Method, system, storage medium and device for extracting fuzzy boundary objects from remote sensing images

Abstract

The present invention provides a method for extracting fuzzy boundary objects in remote sensing images, comprising the following steps: acquiring remote sensing images; performing classification and extraction using a supervised classification algorithm to obtain object data; performing segmentation using a mean shift algorithm to obtain vector surface patches; fusing the object data with the vector surface patches to obtain extracted fuzzy boundary object data of remote sensing images. The present invention also relates to a system, storage medium and device for extracting fuzzy boundary objects in remote sensing images. The present invention obtains extracted fuzzy boundary object data of remote sensing images by fusing the classification result with the segmentation result; in this method, the land class boundaries obtained by using the automatic segmentation capability of the mean shift algorithm are more refined and the morphology is more in line with the actual situation, so that the extraction of fuzzy boundary objects in remote sensing images is more accurate and more in line with reality.

Description

Technical Field

[0001] The invention relates to the technical field of remote sensing image processing, and in particular to a method for extracting fuzzy boundary objects from remote sensing images. Background Art

[0002] The objects in remote sensing images are complex and diverse. The appearance of the same object on the map may be very different, and the appearance of different objects on the map may be relatively similar, which brings great challenges to the interpretation of remote sensing images. In remote sensing images, some objects have regular shapes, clear boundaries, and are easy to segment, such as buildings and roads, while some growth objects such as woodlands and grasslands, or objects that are greatly affected by growth objects such as sandy land, bare land, saline-alkali land, etc., will undergo significant changes in both morphology and sparseness over time, seasons, and other reasons. The morphology of the former is mainly determined by the growth environment, weather factors, etc., while the morphology of the latter is mainly determined by the growth of the former, such as grassland degenerating into sandy land, and bare land losing its nutrients and becoming saline-alkali land. It is worth noting that this type of landform, such as grassland, has obvious differences in its interior due to different growth conditions, which will form relatively independent geographical patches. Similarly, after the grassland degrades, the adjacent sandy land will also be a relatively independent geographical patch. These geographical patches are usually staggered, and there are gradual transitions and blurred boundaries between different categories or even the same categories. This is the weak boundary of remote sensing images. The extraction of these weak boundary features becomes relatively difficult due to blurred boundaries.

[0003] Common boundary extraction methods such as Sobel, Prewitt, and Robert operators mainly obtain boundaries through difference, which have good extraction effects for strong boundary objects (clear boundaries), but poor extraction effects for gradients within objects or between objects or blurred boundaries. In addition, although the canny operator extracts more detailed boundaries than the former, from the boundary extraction results, most of the image patches are not closed, and are not beautiful because they are too fragmented, which is not in line with the actual situation. It is certain that even the current powerful deep learning methods cannot directly obtain segmentation results with boundaries that are more consistent with the morphology of fuzzy boundary objects, because the key to deep learning algorithms lies in the quality and data of samples, but for fuzzy boundary objects, it is impossible to accurately draw the boundaries by the naked eye, and there will always be a certain deviation. Summary of the invention

[0004] In order to overcome the shortcomings of the prior art, the present invention provides a method for extracting fuzzy boundary objects from remote sensing images. The present invention combines the mean shift algorithm for morphological segmentation with a deep learning classification method to achieve a fuzzy boundary extraction method that is morphologically sensitive, i.e., the boundary fits the geographic map.

[0005] The present invention provides a method for extracting fuzzy boundary objects in remote sensing images, comprising the following steps:

[0006] Acquire high-resolution remote sensing images, where the objects to be extracted have fuzzy boundaries;

[0007] Use supervised classification algorithm to classify and extract the target area of ​​remote sensing images, and obtain the object data of the objects to be extracted in the remote sensing images;

[0008] The target area of ​​the remote sensing image is segmented by using a mean shift algorithm to obtain mutually independent vector surface patches;

[0009] The object data is fused with the vector surface patch. When the overlap between the object data and the vector surface patch is greater than or equal to a set threshold range, the object type of the object data is assigned to the vector surface patch to obtain the fuzzy boundary object data of the remote sensing image.

