image is then enhanced by using the dark channel dehazing algorithm to obtain the enhanced remote sensing image. The estimated value of global atmospheric light is a parameter in the dark channel dehazing algorithm, which is a well-known technique and its specific method will not be described here.

[0051] It should be noted that the enhanced remote sensing image after defogging can accurately obtain the relevant geological features of the karst collapse area, so the karst collapse area in the enhanced remote sensing image is segmented.

[0052] Specifically, the remote sensing enhanced images are input into a semantic segmentation neural network to obtain several karst collapse image regions. The neural network model in this embodiment is the U-Net network; other models can be used in other embodiments, and this embodiment is not limited. Specifically, a large number of remote sensing images of karst collapse areas are obtained from a public remote sensing data platform. The remote sensing images of karst collapse areas are manually labeled at the pixel level. The labels are divided into normal areas and karst collapse areas. Pixels belonging to karst collapse areas are labeled with a value of 1, and pixels belonging to normal areas are labeled with a value of 0. The neural network is trained using a dataset composed of a large number of labeled remote sensing images of karst collapse areas. The loss function used by the network is the cross-entropy function. The specific training process is well known and will not be described in detail in this embodiment.

[0053] This embodiment adopts The model is used to represent the inverse proportional relationship and for normalization processing. It is an exponential function with the natural constant as its base. As input to the model, implementers can set inverse proportional functions and normalization functions according to the actual situation.

[0054] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying karst collapse areas from remote sensing images, characterized in that, The method includes the following steps: Acquire remote sensing RGB images, remote sensing HSV images, remote sensing dark channel images, remote sensing grayscale images, and several geographic blocks from remote sensing RGB images; Based on the pixel values ​​of the corresponding pixels in the remote sensing dark channel image of the geographic block, several suspected fog regions are obtained; based on the region edge of the suspected fog region and the H channel value of the corresponding pixels in the remote sensing HSV image of the suspected fog region, the artificial structure index of each suspected fog region is obtained. Obtain the edge lines within the geographic blocks surrounding each suspected fog area. Based on the V channel values ​​of the pixels in the corresponding area of ​​the suspected fog area in the remote sensing HSV image and the edge lines within the geographic blocks surrounding it, obtain the collapse feature index of each suspected fog area. Based on the artificial structure index and collapse feature index of the suspected fog area, the fog analysis weight of each suspected fog area is obtained. Based on the fog analysis weight of the suspected fog area, several atmospheric light analysis areas are obtained. Based on the V channel value and S channel value of the corresponding pixel in the remote sensing HSV image of the atmospheric light analysis area, and the collapse feature index of the atmospheric light analysis area, the fog density index of each atmospheric light analysis area is obtained. Based on the fog density index, remote sensing dark channel image, remote sensing HSV image, and remote sensing grayscale image of the atmospheric light analysis area, the estimated value of global atmospheric light is obtained; based on the estimated value of global atmospheric light, the remote sensing RGB image is defogging, enhanced, and segmented to obtain the karst collapse image area.

2. The method for identifying karst collapse areas from remote sensing images according to claim 1, characterized in that, The method for identifying several suspected fog areas based on the pixel values ​​of pixels within the corresponding region of a geographic block in a remote sensing dark channel image includes: The optimal segmentation threshold of the remote sensing dark channel image is obtained using the Otsu thresholding method. Pixels in the remote sensing dark channel image that are larger than the optimal segmentation threshold are recorded as suspected fog pixels. Pixels at the same position in the remote sensing RGB image corresponding to the suspected fog pixels are recorded as control suspected fog pixels. In the remote sensing RGB image, the geographical blocks containing control suspected fog pixels are recorded as suspected fog blocks. The area formed by several adjacent suspected fog blocks is denoted as the suspected fog area.

3. The method for identifying karst collapse areas from remote sensing images according to claim 1, characterized in that, The method for obtaining the artificial structure index of each suspected fog region based on the region edge of the suspected fog region and the H channel value of the corresponding pixel in the remote sensing HSV image includes the following specific methods: On the edge of any suspected fog area, any edge pixel is recorded as the target edge pixel. The direction from the target edge pixel to the next adjacent edge pixel is recorded as the extension direction of the target edge pixel, along the clockwise direction of the edge of the suspected fog area. The minimum angle between the extension direction of the target edge pixel and the extension direction of the next adjacent region edge pixel is denoted as the degree of extension change of the target edge pixel. The average value of the extension variation of all edge pixels in the suspected fog area is recorded as the edge distortion degree of the suspected fog area; The variance of the H channel values ​​of all pixels in the corresponding region of the suspected fog area in the remote sensing HSV image is recorded as the degree of color disorder in the suspected fog area; The inversely proportional normalized result of the product of the edge distortion degree and the color disorder degree of the suspected fog area is denoted as the artificial structure index of the suspected fog area.

