Water body extraction method based on integration of multiple artificial intelligence algorithms

Abstract

The invention discloses a water body extraction method based on multi-artificial intelligence algorithm integration. The method comprises the following steps: S1, acquiring and preprocessing a remote sensing image; s2, performing multi-feature extraction on the experiment area image; s3, performing water body extraction on the experiment area image by using a water body index algorithm, a random forest RF algorithm, a support vector machine SVM algorithm, a K-means algorithm and a convolutional neural network algorithm; s4, taking binary images of extraction results of the five algorithms in the S3 as input of a data analysis model, and obtaining real precision corresponding to a water body extraction result; s5, calculating weights corresponding to the five algorithms according to the real precision; calculating the water body probability of each pixel in the binary image by using the weight to obtain a water body probability set; and setting a classification threshold to optimize and integrate the water body probability set to obtain a final water body classification binary image. According to the method, advantages of different water body extraction algorithms are complementary, and the water body extraction precision is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of water body extraction from remote sensing images, and in particular to a water body extraction method based on the integration of multiple artificial intelligence algorithms. Background Art

[0002] As an important component of the ecological cycle, water resources play an important role in the hydrological and biochemical cycles at regional and global scales. However, due to climate change, land use, land cover change and other forms of environmental change, surface water bodies are also constantly changing in time and space. Therefore, it is of great significance to carry out large-scale high-precision extraction research on water bodies and realize timely and accurate monitoring of surface water bodies such as rivers. Traditional extraction methods are mostly analyzed through field measurements and on-site sampling. Although this method can obtain more accurate information on surface water bodies, it consumes a lot of manpower and material resources, and the experimental cycle is long, which cannot meet the actual needs of large-scale and multi-frequency surface water extraction. With the rapid development of satellite remote sensing technology and its fast, all-weather, and high-resolution characteristics, satellite remote sensing technology has become the main means of surface water monitoring.

[0003] Satellite remote sensing technology for surface water extraction is mainly divided into three categories: water index method, machine learning method, and deep learning method.

[0004] in:

[0005] The water index method is to construct a mathematical model based on the different reflectances of water bodies in different visible light bands by establishing band ratio operations, and set corresponding thresholds to perform water body identification. The water index method is widely used because of its simple calculation and high efficiency. Commonly used water indexes include: Normalized Difference Water Index (NDWI), Modified Normalized Water Index (MNDWI), etc. However, when the water index method processes higher-resolution images, the more visible spatial details, the greater the spectral variability of the water body, which leads to poor generalization ability of the water index method.

[0006] Water extraction methods based on machine learning include supervised classification and unsupervised classification. The most commonly used supervised classification water extraction methods are support vector machine (SVM) and random forest (RF). They mainly collect water samples and non-water samples to establish a training data set, and extract water from the identified area based on the training model. Clustering algorithm is a common unsupervised classification water extraction method. Its principle is to divide the pixels in the image into different categories according to specific standards. The similarity of pixels in the same category is as large as possible, and the difference between pixels in different categories is as large as possible.

[0007] The water body extraction method based on deep learning mainly uses convolutional neural network (CNN) to learn texture, spectrum and other high-dimensional features from manually annotated data sets (training samples), and then uses the convolutional neural network to extract water bodies in the area of ​​interest of the image.

[0008] Although the water extraction methods based on machine learning and deep learning have improved the extraction accuracy compared with the water index method, there are still two problems:

[0009] 1. The recognition results of machine learning algorithms such as random forests can easily confuse buildings, bridges, roads and rivers, leading to misjudgment, especially in complex urban scenes.

[0010] 2. The water body extraction method based on deep learning has extremely high requirements for training samples and the entire image needs to be marked, otherwise it will seriously affect the accuracy of water body recognition.

[0011] In addition, the water body index method, machine learning method, and deep learning method are all poorly effective in identifying narrow water bodies, and flow interruptions or missed judgments often occur. Summary of the invention

[0012] In order to solve the problems existing in the prior art, the present invention provides a high-precision, time-continuous and space-seamless water body extraction method based on the integration of multiple artificial intelligence algorithms.