[0010] Preferably, the method of classifying and extracting the target area of ​​the remote sensing image using a supervised classification algorithm includes:

[0011] The deep learning classification algorithm is used to classify and extract the target area of ​​the remote sensing image, including the following steps:

[0012] Segmenting and cropping the remote sensing image to obtain a surface vector file containing one or more fuzzy boundary objects, and drawing object labels in the surface vector file, converting vector data of several surface vector samples labeled as objects into raster data to obtain a rasterized object sample set;

[0013] Adjusting the UNET network model parameters, performing UNET-based model training on the ground object sample set, and obtaining a ground object model;

[0014] Input the remote sensing image data to be tested into the object model, and calculate the inspection evaluation function of this test. If the evaluation function value meets the standard, jump to the next step; if the evaluation function value does not meet the standard, adjust the UNET network model parameters and return to the step of sample preparation and modify the sample set to re-iterate training;

[0015] The ground object model evaluated as meeting the standard in the step accuracy evaluation is used to predict the ground objects in the target area of ​​the remote sensing image to obtain the rasterized ground object data in the remote sensing image.

[0016] Preferably, the ground feature data and the vector surface patches are subjected to patch removal and small hole filling processing.

[0017] Preferably, after the step of segmenting the target area of ​​the remote sensing image using the mean shift algorithm to obtain mutually independent vector surface spots, the method further includes:

[0018] The obtained vector surface pattern is simplified and smoothed.

[0019] Preferably, the step of segmenting the remote sensing image target area using a mean shift algorithm to obtain independent spot boundaries comprises the following steps:

[0020] Select the center point and randomly select a point as the center point in the unclassified remote sensing image data;

[0021] Obtain a set, obtain all points within the bandwidth from the center point, record them as set M, and the set M forms a cluster;

[0022] Calculate the vector, calculate the vector from the center point to each element in the set M, add these vectors together to obtain the offset vector;

[0023] Drift movement, the center point moves along the drift direction, and the distance of movement is the modulus of the offset vector;

[0024] Obtain a new center point, repeat the steps of obtaining a set, calculating a vector, and drifting and moving until the size of the offset vector is within a set threshold range, and obtain a new center point at this time;

[0025] Form clusters, repeat the steps of obtaining sets, calculating vectors, and drifting and moving until all points in the remote sensing image data are clustered;

[0026] Determine the cluster to which the current point belongs, and according to the access frequency of each cluster to each point, take the cluster with the largest access frequency as the cluster to which the current point belongs, that is, the independent patch boundary.

[0027] Preferably, the step of segmenting the remote sensing image target area using a mean shift algorithm to obtain mutually independent vector surface patches includes:

[0028] The target area of ​​the remote sensing image is segmented by using a mean shift algorithm to obtain independent image spot boundaries; the independent image spot boundaries are vectorized and faceted to obtain the mutually independent vector surface spots.

[0029] An electronic device, characterized by comprising: a processor;

[0030] A memory; and a program, wherein the program is stored in the memory and is configured to be executed by a processor, and the program includes a method for executing a method for extracting blurred boundary objects in remote sensing images.

[0031] A computer-readable storage medium stores a computer program, wherein the computer program is executed by a processor to extract a method for extracting fuzzy boundary objects in remote sensing images.

[0032] A remote sensing image fuzzy boundary object extraction system includes an acquisition module, a classification module, a segmentation module and a fusion module; wherein,

[0033] The acquisition module is used to acquire high-resolution remote sensing images, where the objects to be extracted in the remote sensing images have fuzzy boundaries;

[0034] The classification module is used to classify and extract the target area of ​​the remote sensing image using a supervised classification algorithm to obtain the object data of the object to be extracted in the remote sensing image;

[0035] The segmentation module is used to segment the target area of ​​the remote sensing image using a mean shift algorithm to obtain mutually independent vector surface spots;

[0036] The fusion module is used to fuse the object data with the vector surface patch. When the overlap between the object data and the vector surface patch is greater than or equal to a set threshold range, the object type of the object data is assigned to the vector surface patch to obtain the blurred boundary object data of the remote sensing image.

