Remote sensing image water body extraction method and system combining global features and local intercommunication

Abstract

The application discloses a kind of remote sensing image water body extraction method and system combined with global feature and local intercommunication, first, on the basis of obtaining water initial result, image is carried out denoising processing using Gaussian filter, the linear feature of strongest response is extracted by multi-scale Frangi filtering, then the linear branch is extracted using OTSU binaryzation segmentation algorithm and combined with initial result to carry out spectral inspection, linear branch is superimposed with initial result to obtain preliminary optimization result;Second, the connection of broken flow part in preliminary optimization result is carried out by combining local SSIM index and intercommunication algorithm, the accurate optimization of linear water body is realized;Finally, water body part is extracted by K-means clustering, then topological inspection and spectral inspection are carried out to water clustering result and accurate optimization graph, based on image size, small area misclassified part is removed, further optimization of blocky water body is realized.The application improves the degree of automation of optimization, and further improves the precision of water extraction.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of remote sensing image application, and relates to a remote sensing image water body extraction method and system, in particular to a remote sensing image water body extraction post-processing method and system combining global features and local intercommunication. BACKGROUND

[0002] Water body extraction plays an important role in ecological environment monitoring. The water body range is an important indicator of the conditions of the water resources of a river basin, and it is of great significance for the rational use of water resources by humans. The information it provides can help the use and protection of water resources, the assessment and prevention of natural disasters such as floods, and guide agricultural production around the water area. Since the beginning of remote sensing, this technology has become increasingly mature, helping us to quickly and comprehensively obtain the distribution of surface water and continuously monitor its dynamic changes. Among them, the complete and accurate extraction of the water body range from high-resolution remote sensing images is of great significance for understanding the current water resource situation, rational planning and management of water resources, and human social activities.

[0003] At present, the methods for water body extraction optimization mainly include the following three kinds: improved water body index method, morphological operation optimization method, and deep learning-based method.

[0004] Among them, the improved water body index method is to distinguish water from background according to its spectral characteristics, and various studies have been conducted on the extraction of water bodies. Based on the analysis of the spectral reflectance characteristics of rocks and soil in Landsat8 multi-band images, Yi Ran et al. [1] improved the normalized difference water index and proposed a mountain water index. Xilina Yidu et al. [2] used four kinds of water body extraction analysis, including normalized difference water index, improved normalized difference water index, automatic water body index, and improved automatic water body index, and concluded that the improved automatic water body index had the highest accuracy. Yang Ji et al. [3] proposed a city water index method and combined it with a fractal geometry algorithm to realize automatic extraction of water body information in complex urban environments. This method overcomes the discontinuity of water body extraction results in the SVM algorithm and the situation of partial river channel missing. This kind of method is easy to produce different results due to different image quality, which has different degrees of influence on the accuracy of water body extraction, and has great limitations.

[0005] In the research of optimizing water extraction results through morphological operation, Zhang Qingchun et al. proposed a method based on multi-feature fusion and soft voting combined with morphological operation to optimize water extraction, but the field and other targets similar to the characteristics of river areas will be misclassified[4-5]. Chen Jingjue et al. designed a morphological optimization method combined with an elevation connectivity calibration algorithm, which sets an elevation connectivity threshold to discuss the rules that water bodies should meet to extract fine water bodies, but small lakes may be deleted by the established rules. Feng Zhaohui et al. studied a method based on spectral information component graph and morphological principle to improve the extraction of water bodies by watershed transformation to solve the problems of incomplete water extraction and incorrect differentiation with surrounding features. This method is easily affected by the surrounding features of water bodies, resulting in incomplete water body edges.

[0006] In recent years, deep learning has also been widely used in water extraction. For example, the fully convolutional neural network (FCN) widely used in semantic segmentation in recent years[8], as well as the multi-scale combined fully convolutional neural network[9] based on its architecture, U-type convolutional neural network and optimization

[10] . However, the above methods pay less attention to the edge, such as water edge blur, vegetation or shadow blocking near the river, causing missed detection, river cutoff, etc., which need further processing to improve the building extraction accuracy.

[0007] [Document 1] Yi Ran, Zhang Li, Geng Qing. Water body boundary extraction method based on mountain water body index[J]. Geospatial Information, 2022, 20(03): 25-29.