[0013] To this end, the present invention adopts the following technical solutions:

[0014] The water extraction method based on the integration of multiple artificial intelligence algorithms includes the following steps:

[0015] S1. Obtain remote sensing images and preprocess them:

[0016] Acquire a remote sensing image set containing the water body area to be extracted within a specified time range; perform cloud screening and cloud removal on the remote sensing image set to reduce the impact of cloud cover; then perform image cropping on the image set according to the experimental area to obtain an image set of the water body area to be extracted; perform median synthesis on the image set to obtain a final image of the water body area to be extracted;

[0017] S2. Perform multi-feature extraction on the final image:

[0018] Acquire optical band information of each pixel in the final image, and calculate index band information of each pixel in the final image according to the optical band information, wherein the index band information includes NDWI, MNDWI, NDVI, EVI, NDBI and BSI;

[0019] S3, using the optical band information and index band information obtained in S2, five water body extraction algorithms are used to extract water from the final image to obtain binary images of water body extraction results, which are x a 、x b 、x c 、x d and x e ;

[0020] S4, solve the true accuracy of water extraction results based on mathematical uncertainty analysis:

[0021] The binary image x obtained from S3 a 、x b 、x c 、x d and x e Calculate the covariance between two different water body extraction results, and then calculate the accuracy vector y; use the accuracy vector y as the input of the data analysis model to obtain the true accuracy corresponding to the water body extraction result;

[0022] S5, multi-algorithm ensemble learning to obtain the optimal water extraction algorithm, including the following steps:

[0023] S51, according to the true accuracy of the binary image obtained by the data analysis model, calculate the weights w corresponding to the five algorithms in S3 respectively a 、w b 、w c 、w d and w e , the calculation formula is as follows:

[0024]

[0025] Among them, v i is the binary image x obtained by different algorithms iThe true accuracy of , i = a ~ e;

[0026] S52, according to the weights corresponding to the five algorithms, traverse all pixels in the binary image, calculate the water body probability of each pixel in the binary image, and obtain the water body probability set M = {m1, m2...m n}, where the water body probability m of the jth pixel in the binary image is j The calculation formula is as follows:

[0027] m j =w a a j +w b b j +w c c j +w d d j +w e e j (8)

[0028] In the formula, a j , b j 、c j ,d j and e j They are binary images x a , binary image x b , binary image x c , binary image x d and the binary image x e The pixel value of the jth pixel in ; w a 、w b 、w c 、w d and w e They are the weights corresponding to the five algorithms; j = 1, 2, ... n;

[0029] S53, setting the classification threshold to 0; masking the pixels whose water body probability is greater than the classification threshold according to the water body probability set M, and optimizing and integrating them to obtain the final water body classification binary map, thereby completing the accurate extraction of water bodies in the experimental area image.

[0030] The remote sensing image collection mentioned in S1 is the Sentinel-2 satellite image collection.

[0031] The method for extracting water bodies using the water index algorithm is as follows: using the experimental area image as input, performing NDWI ratio calculation on each pixel of the experimental area image using the water index method and setting a threshold, if the NDWI of the pixel is greater than or equal to the threshold, it is a water body, if the NDWI of the pixel is less than the threshold, it is a non-water body, and the water index algorithm outputs a binary image x containing only water bodies and non-water bodies a , where the threshold is set to 0.

[0032] S3 obtains the binary image x a 、x b 、x c 、x d and x e The total number of pixels is equal to n.

[0033] The integrated learning algorithm described in S3 is the random forest algorithm RF, the supervised machine learning algorithm is the support vector machine SVM algorithm, the clustering algorithm is the K-means algorithm, and the supervised deep learning algorithm is the convolutional neural network algorithm CNN.

[0034] The method for extracting water bodies using the random forest algorithm RF and the support vector machine SVM algorithm is as follows: based on the actual situation of the experimental area, each pixel in the experimental area image is divided into five types of land objects: water body, crop, building, vegetation, and bare land, and sample pixels are manually selected by visual interpretation on the experimental area image; the optical band information and index band information of each sample pixel are used as feature variables in the training sets of the RF and SVM algorithms, and are respectively input into the random forest RF algorithm and the support vector machine SVM algorithm for training to obtain training results, and the sample pixels whose land object type is water body are masked, and the random forest algorithm RF and the support vector machine SVM algorithm both obtain binary images with only water bodies and non-water bodies, which are x b and x c .

[0035] The sample pixels include 150 water pixels and 382 non-water pixels in the experimental area image.