[0037] Preferably, the classification module includes a deep learning classification algorithm unit, and the deep learning classification algorithm unit is used to classify and extract the target area of ​​the remote sensing image to obtain the rasterized ground object data in the remote sensing image;

[0038] The segmentation module includes a vectorization processing unit and a vectorization post-processing unit. The vectorization processing unit is used to segment the target area of ​​the remote sensing image using a mean shift algorithm to obtain independent image spot boundaries; the independent image spot boundaries are vectorized and faceted to obtain the mutually independent vector surface spots;

[0039] The vectorization post-processing unit is used to simplify and smooth the obtained vector surface pattern.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] The invention discloses a method for extracting fuzzy boundary objects in remote sensing images. The method uses a network model to learn multiple features of objects to obtain a rough classification of object categories, and then uses a mean shift algorithm to extract boundaries of the same remote sensing image. The method obtains extracted fuzzy boundary object data of the remote sensing image by fusing the classification result with the segmentation result. In the method, the land boundary obtained by using the automatic segmentation ability of the mean shift algorithm is more refined and the shape is more in line with the actual situation, so that the extraction of fuzzy boundary objects in remote sensing images is more accurate and more in line with reality.

[0042] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention and implement it according to the contents of the specification, the following is a detailed description of the preferred embodiments of the present invention in conjunction with the accompanying drawings. The specific implementation of the present invention is given in detail by the following embodiments and their accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0044] Figure 1 The overall structure flow chart of a method for extracting fuzzy boundary objects from remote sensing images of the present invention is as follows;

[0045] Figure 2 It is a specific flow chart of the supervised classification algorithm of the present invention;

[0046] Figure 3 It is a specific flow chart of the mean segmentation algorithm of the present invention;

[0047] Figure 4 It is an overall logic schematic diagram of a method for extracting fuzzy boundary objects from remote sensing images of the present invention;

[0048] Figure 5 It is a schematic diagram of the UNET network model of the present invention;

[0049] Figure 6 It is a logic diagram of the mean segmentation algorithm of the present invention;

[0050] Figure 7 It is a simple schematic diagram of the fusion of the ground feature data and the vector map patch of the present invention;

[0051] Figure 8 The remote sensing image target area of ​​the present invention;

[0052] Fig. 9 The segmentation result of the target area using the mean shift algorithm of the present invention;

[0053] Fig.10 for Fig. 9 Schematic diagram of enlarged details of some areas;

[0054] Fig.11 This is a schematic diagram of the grassland extracted using the UNET network model;

[0055] Fig.12 for Fig.11 Part of the area (with Fig.10 (a) is a schematic diagram of the details of the same area;

[0056] Fig.13 For the general Fig.10 and Fig.12 Schematic diagram after fusion;

[0057] Fig.14 For the general Fig.13 Schematic diagram after assigning grassland categories;

[0058] Fig.15 A schematic diagram of extracting fuzzy boundary objects in a target area of ​​a remote sensing image according to the present invention;

[0059] Fig.16 This is a module diagram of a remote sensing image fuzzy boundary object extraction system of the present invention. DETAILED DESCRIPTION

[0060] The present invention is further described below in conjunction with the accompanying drawings and specific implementation methods. It should be noted that, under the premise of no conflict, the various embodiments or technical features described below can be arbitrarily combined to form a new embodiment.

[0061] The extraction of remote sensing image objects in the present invention (such as grasslands and sandy land) is different from conventional objects (such as roads and buildings). Objects with fuzzy boundaries such as grasslands, sandy lands, and woodlands, i.e., weak boundary objects, are difficult to define, mainly because the generation or change of these objects is gradual with the adjacent objects, and there is no obvious boundary. In addition, the definition of the minimum geographic spot of remote sensing images, i.e., the minimum independent spot at a certain scale, often appears in the same category. Taking grassland as an example, it is impossible for a large area of ​​grassland to grow exactly the same. There must be a sparse growth of the grass in the middle part. Then this sparse grassland is also an independent spot relative to other grasslands, and it is also separable from the perspective of segmentation. Although convolutional neural networks are relatively mature in semantic segmentation and are widely used in object extraction, there is still a lot of room for improvement in the segmentation ability of objects with fuzzy boundaries, because the growth distribution of these objects is relatively complex, especially the form constructed by the edge is often not in line with reality visually, and the accuracy is naturally very low.