[0008] [Document 2] Xilinaiyi Duolai, Alimujiang Kasimu, Rukeya Rehemann, Liang Hongwu. Extraction of Ebinur Lake water surface based on four water body indices and spatiotemporal variation analysis[J]. Journal of Yangtze River Scientific Research Institute, 2022, 39(10): 134-140.

[0009] [Document 3] Yang Ji, Han Liusheng, et al. A method for extracting water bodies from OLI remote sensing images based on urban water body index and fractal geometry algorithm[J]. Bulletin of Surveying and Mapping, 2018(04): 44-49. DOI: 10.13474 / j.cnki.11-2246.2018.0108.

[0010] [Document 4] Zhang Qingchun, Tong Guofeng, Li Yong, Gao Liwei, Chen Huai Rong. Remote sensing image river detection based on multi-feature fusion and soft voting[J]. Acta Optica Sinica, 2018, 38(06): 320-326.

[0011] [Document 5] Fu Baojing, Li Zili. Remote sensing image river extraction based on multi-feature fusion [J]. China Rural Water Resources and Hydropower, 2022 (12): 53-58.

[0012] [Document 6] Chen Jingjue, Liu Rui, Yang Xin, Yang Mei, Yang Yuantao. Improved Otsu and morphological water information extraction [J]. Remote sensing information, 2022, 37 (01): 101-109.

[0013] [Document 7] Feng Zhaohui, Li Qian, Han Liusheng, Xue Guochao, Zhao Hongying. Remote sensing image water extraction based on improved watershed method [J]. Surveying and mapping bulletin, 2019 (06): 11-15. DOI: 10.13474 / j.cnki.11-2246.2019.0175.

[0014] [Document 8] Shen Junao, Ma Mengting, Song Zhiyuan, Liu Tingzhou, Zhang Wei. High resolution remote sensing image water extraction based on deep learning semantic segmentation model [J]. Natural resources remote sensing, 2022, 34 (04): 129-135.

[0015] [Document 9] Kang Jian. High resolution remote sensing image water extraction based on multi-scale feature enhancement network [D]. Nanjing University of Information Engineering, 2022. DOI: 10.27248 / d.cnki.gnjqc.2022.000355.

[0016] [Document 10] Zheng Taihao, Wang Qingtao, Li Jiagu, Zheng Fengbin, Zhang Yonghong, Zhang Ning. High resolution six image water body automatic extraction based on deep learning [J]. Science, technology and engineering, 2021, 21 (04): 1459-1470. SUMMARY

[0017] In order to solve the influence of water body extraction in high resolution remote sensing image, combined with the spectrum and texture characteristics of water body itself, the application provides a kind of remote sensing image water body extraction method and system combining global feature and local intercommunication.

[0018] The technical scheme adopted by the method of the application is: a kind of remote sensing image water body extraction method combining global feature and local intercommunication, comprising the following steps:

[0019] Step 1: using deep learning method to extract water body from original high resolution remote sensing image, to automatically obtain water body initial extraction result;

[0020] Step 2: for original high resolution remote sensing image, using multi-scale Frangi filtering algorithm and segmentation algorithm to obtain linear feature in image, and combining water body initial extraction result to carry out spectral inspection, to obtain linear branch in image, and superimpose linear branch with initial extraction result, to obtain preliminary optimization result;

[0021] Step 3: Combine the local SSIM index with the interworking algorithm to connect and supplement the broken flow existing in the preliminary optimization result, and obtain the accurate optimization graph of the linear branch flow;

[0022] Step 4: Extract the water part in the high-resolution remote sensing image through the clustering method, then perform topological inspection and spectral inspection on the water clustering result and the accurate optimization graph obtained in step 3, remove the small area of the wrong part based on the image size, realize the further optimization of the block-shaped water body, and obtain the final water body optimization result.