[0036] The method of using K-means algorithm to extract water bodies is as follows: all pixels in the experimental area image are divided into K given clusters, and pixels with the ground feature type of water bodies are masked to obtain a binary image x containing only water bodies and non-water bodies. d , where K=5.

[0037] The method of using the convolutional neural network algorithm CNN to extract water bodies is as follows: perform water body mask processing on the Dynamic World (DW) dataset, extract pixels classified as water bodies, and obtain a binary image x containing only water bodies and non-water bodies. e .

[0038] S4 includes the following steps:

[0039] S41, calculating the covariance r between two different water body extraction results based on the water body extraction results obtained by the five algorithms ad , covariance r ae , covariance r de , covariance rbd , covariance r be , covariance r ce and covariance r cd ;

[0040] With covariance r ad For example, the calculation formula is:

[0041]

[0042] In the formula, a j is a binary image x a The pixel value of the jth pixel in , d j is a binary image x d The pixel value of the jth pixel in , is a binary image x a The mean of all pixel values, is a binary image x d The mean of all pixel values, n is the total number of pixels in the binary image;

[0043] S42, perform combined calculation on the covariance obtained in S41 to obtain the precision vector y, and the calculation formula is as follows:

[0044]

[0045] S43, construct a data parsing model based on the precision vector y:

[0046] Av=y (3)

[0047] Among them, v represents the true accuracy of the binary image obtained by the five algorithms, y is the accuracy vector, A is the constant vector matrix connecting the true accuracy v with the accuracy vector y, and the calculation formula of the true accuracy v and the constant vector matrix A is as follows:

[0048] v=[v a v b v c v d v e ] T (4)

[0049]

[0050] In formula (4), v a is a binary image x a True accuracy in data parsing models; v b is a binary image x b True accuracy in data parsing models; v c is a binary image x c True accuracy in data parsing models; vd is a binary image x d True accuracy in data parsing models; v e is a binary image x e True accuracy in data parsing models;

[0051] S44, calculate the least squares solution of equation (3) to obtain the true accuracy v, which is calculated as follows:

[0052] v=(A T A) -1 A T y (6).

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] 1. Based on the theory of mathematical uncertainty analysis, the present invention realizes the accuracy evaluation and integrated learning of multiple intelligent algorithms that are independent of the "ground truth value", overcomes the problem of misjudgment or omission of water bodies by a single water body extraction algorithm, enables the complementary advantages of different water body extraction algorithms, and improves the accuracy of water body recognition in complex backgrounds.

[0055] 2. The water body extraction method of the present invention addresses the problems of incomplete water body boundary extraction and discontinuous narrow water bodies in some water body extraction algorithms. By fusing different water body extraction algorithms to determine the corresponding weight ratios, the slender water bodies can be accurately extracted, and the high precision of the water body boundaries can be maintained.

[0056] 3. The water extraction method of the present invention effectively avoids the influence of shadows from buildings, bridges, roads and mountains, and this advantage is particularly evident in urban areas.

[0057] 4. The water extraction method of the present invention can achieve high-precision, time-continuous and spatially seamless water extraction, becoming an important means of quickly and accurately monitoring the scope of flood inundation. It is of great significance for understanding the evolution of flood disasters, and provides data support and guarantee for disaster assessment, post-disaster reconstruction and water resources management. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 A flow chart of a water extraction method according to an embodiment of the present invention;

[0059] Figure 2 The results of extracting narrow water bodies in the same area by the present invention and other algorithms are shown in the figure, where:

[0060] Figure (a) is a Gaofen-6 (GF6) image;

[0061] Figure (b) is the true value image used for verification;

[0062] Figure (c) shows the image of the experimental area;

[0063] Figure (d) shows the water extraction result of the convolutional neural network algorithm;

[0064] Figure (e) shows the water body extraction result of the water body index algorithm;

[0065] Figure (f) shows the water extraction result of the SVM algorithm;

[0066] Figure (g) shows the water extraction result of K-means algorithm;

[0067] Figure (h) shows the water extraction result of RF algorithm;

[0068] Figure (i) shows the water extraction result of the water extraction method of the present invention. DETAILED DESCRIPTION

[0069] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings and embodiments.