[0062] The distribution of objects in remote sensing images is spatially complex. At the same time, the intra-class and inter-class differences of objects are also very large. It is impossible to solve the classification problem of objects with one method, and the fuzzy boundary objects are even more difficult to extract with ordinary methods. The present invention proposes a method of combining classification with segmentation to constrain the distribution of fuzzy boundary objects, fully explores the characteristics of weak boundary objects, and obtains beautiful patches.

[0063] The present invention provides a method for extracting fuzzy boundary objects in remote sensing images. Figure 1-7 As shown, the following steps are included:

[0064] S1. Obtain a high-resolution remote sensing image, in which the objects to be extracted have fuzzy boundaries. In one embodiment, obtain a 4-band 2-meter spatial resolution high-resolution remote sensing image of Gaofen No. 1, and select one or more objects with fuzzy boundaries, such as grassland, sand, saline-alkali land, etc.

[0065] S2. Classify and extract the target area of ​​the remote sensing image using a supervised classification algorithm to obtain the object data of the object to be extracted in the remote sensing image. In one embodiment, supervised classification is a type of remote sensing image classification, that is, the process of using sample pixels of a confirmed category to identify other unknown category pixels. Sample pixels of confirmed categories refer to those located in the training area. In this classification, the analyst selects a certain number of training areas for each category on the image, and the computer calculates the statistics or other information of each training sample area. Each pixel is compared with the training sample and divided into the sample class that is most similar to it according to different rules. Supervised classification can be divided into two basic steps: selecting training samples and extracting statistical information, and selecting a classification algorithm. The deep learning classification model is used to roughly extract the fuzzy boundary object category. The deep learning method is used because the extraction effect of the deep learning classification model is more stable than the traditional method.

[0066] Specifically, the method of classifying and extracting the target area of ​​the remote sensing image using a supervised classification algorithm includes:

[0067] The deep learning classification algorithm is used to classify and extract the target area of ​​remote sensing images, such as Figure 2 As shown, the following steps are included:

[0068] S21, segment and crop the remote sensing image to obtain a surface vector file containing one or more fuzzy boundary objects, draw the object labels in the surface vector file, convert the vector data of the surface vector samples labeled as objects into raster data, and obtain a rasterized object sample set. A small number of 1000*1000 samples are selected from the acquired Gaofen No. 1 2-meter spatial resolution 4-band remote sensing image and drawn and converted to obtain a sample set containing images and labels.

[0069] S22, adjusting the UNET network model parameters, performing UNET-based model training on the ground object sample set, and obtaining a ground object model. Figure 5As shown in the figure, the sample set is input into the convolutional neural network. The convolutional neural network model is preferably UNET (Fully Convolutional Neural Network Image Segmentation). UNET is a convolutional neural network with multi-scale and suitable for semantic segmentation of remote sensing large images by adding an upsampling stage on the basis of the FCN network (Fully Convolutional Network). The UNET network structure is mainly divided into two parts. The left part is a convolutional network for extracting image features like FCN, and the right part is the deconvolution operation of the upsampling part. Many feature channels are added in this stage to allow more original image texture information to be propagated in the high-resolution layer. It is particularly noted that UNET has no fully connected layer, and the whole process uses valid convolution to ensure that the segmentation results are obtained based on the non-missing context features. Therefore, the image sizes of input and output will be inconsistent. For large image input, overlap-strategy can be used for seamless image output. In addition, since the upsampling part will fuse the output of the feature extraction part, this is actually a multi-scale fusion. The features of the object come from the output of the first convolution block and the output of the upsampling. Such a connection runs through the entire network and can fully extract image features.