[0023] The technical scheme of the system of the application is: a remote sensing image water body extraction system combining global features and local interworking, comprising the following modules:

[0024] Module 1 is used for water body extraction of the original high-resolution remote sensing image by using a deep learning method, and automatically obtaining the initial extraction result of the water body;

[0025] Module 2 is used for obtaining the linear features in the image by using a multi-scale Frangi filtering algorithm and a segmentation algorithm on the original high-resolution remote sensing image, and performing spectral inspection in combination with the initial extraction result of the water body, obtaining the linear branch flow in the image, and superimposing the linear branch flow and the initial extraction result to obtain a preliminary optimization result;

[0026] Module 3 is used for connecting and supplementing the broken flow existing in the preliminary optimization result by combining the local SSIM index with the interworking algorithm, and obtaining the accurate optimization graph of the linear branch flow;

[0027] Module 4 is used for extracting the water part in the high-resolution remote sensing image through the clustering method, then performing topological inspection and spectral inspection on the water clustering result and the accurate optimization graph obtained in module 3, removing the small area of the wrong part based on the image size, realizing the further optimization of the block-shaped water body, and obtaining the final water body optimization result.

[0028] Compared with the prior art, the application has the following advantages:

[0029] 1. The water body optimization effect is more targeted. Due to the different complexity of high-resolution image backgrounds, some water bodies are in complex scenes, and the size of the water body is different and the shape is various. The previous method can extract most of the water body, but some small branch flows are missed. The global-local linkage optimization watershed water body post-processing method based on spatial clustering and multi-scale Frangi filtering fusion in the present application can realize active perception and indexing of small branch flows in a large watershed range.

[0030] 2. Strong universality, wide application range. Because there are many types of data sources that can be used for water extraction, existing methods often combine multiple data sources to extract together to improve extraction accuracy. However, this will make the method too dependent on data sources, and the absence of any data will cause the method to be unable to be applied. Therefore, in order to improve the universality of the method and make it truly applied to more remote sensing applications, only a single high-resolution visible light image is used as the basic data source without any auxiliary data and prior knowledge. Visible light data includes R, G, B three bands, which is the most common image data, which can be easily obtained through aerial, space and ground remote sensing methods, and has wide coverage. Therefore, based on visible light remote sensing data, the building extraction method researched expands the application range from the data source itself, which helps to enhance the universality of the method.

[0031] 3. High optimization efficiency. Compared with other methods, the method has high automation degree and the water body optimization contour is closer to the real contour of the water body. The edge is smooth, without fine strip-shaped burrs, and can better preserve small concave edges. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 The method flowchart of the embodiment of the present application is shown in the figure;

[0033] Figure 2 The water body optimization result and deep learning extraction result comparison chart of the embodiment of the present application is shown in the figure, wherein (a) is the original image, (b) is the deep learning result, and (c) is the optimization result of the present method. DETAILED DESCRIPTION

[0034] In order to facilitate those skilled in the art to understand and implement the present application, the present application is further described in detail below in combination with the drawings and embodiments. It should be understood that the embodiments described herein are only used to illustrate and explain the present application, and are not used to limit the present application.

[0035] See Figure 1 The present application provides a remote sensing image water body extraction method combining global features and local intercommunication, which comprises the following steps:

[0036] Step 1: using a deep learning method to extract water body from the original high-resolution remote sensing image to automatically obtain the initial extraction result of the water body;

[0037] Step 2: using a multi-scale Frangi filtering algorithm and a segmentation algorithm to obtain linear features in the image from the original high-resolution remote sensing image, and combining the initial extraction result of the water body to perform spectral inspection to obtain linear branch streams in the image, and superimposing the linear branch streams and the initial extraction result to obtain a preliminary optimization result;

[0038] In this embodiment, Gaussian filter is used to denoise the image, and multi-scale Frangi filter is used to extract the strongest linear features, and then OTSU binary segmentation algorithm is used to obtain the linear feature map.

[0039] In this embodiment, the linear tributaries in the image are obtained by combining the initial extraction result of the water body with spectral inspection. The average RGB value of the water body in the whole image is calculated by respectively accumulating the RGB values of the water body pixel points corresponding to the initial extraction result in the whole image. A certain error parameter (the optimal value is [0.8, 1.2]) is set, and then the average RGB value is used to scan the linear feature map. The linear features that do not exceed the error parameter are considered as linear tributaries.

[0040] Step 3: Combine the local SSIM index and the interworking algorithm to connect and supplement the broken flow in the preliminary optimization result, and obtain the accurate optimization map of the linear tributaries.