[0070] Example

[0071] like Figure 1 As shown, a water body extraction method based on the integration of multiple artificial intelligence algorithms specifically includes the following steps:

[0072] S1, obtain remote sensing images and preprocess them:

[0073] Based on the Google Earth Engine (GEE) cloud platform, a Sentinel-2 satellite image collection within a specified time range including the experimental area is obtained; cloud amount screening and cloud removal processing are performed on the Sentinel-2 satellite image collection to reduce the impact of cloud cover; then the Sentinel-2 satellite image collection is cropped according to the experimental area to obtain an image collection of the experimental area; median synthesis is performed on the image collection of the experimental area to obtain the final domain image of the experimental area.

[0074] Among them, in this embodiment, the Jiuxuanzha area in Jinghai District, Tianjin City is taken as the experimental area, the pixel size of the image in the experimental area is 10m×10m, and the Sentinel-2 satellite image collection in October 2022 is selected for preprocessing.

[0075] S2, extract multiple features from the final image of the experimental area:

[0076] The optical band information of each pixel in the final image of the experimental area is obtained, and the index band information of each pixel in the final image of the experimental area is calculated according to the optical band information and the calculation method provided in Table 1. The index band information includes NDWI, MNDWI, NDVI, EVI, NDBI and BSI. The calculation formula of the index band is shown in Table 1.

[0077] In Table 1, Red represents the red band (B4), Green represents the green band (B3), Blue represents the blue band (B2), NIR represents the near infrared band (B8), and SWIR1 represents the mid-infrared band (B11).

[0078] Table 1 Characteristic variables

[0079]

[0080] S3, respectively uses the water index algorithm, random forest RF algorithm, support vector machine SVM algorithm, K-means algorithm and convolutional neural network algorithm to extract water bodies from the experimental area images, as follows:

[0081] Water index algorithm: Take the experimental area image as input, use the water index method to perform NDWI ratio calculation on each pixel of the experimental area image and set a threshold. If the NDWI of the pixel is greater than or equal to the threshold, it is a water body. If the NDWI of the pixel is less than the threshold, it is a non-water body. The water index algorithm outputs a binary image x containing only water bodies and non-water bodies. a In this embodiment, the threshold is set to 0.

[0082] Random forest algorithm (RF) and support vector machine (SVM) algorithm: Combined with the actual situation of the experimental area, each pixel in the experimental area image is divided into five types of land features: water bodies, crops, buildings, vegetation, and bare land. Visually interpret and manually select sample pixels on the experimental area image; the optical band information and index band information of each sample pixel are used as feature variables in the training set of the RF and SVM algorithms, and are input into the random forest RF algorithm and the support vector machine (SVM) algorithm for training, respectively, to obtain the training results, and mask the sample pixels whose land feature type is water bodies. Both the random forest algorithm (RF) and the support vector machine (SVM) algorithm obtain binary images with only water bodies and non-water bodies, which are x and y, respectively. b and x c In this embodiment, the sample pixels include 150 water pixels and 382 non-water pixels in the experimental area image.

[0083] K-means algorithm: All pixels in the experimental area image are divided into K given clusters. In combination with the land use / cover characteristics of the study area, the classification cluster K value is set to 5, and the pixels with the land object type of water body are masked to obtain a binary image x containing only water body and non-water body. d .

[0084] Convolutional Neural Network (CNN): The Dynamic World (DW) dataset (a near real-time LULC product developed by Google based on Sentinel-2 data with a resolution of 10 meters, which uses a fully convolutional neural network (FCNN) to classify Sentine-2 images containing the experimental area into nine categories: water bodies, trees, grass, flooded vegetation, crops, bushes / shrubs, built-up areas, bare land, snow / ice) is processed with water body masks, pixels classified as water bodies are extracted, and a binary image x containing only water bodies and non-water bodies is obtained. a .

[0085] Among them, the binary images obtained by the above five algorithms are x a 、x b 、x c 、x d and x e , their total number of pixels is equal, both are n.

[0086] S4, solve the true accuracy of water extraction results based on mathematical uncertainty analysis:

[0087] The binary images x obtained by the five algorithms in S3 a 、x b 、x c 、x d and x e As the input of the classification triple correction fusion method (data analysis model) based on mathematical uncertainty analysis, the true accuracy corresponding to the water body extraction result is obtained. The specific steps are as follows:

[0088] S41, based on the water body extraction results obtained by the five algorithms, calculate the covariance r between two different water body extraction results (binary images) ad , covariance r ae , covariance r de , covariance r bd , covariance r be , covariance r ce and covariance r cd .