[0070] It should be noted that the rough classification results obtained by the UNET network model do not require fine classification results; general network models are improved and optimized by referring to the UNET model, so the results obtained by adjusting the parameters of the UNET network model can meet the requirements of this embodiment.

[0071] S23, input the remote sensing image data to be tested into the object model, and calculate the inspection evaluation function of this test. If the evaluation function value meets the standard, jump to the next step; if the evaluation function value does not meet the standard, adjust the UNET network model parameters and return to the step sample preparation and modify the sample set to re-iterate the training. The evaluation function preferably selects the MIoU (mean intersection-over-union) value and the MPA (mean pixel accuracy). If the MIoU value and the MPA value meet the standard, jump to the next step. If the MIoU value or the MPA value does not meet the standard, adjust the UNET network model parameters and return to S1 and modify the sample to re-iterate the training; for example, set the MIoU value range to 0.70-1, and set the MPA value range to 0.75-1 to be considered as meeting the standard. When it exceeds the range, adjust the UNET network model parameters and return to S21 for re-training.

[0072] S24, using the object model that has been evaluated as meeting the standard in step accuracy evaluation to predict the object in the target area of ​​the remote sensing image, and obtain the rasterized object data in the remote sensing image. The target area of ​​the remote sensing image is input into the object model that has been evaluated as meeting the standard in step S23, and a classification result is obtained, and the raster result is post-processed. The raster post-processing mainly removes spots and fills small holes in the classification result to ensure the aesthetics of the result.

[0073] S3. Use the mean shift algorithm to segment the target area of ​​the remote sensing image to obtain independent vector surface patches. In one embodiment, the mean shift algorithm is used to segment the same remote sensing image target area as in step S2. The mean shift algorithm has automatic segmentation capabilities, does not require manual parameter setting, and is better than general clustering methods. The mean shift algorithm is based on a sliding window and iterates out areas with dense data points. The most important thing is that compared with other clustering algorithms, it does not need to set the number of categories, but determines the number of categories through iteration, and this time the algorithm is mainly used to segment to obtain reasonable edges, and the categories are finally assigned through a convolutional network.

[0074] The basic formula of mean shift algorithm is: S h :x i The high-dimensional sphere region with a center point and a radius of h; k: contained in S h The number of points in the range; x i : Included in S h The points within the range are used to calculate the drift mean. The center update formula is x t+1 =M t +x t , used to move the center point to the offset mean position, where Mt is the offset mean obtained under state t; xt is the center under state t.

[0075] Specifically, Figure 3 , Figure 6 As shown, the following steps are included:

[0076] S31, selecting a center point, randomly selecting a point in the unclassified remote sensing image data as the center point;

[0077] S32, obtaining a set, obtaining all points within the bandwidth from the center point, recorded as a set M, and the set M forms a cluster;

[0078] S33, calculating a vector, calculating a vector from the center point to each element in the set M, and adding these vectors to obtain an offset vector;

[0079] S34, drift movement, the center point moves along the drift direction, and the moving distance is the modulus of the offset vector;

[0080] S35, obtaining a new center point, repeating the steps of obtaining a set, calculating a vector, and drifting and moving, until the size of the offset vector is within a set threshold range, and obtaining a new center point at this time;

[0081] S36, forming clusters, repeating the steps of obtaining sets, calculating vectors, and drifting and moving until all points in the remote sensing image data are clustered;

[0082] S37, determining the cluster to which the current point belongs, and taking the cluster with the largest access frequency as the cluster to which the current point belongs, i.e., the independent patch boundary, according to the access frequency of each cluster to each point.

[0083] After step S37, the method further includes the following steps:

[0084] S38. Use the mean shift algorithm to segment the target area of ​​the remote sensing image to obtain independent spot boundaries; vectorize the obtained independent spot boundaries and perform faceting to obtain the mutually independent vector surface spots. In one embodiment, the mean shift algorithm is used to perform multi-scale segmentation on the target area of ​​the remote sensing image to obtain boundaries that fit the independent spots and achieve the best morphological state. The obtained independent spots are vectorized, and the isolated line segments are extended and faceted to obtain mutually independent vector surface spots. The vector surface spots obtained by the mean shift algorithm are rasterized to remove broken spots and fill small holes to ensure the aesthetics of the result.