[0041] In this embodiment, the specific implementation of step 3 includes the following sub-steps:

[0042] Step 3.1: Extract the skeleton and endpoint of the river through the preliminary optimization result.

[0043] Step 3.2: According to the river endpoint, obtain the circumscribed rectangle of the possible broken flow area and the corresponding image subgraph.

[0044] In this embodiment, a certain distance threshold (the optimal value is [5, 10]) is set to traverse the river endpoints. Two endpoints are taken as a group, and the endpoints are taken as the diagonal lines to extend outward by a predetermined distance, which is taken as the two diagonal angles of the circumscribed rectangle. In this way, the circumscribed rectangle and the corresponding image subgraph are obtained.

[0045] Step 3.3: Combine the local SSIM index and the interworking algorithm to connect and supplement the broken flow area.

[0046] In this embodiment, for each image subgraph, the SSIM value of each pixel in the subgraph is calculated. A certain index threshold is set. If the SSIM value of a certain pixel point is within the index threshold range of the SSIM values of the two endpoints, it is considered that the pixel point is similar in structure to the two endpoints. In this way, the pixel points similar in structure to the two endpoints in the subgraph are obtained.

[0047] In this embodiment, the obtained pixel points are used to obtain different connected regions in the subgraph by using the connected domain algorithm and marking. If the two endpoints are the same label, it is considered that the two endpoints are the same connected domain. The connected region is supplemented to the preliminary optimization result of step 1 to obtain the accurate optimization result of the water body, and the accurate optimization of the linear tributaries is realized.

[0048] Step 4: Extract the water body part in the high-resolution remote sensing image by a clustering method, then perform topological inspection and spectral inspection on the water body clustering result and the precise optimization map obtained in step 3, remove the small-area misclassified part based on the image size, realize further optimization of the block-shaped water body, and obtain the final water body optimization result.

[0049] In this embodiment, the specific implementation of step 4 includes the following sub-steps:

[0050] Step 4.1: Classify the image map by using K-means clustering, calculate the coincidence degree of the precise optimization result and each category, and determine the category with the highest coincidence degree as the water body category to obtain the clustering water body result.

[0051] Step 4.2: Perform topological analysis on the clustering water body result and the precise optimization result to obtain the intersecting part and the non-intersecting part, and retain the intersecting part of the two.

[0052] Step 4.3: Perform spectral inspection on the non-intersecting part. First, accumulate the RGB values of the water body pixel points corresponding to the precise optimization result in the entire image, and calculate the average RGB value of the water body in the entire image. Set a certain error parameter (the optimal value is [0.8, 1.2]), then scan the non-intersecting part according to the average RGB value, and determine the non-intersecting part that does not exceed the error parameter as the water body. Then output the non-intersecting water body that passes the spectral inspection to the precise optimization result to obtain the further optimization result.

[0053] Step 4.4: Set a certain area threshold (the optimal value is the image size *[0.001, 0.005]) according to the size of the image, remove the part of the water body in the optimization result obtained in step 4.3 whose area is less than the above area threshold, and obtain the final water body optimization result.

[0054] See Figure 2 The comparison chart of the water body optimization result obtained by using the method of the present application and the deep learning extraction result shows that, with the original image as a reference, after the water body is optimized by the method of the present application, the small tributaries that are not extracted by the deep learning method are extracted by the present application, and the water body basin is more perfect and accurate.

[0055] Firstly, on the basis of the initial result of water body obtained by the deep learning method, the image graph is denoised by using Gaussian filtering, the linear features with the strongest response are extracted by multi-scale Frangi filtering, the linear branch is extracted by using OTSU binary segmentation algorithm and combining the initial result for spectral inspection, the linear branch is superimposed with the initial result to obtain the preliminary optimized result; secondly, the disconnected part in the preliminary optimized result is connected by combining the local SSIM index with the interoperability algorithm, so that the linear water body is accurately optimized; finally, the water body part is extracted by K-means clustering, then the water body clustering result and the accurate optimization graph are topologically inspected and spectrally inspected, the small area of the wrong part is removed based on the image size, so that the blocky water body is further optimized. The present application improves the automation degree of optimization and further improves the precision of water body extraction.