[0089] With covariance r ad For example, the calculation formula is:

[0090]

[0091] In the formula, a j is a binary image x a The pixel value of the jth pixel in , d j is a binary image x d The pixel value of the jth pixel in , is a binary image xa The mean of all pixel values, is a binary image x d is the mean of all pixel values, and n is the total number of pixels in the binary image.

[0092] S42, perform combined calculation on the covariance obtained in S41 to obtain the precision vector y, and the calculation formula is as follows:

[0093]

[0094] S43, construct a data parsing model based on the precision vector y:

[0095] Av=y (3)

[0096] Among them, v represents the true accuracy of the binary image obtained by the five algorithms, y is the accuracy vector, A is the constant vector matrix connecting the true accuracy v with the accuracy vector y, and the calculation formula of the true accuracy v and the constant vector matrix A is as follows:

[0097] v=[v a v b v c v d v e ] T (4)

[0098]

[0099] In formula (4), v a is a binary image x a True accuracy in data parsing models; v b is a binary image x b True accuracy in data parsing models; v c is a binary image x c True accuracy in data parsing models; v d is a binary image x d True accuracy in data parsing models; v e is a binary image x e True accuracy in data parsing models.

[0100] S44, calculate the least square solution of equation (3) to obtain the true accuracy v. The true accuracy v reflects the accuracy of the water extraction results of the five algorithms, and its calculation formula is as follows:

[0101] v=(A T A) -1 A T y (6)

[0102] S5, multi-algorithm ensemble learning to obtain the optimal water extraction algorithm:

[0103] According to the real accuracy v of the binary image obtained by the data analysis model, the weights corresponding to the five algorithms are calculated; all pixels in the binary image are traversed, and the water body probability of each pixel in the binary image is calculated using the weight to obtain the water body probability set M; the classification threshold is set, and the water body probability set M is optimized and integrated to obtain the final water body classification binary map, completing the accurate extraction of water bodies in the experimental area image. The specific steps are as follows:

[0104] S51, according to the true accuracy v of the binary image obtained by the data analysis model, the weights w corresponding to the five algorithms are calculated respectively a 、w b 、w c 、w s and w e , the calculation formula is as follows:

[0105]

[0106] Among them, v i is the binary image x obtained by different algorithms i The true accuracy of , i = a ~ e.

[0107] S52, according to the weights corresponding to the five algorithms, traverse all pixels in the binary image, calculate the water body probability of each pixel in the binary image, and obtain the water body probability set M = {m1, m2...n n}, where the water body probability m of the jth pixel in the binary image is j The calculation formula is as follows:

[0108] m j =w a a j +w b b j +w c c j +w d d j +w e e j (8)

[0109] In the formula, a j , b j 、c j ,d j and e j They are binary images x a , binary image x b , binary image x c , binary image x d and the binary image x e The pixel value of the jth pixel in ; w a 、w b、w c 、w d and w e They are the weights corresponding to the five algorithms; j=1,2,…n.

[0110] S53, setting the classification threshold to 0, that is, the threshold for water body discrimination is 0; according to the water body probability set M, masking is performed on pixels whose water body probability is greater than the classification threshold, and optimization and integration are performed to obtain the final water body classification binary map, thereby completing the accurate extraction of water bodies in the experimental area image.

[0111] In order to verify the accuracy of this method, the accuracy of the water body extraction results (water body classification binary map) of this method is evaluated from different angles as follows:

[0112] (1) Evaluate the accuracy of water extraction results from the perspective of visual interpretation:

[0113] The water extraction results of the five algorithms in the embodiment are as follows Figure 2 As shown in (d) to (h) in the figure, it can be seen that the NDWI algorithm is not good enough to extract water bodies in small areas and long rivers, and there is a serious phenomenon of missing extraction. The SVM algorithm and RF algorithm mistakenly extract a large number of buildings, roads, bridges, and farmlands as water bodies. The K-means algorithm mistakenly extracts roads as water bodies. The boundaries of the water body extraction results of the CNN algorithm are fuzzy.