[0085] It should be noted that the mean shift algorithm obtains the edge morphology of the smallest indivisible patch of remote sensing images at a certain scale. When the obtained patches are not detailed enough, the segmentation scale can be adjusted to obtain more detailed patch morphology.

[0086] S39, vector post-processing, the obtained vector surface pattern has serious jaggedness, and the vector surface pattern needs to be simplified and smoothed to obtain a vector surface pattern with beautiful appearance and little change in overall shape.

[0087] S4, fusing the object data with the vector surface patch, and when the overlap between the object data and the vector surface patch is greater than or equal to a set threshold range, assigning the object type of the object data to the vector surface patch to obtain the fuzzy boundary object data of the remote sensing image. Figure 7As shown, the feature data obtained in step S2 is merged with the vector surface patch obtained in step S3; generally, the area size of each patch is calculated, and when the overlap between the feature data obtained and the area of ​​the vector surface patch is greater than or equal to the set threshold range, the threshold range is generally set to 70%-90%, that is, when the overlap between the feature data obtained through the network model and the vector surface patch at the same position obtained through the mean shift algorithm is within 70%-90%, the boundary of the vector surface patch obtained by the mean shift algorithm is used as the extracted feature boundary and the feature type of the vector surface patch is assigned. For example, if the feature data at the position obtained through the network model is grassland, the grassland is assigned to the vector surface patch, and finally an extraction result of a fuzzy boundary land class whose morphology conforms to the actual one is obtained.

[0088] For example, Figure 8-15 As shown, taking grassland as an example, high-density grassland patches are found and assigned values ​​through overlapping statistics. Through actual production, it only takes about 4 hours to extract all weak boundary objects including grassland, saline-alkali land, bare land, etc. in the target area of ​​169 square kilometers. Manual drawing is not only time-consuming, but also very difficult to distinguish the minimum separable patches of remote sensing images at different scales with the naked eye, which is basically impossible to implement manually. The present invention effectively improves the efficiency of extracting fuzzy boundary objects in remote sensing images while ensuring accuracy. Among them, Figure 8 The target area of ​​the remote sensing image of the present invention is a four-band remote sensing image mainly composed of weak boundary objects, including grassland, bare land, saline-alkali land, etc., among which the distribution of grassland is also relatively complex, and the sparseness of different areas varies greatly; Fig. 9 In order to use the mean shift algorithm to segment the entire target area, the mean shift algorithm can reduce the remote sensing image segmentation to the smallest indivisible patch at the set segmentation scale; Fig.10 for Fig. 9 The enlarged image is the detailed display image. It can be seen that the algorithm has obtained the edge morphology of the smallest indivisible patch of remote sensing images at a certain scale, whether between categories or within categories. In addition, if you feel that the patch is not detailed enough, you can adjust the segmentation scale to obtain a more detailed patch morphology; Fig.11 This is the classification result of the UNET network model. This test takes high-density grassland as an example and extracts the high-density grassland in the entire target area through the network model. Fig.12 This is a detailed display of the classification results of high-density grassland by the UNET network model. At this time, the land class boundary is not close enough to the actual boundary and needs to be optimized in combination with the segmentation results. Fig.13 It is the result of fusing the same type of patches after combining the segmentation result of the mean shift algorithm with the classification result of weak boundary objects. After the grassland, sandy land, bare land and other categories are fused with the segmentation result, the result is obtained by fusing the segmented patches of the same category. Fig.14The result of high-density grassland clustering areas after assigning categories to segmented patches through calculation of overlap; Fig.15 This is the extraction result of fuzzy boundary objects in the target area of ​​the remote sensing image.