4. The method for identifying karst collapse areas from remote sensing images according to claim 1, characterized in that, The specific method for obtaining the edge lines within the geographic blocks surrounding each suspected fog area, and obtaining the collapse feature index of each suspected fog area based on the V channel values ​​of the pixels within the corresponding region in the remote sensing HSV image and the edge lines within the surrounding geographic blocks, includes: The average V channel value of all pixels in the corresponding region of any suspected fog region in the remote sensing HSV image is recorded as the regional brightness of the suspected fog region. For any suspected fog area, the region consisting of all geographical blocks adjacent to the suspected fog area is denoted as the crack analysis region of the suspected fog area; The Canny edge detection algorithm was used to perform edge detection on the crack analysis area of ​​the suspected fog area, and several edge lines were obtained. For any edge line, among all endpoints of the edge line, select the endpoint with the greatest Euclidean distance from the centroid of the suspected fog region, and denote it as the characteristic endpoint of the edge line; denote the shortest length along the edge line between any endpoint of the edge line and the characteristic endpoint as the characteristic distance of the endpoint; denote the maximum value among all characteristic distances of the edge line as the characteristic length of the edge line; denote the projection length of the edge line onto the line connecting the characteristic endpoint of the edge line and the centroid of the suspected fog region as the extension length of the edge line; denote the ratio of the extension length of the edge line to the characteristic length as the texture extension degree of the edge line. Among all the endpoints of the edge line, the endpoint that is farthest from the centroid of the suspected fog region by the Euclidean distance is marked as the endpoint of the edge line; Use the Sobel operator to obtain the gradient values ​​of all pixels on the edge line; Let any pixel on the edge line be the target pixel. Let the absolute value of the difference between the gradient value of the next pixel along the direction from the target pixel to the marked endpoint and the gradient value of the target pixel be the gradient weakening contribution of the target pixel. The average value of the gradient weakening contribution of all pixels on the edge line is denoted as the texture weakening degree of the edge line. Based on the regional brightness of the suspected fog area and the degree of texture extension and texture weakening of the edge lines within the crack analysis area, the collapse characteristic index of each suspected fog area is obtained.

5. The method for identifying karst collapse areas from remote sensing images according to claim 4, characterized in that, The collapse characteristic index of each suspected fog region is obtained based on the regional brightness of the suspected fog region and the texture extension and texture weakening degree of the edge lines within the crack analysis region. The specific method for obtaining the index is as follows: In the formula, For the first Collapse characteristic index of a suspected fog area; For the first The number of edge lines within the crack analysis area of ​​a suspected foggy area; For the first The crack analysis area within the suspected fog area is the first... The extent of texture extension of each edge line; For the first The crack analysis area within the suspected fog area is the first... The degree of textural weakening of each edge line; For the first Brightness of the area suspected to be foggy; Represents the sigmoid function; This is a hyperparameter.

6. The method for identifying karst collapse areas from remote sensing images according to claim 1, characterized in that, The method for obtaining a fog analysis weight for each suspected fog region based on its artificial structure index and collapse characteristic index, and then obtaining several atmospheric light analysis regions based on these weights, includes the following specific methods: For any suspected fog area, the product of the inversely proportional normalized result of the artificial structure index of the suspected fog area and the collapse characteristic index is recorded as the fog analysis weight of the suspected fog area. The sequence of fog analysis weights for all suspected fog areas, arranged in ascending order, is denoted as the analysis weight sequence. The difference between each fog analysis weight in the analysis weight sequence and its preceding neighbor is recorded as the degree of increase of that fog analysis weight. In the analysis weight sequence, the fog analysis weight with the largest increase and all subsequent fog analysis weights corresponding to the suspected fog regions are respectively recorded as atmospheric light analysis regions.

7. The method for identifying karst collapse areas from remote sensing images according to claim 1, characterized in that, The fog density index for each atmospheric light analysis region is obtained based on the V channel and S channel values ​​of the pixels in the corresponding region of the atmospheric light analysis region in the remote sensing HSV image, as well as the collapse characteristic index of the atmospheric light analysis region. The specific method for obtaining this index is as follows: In the formula, For the first Haze density index of one atmospheric light analysis region; For the first Collapse characteristic index of an atmospheric light analysis region; For the first The number of pixels within an atmospheric light analysis region; For the first The first atmospheric light analysis region The V channel value of each pixel at the corresponding position in the remote sensing HSV image; For the first The first atmospheric light analysis region The S-channel value of a pixel at the corresponding position in the remote sensing HSV image; This is the weight normalization function; This is a hyperparameter.

8. The method for identifying karst collapse areas from remote sensing images according to claim 1, characterized in that, The method for obtaining the estimated value of global atmospheric light based on the fog density index, remote sensing dark channel image, remote sensing HSV image, and remote sensing grayscale image of the atmospheric light analysis area includes the following specific methods: The atmospheric light analysis region with the highest fog density index is designated as the atmospheric light determining region; the dark channel value of the corresponding pixel in the remote sensing dark channel image for any pixel within the atmospheric light determining region is designated as the analysis index of that pixel; the region with the highest analysis index within the atmospheric light determining region is designated as the atmospheric light determining region. The pixels representing the proportion are denoted as atmospheric light analysis pixels; among them, This is the preset screening ratio; For any atmospheric light analysis pixel, the product of the inversely proportional normalized result of the S channel value of the corresponding pixel in the remote sensing HSV image and the V channel value is recorded as the analysis degree of the atmospheric light analysis pixel. The grayscale value of the corresponding pixel in the remote sensing grayscale image of the atmospheric light analysis pixel with the highest degree of analysis is used as the estimated value of global atmospheric light.

9. The method for identifying karst collapse areas from remote sensing images according to claim 1, characterized in that, The specific method for dehazing, enhancing, and segmenting remotely sensed RGB images based on estimated global atmospheric light values ​​to obtain karst collapse image regions includes: Based on the estimated global atmospheric light, the remote sensing RGB image is enhanced by using a dark channel dehazing algorithm to obtain an enhanced remote sensing image. The enhanced remote sensing image is then input into a semantic segmentation neural network to obtain several karst collapse image regions.

10. A method for identifying karst collapse areas from remote sensing images according to claim 1, characterized in that, The specific method for obtaining several geographic blocks in the remote sensing RGB image is as follows: The remote sensing RGB image is segmented using a superpixel segmentation algorithm to obtain several image sub-regions in the remote sensing RGB image; each image sub-region is recorded as a geographic block.

Images