[0114] The water extraction results obtained according to the method of the present invention are as follows Figure 2 As shown in (i), it can be seen from the figure that the water body extraction method of the present invention has an ideal extraction result for the edge of the river bank. Compared with the water body extraction results of the five algorithms, the water body extraction result of the water body extraction method of the present invention is smoother and more accurate. The water body extraction method of the present invention can not only distinguish between water bodies and non-water body areas well, but also can more accurately extract small targets, small-area rivers and slender rivers, and distinguish water bodies from shadows.

[0115] (2) Evaluate the accuracy of water extraction results from the perspective of evaluation indicators:

[0116] Calculate the evaluation index: Get the Gaofen-6 (GF6) image that includes the experimental area and was imaged on October 7, 2022, such as Figure 2 As shown in (a), the pixel size of the GF6 image is 2m×2m. A water mask is manually created, and the water pixel value is marked as 1 and the non-water pixel value is marked as 0. The water mask will be used as the true value image for verification, as shown in Figure 2As shown in (b), water sample points and non-water sample points are selected in the true value image, and a confusion matrix is ​​constructed based on the selected water sample points and non-water sample points; the confusion matrix is ​​used to calculate evaluation indicators, which include Kappa coefficient (KC), F1 score and balanced accuracy.

[0117] According to the evaluation indexes calculated, the accuracy of the results of the RF algorithm and the water extraction method of the present invention is better, the evaluation index of the water extraction result of the water index algorithm is the lowest, and the evaluation indexes of the water extraction results of the K-means algorithm and the SVM algorithm are the same. The water extraction result of the water extraction method of the present invention has the highest F1 score, both exceeding 0.95; the F1 score of the water extraction result of the method of the present invention is 4.49% and 15.14% higher than that of the RF method and the CNN method, respectively; the Kappa coefficient of the water extraction result of the method of the present invention is improved by 11.90% and 30.56% compared with the RF algorithm and the convolutional neural network algorithm, respectively; the Balanced accuracy of the water extraction result of the method of the present invention is improved by 5.43% and 12.79%, respectively.

[0118] In summary, the water body extraction method of the present invention has the highest accuracy in extracting water bodies from remote sensing images, and the water body extraction results perform best.

[0119] The water body extraction method of the present invention is applicable to water body extraction data sets of any region and any year and are continuous in time and space.

Claims

1. A water extraction method based on the integration of multiple artificial intelligence algorithms, characterized by The following steps are involved: S1. Obtain remote sensing images and preprocess them: Acquire a remote sensing image set containing the water body area to be extracted within a specified time range; perform cloud screening and cloud removal on the remote sensing image set to reduce the impact of cloud cover; then perform image cropping on the image set according to the experimental area to obtain an image set of the water body area to be extracted; perform median synthesis on the image set to obtain a final image of the water body area to be extracted; S2. Perform multi-feature extraction on the final image: Acquire optical band information of each pixel in the final image, and calculate index band information of each pixel in the final image according to the optical band information, wherein the index band information includes NDWI, MNDWI, NDVI, EVI, NDBI and BSI; S3, using the optical band information and index band information obtained in S2, five water body extraction algorithms are used to extract water from the final image to obtain binary images of water body extraction results, which are x a 、x b 、x c 、x d and x e ; S4, solve the true accuracy of water extraction results based on mathematical uncertainty analysis: The binary image x obtained from S3 a 、x b 、x c 、x d and x e Calculate the covariance between two different water body extraction results, and then calculate the accuracy vector y; use the accuracy vector y as the input of the data analysis model to obtain the true accuracy corresponding to the water body extraction result; S5, multi-algorithm ensemble learning to obtain the optimal water extraction algorithm, including the following steps: S51, according to the true accuracy of the binary image obtained by the data analysis model, calculate the weights w corresponding to the five algorithms in S3 respectively a 、w b 、w c 、w d and w e , the calculation formula is as follows: Among them, v i is the binary image x obtained by different algorithms i The true accuracy of , i = a ~ e; S52, according to the weights corresponding to the five algorithms, traverse all pixels in the binary image, calculate the water body probability of each pixel in the binary image, and obtain the water body probability set M = {m1, m2...m n }, where the water body probability m of the jth pixel in the binary image is j The calculation formula is as follows: m j =w a a j +w b b j +w c c j +w d d j +w e e j (8) In the formula, a j 、b j 、c j ,d j and e j They are binary images x a , binary image x b , binary image x c , binary image x d and the binary image x e The pixel value of the jth pixel in ; w a 、w b 、w c 、w d and w e They are the weights corresponding to the five algorithms; j = 1, 2, ... n; S53, setting the classification threshold to 0; masking the pixels whose water body probability is greater than the classification threshold according to the water body probability set M, and optimizing and integrating them to obtain the final water body classification binary map, thereby completing the accurate extraction of water bodies in the experimental area image.