[0089] An electronic device, characterized in that it comprises: a processor; a memory; and a program, wherein the program is stored in the memory and is configured to be executed by the processor, and the program includes a method for extracting a fuzzy boundary object from a remote sensing image. A computer-readable storage medium, having a computer program stored thereon, characterized in that the computer program is executed by the processor to perform a method for extracting a fuzzy boundary object from a remote sensing image.

[0090] A remote sensing image fuzzy boundary object extraction system, such as Fig.16 As shown, it includes an acquisition module, a classification module, a segmentation module and a fusion module; wherein,

[0091] The acquisition module is used to acquire high-resolution remote sensing images, where the objects to be extracted in the remote sensing images have fuzzy boundaries;

[0092] The classification module is used to classify and extract the target area of ​​the remote sensing image using a supervised classification algorithm to obtain the object data of the object to be extracted in the remote sensing image;

[0093] The segmentation module is used to segment the target area of ​​the remote sensing image using a mean shift algorithm to obtain mutually independent vector surface spots;

[0094] The fusion module is used to fuse the object data with the vector surface patch. When the overlap between the object data and the vector surface patch is greater than or equal to a set threshold range, the object type of the object data is assigned to the vector surface patch to obtain the blurred boundary object data of the remote sensing image.

[0095] Furthermore, the classification module includes a deep learning classification algorithm unit, which uses the deep learning classification algorithm unit to classify and extract the target area of ​​the remote sensing image to obtain the rasterized ground object data in the remote sensing image;

[0096] The segmentation module includes a vectorization processing unit and a vectorization post-processing unit. The vectorization processing unit is used to segment the target area of ​​the remote sensing image using a mean shift algorithm to obtain independent image spot boundaries; the independent image spot boundaries are vectorized and faceted to obtain the mutually independent vector surface spots;

[0097] The vectorization post-processing unit is used to simplify and smooth the obtained vector surface pattern.

[0098] The above are only preferred embodiments of the present invention and are not intended to limit the present invention in any form. Any ordinary technician in the industry can smoothly implement the present invention as shown in the drawings and above. However, any equivalent changes, modifications and evolutions made by technicians familiar with the profession without departing from the scope of the technical solution of the present invention using the technical content disclosed above are all equivalent embodiments of the present invention. At the same time, any equivalent changes, modifications and evolutions made to the above embodiments based on the essential technology of the present invention are still within the protection scope of the technical solution of the present invention.

Claims

1. A method for extracting fuzzy boundary objects from remote sensing images, characterized in that: The steps include: Acquire high-resolution remote sensing images, where the objects to be extracted have fuzzy boundaries; Use supervised classification algorithm to classify and extract the target area of ​​remote sensing images, and obtain the object data of the objects to be extracted in the remote sensing images; The target area of ​​the remote sensing image is segmented by using a mean shift algorithm to obtain mutually independent vector surface patches; The object data is merged with the vector surface patch, and when the overlap between the object data and the vector surface patch is greater than or equal to a set threshold range, the object type of the object data is assigned to the vector surface patch to obtain fuzzy boundary object data of the remote sensing image; The method of using a supervised classification algorithm to classify and extract the target area of ​​the remote sensing image includes: using a deep learning classification algorithm to classify and extract the target area of ​​the remote sensing image, including the following steps: Segmenting and cropping the remote sensing image to obtain a surface vector file containing one or more fuzzy boundary objects, and drawing object labels in the surface vector file, converting vector data of several surface vector samples labeled as objects into raster data to obtain a rasterized object sample set; Adjusting the UNET network model parameters, performing UNET-based model training on the ground object sample set, and obtaining a ground object model; Input the remote sensing image data to be tested into the ground object model, and calculate the evaluation function of this test. If the evaluation function value meets the standard, jump to the next step; if the evaluation function value does not meet the standard, adjust the UNET network model parameters and return to the step of sample preparation and modify the sample set to re-iterate training; Using the evaluated qualified ground object model to predict the ground object in the target area of ​​the remote sensing image, obtain the rasterized ground object data in the remote sensing image, and remove the broken spots and fill the small holes of the ground object data and the vector surface spots; Wherein, the target area of ​​the remote sensing image is segmented by using the mean shift algorithm to obtain Independent vector surface patches, including: The target area of ​​the remote sensing image is segmented by using a mean shift algorithm to obtain independent image spot boundaries; the independent image spot boundaries are vectorized and faceted to obtain the mutually independent vector surface spots; The obtained vector surface pattern is simplified and smoothed.