[0056] It should be understood that the above description of the preferred embodiments is more detailed and should not be considered as limiting the scope of patent protection of the present application. Those skilled in the art can make substitutions or modifications without departing from the scope of protection claimed by the present application under the inspiration of the present application, and all fall within the scope of protection of the present application. The scope of protection claimed by the present application shall be subject to the appended claims.

Claims

1. A method for water body extraction from remote sensing images by combining global features and local intercommunication, characterized in that, The method comprises the following steps: Step 1: water body extraction is performed on the original high-resolution remote sensing image to automatically obtain an initial extraction result of the water body; Step 2: linear features in the original high-resolution remote sensing image are obtained by using a filtering algorithm and a segmentation algorithm, and spectral inspection is performed in combination with the initial extraction result of the water body to obtain linear branch streams in the image, and the linear branch streams are superimposed on the initial extraction result to obtain a preliminary optimization result; Step 3: disconnected streams in the preliminary optimization result are connected and supplemented to obtain an accurate optimization map of the linear branch streams; The specific implementation of Step 3 comprises the following sub-steps: Step 3.1: the skeleton and endpoints of the river are extracted from the preliminary optimization result; Step 3.2: according to the river endpoints, an inscribed rectangle of a possible disconnected stream region and a corresponding image sub-map are obtained; A certain distance threshold is set to traverse the river endpoints, two endpoints are taken as a group, and the endpoints are taken as diagonal lines to extend outward by a preset distance, which is taken as two diagonal corners of the inscribed rectangle, so as to obtain the inscribed rectangle and the corresponding image sub-map; Step 3.3: the disconnected stream region is connected and supplemented in combination with a local SSIM index and an intercommunication algorithm; For each image sub-map, the SSIM value of each pixel in the sub-map is calculated, a certain index threshold is set, and if the SSIM value of a certain pixel point is within the index threshold range of the SSIM values of the two endpoints, it is considered that the pixel point is similar in structure to the two endpoints, and in this way, the pixel points similar in structure to the two endpoints in the sub-map are obtained; By using the obtained pixel points, different connected regions in the sub-map are obtained by using a connected domain algorithm and are labeled, if the two endpoints are of the same label, it is considered that the two endpoints are of the same connected domain, and the connected region is supplemented to the preliminary optimization result of Step 1 to obtain an accurate optimization result of the water body, and accurate optimization of the linear branch streams is realized; Step 4: the water body part in the high-resolution remote sensing image is extracted, and then topological inspection and spectral inspection are performed on the water body clustering result and the accurate optimization map obtained in Step 3, small area misclassified parts are removed based on the image size, further optimization of the block-shaped water body is realized, and a final water body optimization result is obtained; The specific implementation of Step 4 comprises the following sub-steps: Step 4.1: the image map is classified by using K-means clustering, the coincidence degree of the accurate optimization result and each category is calculated, the category with the highest coincidence degree is determined as the water body category, and a clustering water body result is obtained; Step 4.2: topological analysis is performed on the clustering water body result and the accurate optimization result to obtain intersecting parts and non-intersecting parts, and the intersecting parts of the two are retained; Step 4.3: spectral inspection is performed on the non-intersecting parts, first, the RGB values of the water body pixel points corresponding to the accurate optimization result in the whole image are accumulated respectively, and the average RGB value of the water body in the whole image is calculated, a certain error parameter is set, then the non-intersecting parts are scanned according to the average RGB value, and the non-intersecting parts that do not exceed the error parameter are determined as the water body; then the non-intersecting water body passing the spectral inspection is output to the accurate optimization result to obtain a further optimization result; Step 4.4: According to the size of the image, a certain area threshold is set, and the part of the optimized result obtained in step 4.3 whose water area is less than the above area threshold is removed, to obtain the final water body optimization result. 2.The method according to claim 1, wherein: In step 2, Gaussian filtering is used to denoise the image, and multi-scale Frangi filtering is used to extract the strongest linear features, and then OTSU binary segmentation algorithm is used to obtain the linear feature map. 3.The method of claim 1, wherein the method further comprises: In step 2, the initial extraction result of the water body is combined with the spectral inspection to obtain the linear branch in the image; the RGB values of the water body pixel points corresponding to the initial extraction result in the whole image are added respectively, and the average RGB value of the water body in the whole image is calculated, a certain error parameter is set, and then the average RGB value is scanned according to the linear feature map, and the linear feature that does not exceed the error parameter is considered as the linear branch.