2. The water extraction method according to claim 1, characterized in that: The remote sensing image collection mentioned in S1 is the Sentinel-2 satellite image collection.

3. The water extraction method according to claim 1, characterized in that: The method for extracting water bodies using the water index algorithm is as follows: using the experimental area image as input, performing NDWI ratio calculation on each pixel of the experimental area image using the water index method and setting a threshold, if the NDWI of the pixel is greater than or equal to the threshold, it is a water body, if the NDWI of the pixel is less than the threshold, it is a non-water body, and the water index algorithm outputs a binary image x containing only water bodies and non-water bodies a , where the threshold is set to 0.

4. The water extraction method according to claim 1, characterized in that: S3 obtains the binary image x a 、x b 、x c 、x d and x e The total number of pixels is equal to n.

5. The water extraction method according to claim 1, characterized in that: The five water body extraction algorithms described in S3 are water body index method, support vector machine SVM algorithm, random forest algorithm RF, K-means algorithm and convolutional neural network algorithm CNN.

6. The water extraction method according to claim 5, characterized in that: The method for extracting water bodies using the random forest algorithm RF and the support vector machine SVM algorithm is as follows: based on the actual situation of the experimental area, each pixel in the experimental area image is divided into five types of land objects: water body, crop, building, vegetation, and bare land, and sample pixels are manually selected by visual interpretation on the experimental area image; the optical band information and index band information of each sample pixel are used as feature variables in the training sets of the RF and SVM algorithms, and are respectively input into the random forest RF algorithm and the support vector machine SVM algorithm for training to obtain training results, and the sample pixels whose land object type is water body are masked, and the random forest algorithm RF and the support vector machine SVM algorithm both obtain binary images with only water bodies and non-water bodies, which are x b and x c .

7. The water extraction method according to claim 6, characterized in that: The sample pixels include 150 water pixels and 382 non-water pixels in the experimental area image.

8. The water extraction method according to claim 5, characterized in that: The method of using K-means algorithm to extract water bodies is as follows: all pixels in the experimental area image are divided into K given clusters, and pixels with the ground feature type of water bodies are masked to obtain a binary image x containing only water bodies and non-water bodies. d , where K=5.

9. The water extraction method according to claim 5, characterized in that: The method of using the convolutional neural network algorithm CNN to extract water bodies is as follows: perform water body mask processing on the Dynamic World (DW) dataset, extract pixels classified as water bodies, and obtain a binary image x containing only water bodies and non-water bodies. e .

10. The water extraction method according to claim 1, characterized in that: S4 includes the following steps: S41, calculating the covariance r between two different water body extraction results based on the water body extraction results obtained by the five algorithms ad , covariance r ae , covariance r de , covariance r bd , covariance r be , covariance r ce and covariance r cd ; With covariance r ad For example, the calculation formula is: In the formula, a j is a binary image x a The pixel value of the jth pixel in , d j is a binary image x d The pixel value of the jth pixel in , is a binary image x a The mean of all pixel values, is a binary image x d The mean of all pixel values, n is the total number of pixels in the binary image; S42, perform combined calculation on the covariance obtained in S41 to obtain the precision vector y, and the calculation formula is as follows: S43, construct a data parsing model based on the precision vector y: Av=y (3) Among them, v represents the true accuracy of the binary image obtained by the five algorithms, y is the accuracy vector, A is the constant vector matrix connecting the true accuracy v with the accuracy vector y, and the calculation formula of the true accuracy v and the constant vector matrix A is as follows: v=[v a v b v c v d v e ] T (4) In formula (4), v a is a binary image x a True accuracy in data parsing models; v b is a binary image x b True accuracy in data parsing models; v c is a binary image x c True accuracy in data parsing models; v d is a binary image x d True accuracy in data parsing models; v e is a binary image x e True accuracy in data parsing models; S44, calculate the least squares solution of equation (3) to obtain the true accuracy v, which is calculated as follows: v=(A T TO) -1 TO T and (6)。

Images