2. The method for extracting fuzzy boundary objects in remote sensing images according to claim 1, characterized in that: In step 1, the target area of ​​the remote sensing image is segmented using the mean shift algorithm to obtain independent patch boundaries, including the following steps: Select the center point and randomly select a point as the center point in the unclassified remote sensing image data; Obtain a set, obtain all points within the bandwidth from the center point, record them as set M, and the set M forms a cluster; Calculate the vector, calculate the vector from the center point to each element in the set M, add these vectors together to obtain the offset vector; Drift movement, the center point moves along the drift direction, and the distance of movement is the modulus of the offset vector; Obtain a new center point, repeat the steps of obtaining a set, calculating a vector, and drifting and moving until the size of the offset vector is within a set threshold range, and obtain a new center point at this time; Form clusters, repeat the steps of obtaining sets, calculating vectors, and drifting and moving until all points in the remote sensing image data are clustered; Determine the cluster to which the current point belongs, and according to the access frequency of each cluster to each point, take the cluster with the largest access frequency as the cluster to which the current point belongs, that is, the independent patch boundary.

3. An electronic device, characterized in that include: processor; Memory; and a program, wherein the program is stored in the memory and configured to be executed by a processor, the program including instructions for executing the method according to claim 1.

4. A computer-readable storage medium having a computer program stored thereon, characterized in that: The computer program is used by a processor to execute the method according to claim 1 .

5. A remote sensing image fuzzy boundary object extraction system, characterized in that: It includes acquisition module, classification module, segmentation module and fusion module; among them, The acquisition module is used to acquire high-resolution remote sensing images, where the objects to be extracted in the remote sensing images have fuzzy boundaries; The classification module is used to classify and extract the target area of ​​the remote sensing image using a supervised classification algorithm to obtain the object data of the object to be extracted in the remote sensing image; The segmentation module is used to segment the target area of ​​the remote sensing image using a mean shift algorithm to obtain mutually independent vector surface spots; The fusion module is used to fuse the object data with the vector surface patch, and when the overlap between the object data and the vector surface patch is greater than or equal to a set threshold range, the object type of the object data is assigned to the vector surface patch to obtain the remote sensing image fuzzy boundary object data; The classification module includes a deep learning classification algorithm unit, and the deep learning classification algorithm unit is used to classify and extract the target area of ​​the remote sensing image to obtain the rasterized ground object data in the remote sensing image, including the following steps: Segmenting and cropping the remote sensing image to obtain a surface vector file containing one or more fuzzy boundary objects, and drawing object labels in the surface vector file, converting vector data of several surface vector samples labeled as objects into raster data to obtain a rasterized object sample set; Adjusting the UNET network model parameters, performing UNET-based model training on the ground object sample set, and obtaining a ground object model; Input the remote sensing image data to be tested into the ground object model, and calculate the evaluation function of this test. If the evaluation function value meets the standard, jump to the next step; if the evaluation function value does not meet the standard, adjust the UNET network model parameters and return to the step of sample preparation and modify the sample set to re-iterate training; Using the evaluated qualified ground object model to predict the ground object in the target area of ​​the remote sensing image, obtain the rasterized ground object data in the remote sensing image, and remove the broken spots and fill the small holes of the ground object data and the vector surface spots; The segmentation module includes a vectorization processing unit and a vectorization post-processing unit. The vectorization processing unit is used to segment the target area of ​​the remote sensing image using a mean shift algorithm to obtain independent image spot boundaries; the independent image spot boundaries are vectorized and faceted to obtain the mutually independent vector surface spots; The vectorization post-processing unit is used to simplify and smooth the obtained vector surface pattern.

Images