4. A system for extracting water bodies from remote sensing images by combining global features and local intercommunication, characterized in that, The following modules are included: Module 1, for water body extraction from original high-resolution remote sensing image, to automatically obtain the initial extraction result of water body; Module 2, for original high-resolution remote sensing image, using filtering algorithm and segmentation algorithm to obtain linear features in the image, and combining with the initial extraction result of water body to perform spectral inspection to obtain linear branch in the image, and superimposing the linear branch and the initial extraction result to obtain the preliminary optimization result; Module 3, for connecting and supplementing the broken flow in the preliminary optimization result to obtain the accurate optimization map of the linear branch; Module 3 includes the following sub-modules: Module 3.1, for extracting the skeleton and endpoint of the river from the preliminary optimization result; Module 3.2, for obtaining the circumscribed rectangle of the possible broken flow area and the corresponding image subgraph according to the river endpoint; A certain distance threshold is set to traverse the river endpoints, two endpoints are taken as a group, and the endpoints are taken as the diagonal line to extend outward by a predetermined distance, which is taken as the two diagonal of the circumscribed rectangle, to obtain the circumscribed rectangle and the corresponding image subgraph; Module 3.3, for connecting and supplementing the broken flow area by combining local SSIM index and interconnectivity algorithm; For each image subgraph, the SSIM value of each pixel in the subgraph is calculated, a certain index threshold is set, and if the SSIM value of a certain pixel is within the index threshold range of the SSIM values of the two endpoints, it is considered that the pixel is similar in structure to the two endpoints, and the same is applied to obtain the pixels similar in structure to the two endpoints in the subgraph; Using the obtained pixel points, different connected regions in the subgraph are obtained by using connected domain algorithm and are labeled, if the two endpoints are the same label, it is considered that the two endpoints are the same connected domain, and the connected region is supplemented to the preliminary optimization result of module 1 to obtain the accurate optimization result of the water body, and the accurate optimization of the linear branch is realized; Module 4, for extracting the water body part in the high-resolution remote sensing image, and then performing topological inspection and spectral inspection on the water body clustering result and the accurate optimization map obtained in module 3, removing the small area of the wrong part based on the image size, to realize further optimization of the blocky water body, and obtain the final water body optimization result; Module 4 includes the following sub-modules: Module 4.1, for classifying the image by K-means clustering, calculating the coincidence degree of the accurate optimization result with each category, and determining the category with the highest coincidence degree as the water body category to obtain the clustering water body result; Module 4.2, for performing topological analysis on the clustering water body result and the accurate optimization result to obtain intersecting and non-intersecting parts, and reserving the intersecting parts of the two; Module 4.3, for performing spectral examination on the non-intersecting part, first accumulating the RGB values of the water body pixel points corresponding to the accurate optimization result in the whole image respectively, and calculating the average RGB value of the water body in the whole image, setting a certain error parameter, then scanning the non-intersecting part according to the average RGB value, and determining the non-intersecting part not exceeding the error parameter as the water body; then outputting the non-intersecting water body passing the spectral examination to the accurate optimization result to obtain a further optimized result; Module 4.4, for setting a certain area threshold according to the size of the image, and removing the part of the optimized result obtained by module 4.3 with the water body area less than the above area threshold to obtain the final water body optimization result.

5. The system according to claim 4, wherein the system further comprises: In module 2, Gaussian filtering is used to denoise the image, and multi-scale Frangi filtering is used to extract the strongest linear features, and then OTSU binary segmentation algorithm is used to obtain the linear feature map. 6.The system for water body extraction from remote sensing images by combining global features and local interactions according to claim 4, wherein: In module 2, the linear branch stream in the image is obtained by combining the initial water body extraction result with spectral examination; that is, by accumulating the RGB values of the water body pixel points corresponding to the initial extraction result in the whole image respectively, calculating the average RGB value of the water body in the whole image, setting a certain error parameter, and then scanning the linear feature map according to the average RGB value, the linear feature not exceeding the error parameter is considered as the linear branch stream.

Images