Method, device and storage medium for constructing remote sensing image water body sample set

Abstract

This application relates to the technical field of water body samples and discloses a method, device, and storage medium for constructing a remote sensing image water body sample set. The method includes obtaining multiple original satellite remote sensing images and preprocessing each original satellite remote sensing image to obtain an initial water body sample set; performing noise cleaning on the initial water body sample set based on a cross-validation and confidence iteration learning method to obtain a water body denoised sample set; and classifying the water body denoised sample set based on water body samples in RGB mode and HSV color space to obtain a remote sensing image water body sample set. This application can obtain a high-quality, refined water body sample set, solving the problems of noise, outliers, and insufficiently refined sample classification that exist in large-scale water body sample sets obtained based on traditional solutions.

Description

Technical Field

[0001] The present application relates to the technical field of water body samples, and in particular to a method, device and storage medium for constructing a remote sensing image water body sample set. Background Art

[0002] Information extraction through remote sensing images is an important technical means of monitoring water resources; that is, deep learning is used to intelligently interpret the water body information of remote sensing images, namely image interpretation (such as water body classification, post-classification processing, accuracy verification, etc.), so as to deeply mine the water resource element information in remote sensing images, thereby obtaining water body distribution information in a timely and accurate manner, which is of great significance to the monitoring and management of water resources.

[0003] Deep learning is essentially a supervised learning method, which requires a large amount of sample data to learn the essential characteristics of the target. The accuracy of the prediction results of the deep learning network model is most directly related to the quality of the sample data set. Therefore, good sample quality is an important guarantee and support for the accuracy of the deep learning model. High-quality samples can improve the accuracy and versatility of the deep learning model. Summary of the Invention

[0004] In view of this, embodiments of the present application provide a method, device, and storage medium for constructing a remote sensing image water sample set to improve the quality of water samples.

[0005] In a first aspect, an embodiment of the present application provides a method for constructing a remote sensing image water body sample set, comprising:

[0006] Acquire multiple original satellite remote sensing images, and preprocess each of the original satellite remote sensing images to obtain an initial water body sample set;

[0007] The water body initial sample set is subjected to noise cleaning based on a cross-validation and confidence iteration learning method to obtain a water body denoised sample set;

[0008] Based on the water body samples in RGB mode and HSV color space, the water body denoising sample set is classified to obtain a remote sensing image water body sample set.

[0009] In some embodiments, preprocessing each of the original satellite remote sensing images to obtain an initial water sample set includes:

[0010] Gridding the original satellite remote sensing image based on geographic coordinates to obtain hierarchical grids corresponding to different slice sizes;

[0011] Performing grid positioning and cropping processing on the original satellite remote sensing image with a hierarchical grid based on image resolution and water body spot vectors to extract a plurality of water body spot sample slice images;

[0012] Performing position marking processing on each of the water body spot sample slice images to obtain a corresponding water body label image;

[0013] A water body initial sample set is constructed based on all the water body spot sample slice images and the water body label images; the water body initial sample set includes multiple water body samples, and each water body sample includes a water body spot sample slice image and a corresponding water body label image.

[0014] In some embodiments, the gridding of the original satellite remote sensing image based on geographic coordinates includes:

[0015] Gridding the original satellite remote sensing image, and grading all grid slices after the gridding based on image resolution and slice size;

[0016] The row and column numbers of each level of the slicing grid are determined according to the latitude and longitude coordinates and the slicing grid level.

[0017] In some embodiments, the method based on cross-validation and confidence iteration learning performs noise cleaning on the initial water sample set to obtain a water denoised sample set, including:

[0018] Dividing the water body samples in the initial water body sample set into multiple parts, selecting one of them as a validation subset and the remaining parts as training subsets, so as to determine the first round of training subsets and validation subsets;

[0019] The semantic segmentation model is trained using the first-round training subset, and the first-round validation subset is screened according to the semantic segmentation model to obtain an updated subset;

[0020] The updated subset is used as a training subset for the next round, and a subset is selected from the training subset of the first round as a validation subset for the next round, so as to determine the training subset and validation subset for the next round and perform iterative training until each water sample in the initial water sample set has been screened once;

[0021] All the updated subsets after screening are used as water body denoising sample sets.

[0022] In some embodiments, screening the first-round validation subset according to the semantic segmentation model to obtain an updated subset includes:

[0023] Inputting the validation subset into the semantic segmentation model to obtain the pixel probability of each pixel in each water body sample in the validation subset;

[0024] Calculating the confidence of each water body spot sample in the water body label image according to the pixel probability of each pixel point in the water body sample;

[0025] Calculating the confidence of the water body label image according to the confidence of all the water body spot samples in the water body label image;

[0026] The confidence level of each water body label image is compared with a set threshold, and water body samples corresponding to the confidence level being greater than the set threshold are taken as the updated subset.

[0027] In some embodiments, the water body samples based on the RGB mode and HSV color space are classified into the water body denoising sample set to obtain the remote sensing image water body sample set, including:

[0028] Randomly sampling all water sample pixels in the water denoising sample set;

[0029] The mode of the RGB values ​​of all sampled water sample pixels is calculated to obtain the RGB value of the water sample;

[0030] The RGB values ​​are converted into HSV values, and the water body denoising sample set is classified based on the color classification rule of the HSV values ​​to obtain a remote sensing image water body sample set.

[0031] In some embodiments, the randomly sampling all pixel points in the water body denoising sample set includes:

[0032] Calculate the proportion of each water body spot in the water body spot sample slice image according to each water body label image in the water body denoising sample set;

[0033] Determine the sampling number of water body sample pixel points in the corresponding water body spot sample slice image according to the proportion value;

[0034] When randomly sampling the water body sample pixel points in the water body patch sample slice image, the sampling number of small water body samples and large water body samples in the water body denoising sample set satisfies the set relationship; wherein, the water body samples with a proportion value less than 1% are defined as small water body samples, and the water body samples with a proportion value greater than 1% are defined as large water body samples.

[0035] In a second aspect, an embodiment of the present application provides a device for constructing a remote sensing image water sample set, comprising:

[0036] A preprocessing module is used to obtain a plurality of original satellite remote sensing images and preprocess each of the original satellite remote sensing images to obtain an initial water body sample set;

[0037] A denoising module is used to perform noise cleaning on the initial water sample set based on a cross-validation and confidence iteration learning method to obtain a water denoising sample set;

[0038] The classification module is used to classify the water body denoising sample set based on the water body samples in the RGB mode and HSV color space to obtain the remote sensing image water body sample set.

[0039] In a third aspect, an embodiment of the present application provides a computer terminal comprising a processor and a memory, wherein the memory stores a computer program, and the processor is configured to execute the computer program to implement the above-mentioned method for constructing a remote sensing image water body sample set.

[0040] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium storing a computer program. When the computer program is executed on a processor, the method for constructing a remote sensing image water body sample set is implemented.

[0041] The embodiments of the present application have the following beneficial effects:

[0042] The method for constructing a remote sensing image water body sample set in this application first performs noise cleaning on the preprocessed initial water body sample set based on the cross-validation and confidence iterative learning methods, and then automatically classifies it based on the RGB mode and HSV color space, thereby obtaining a high-quality and refined water body sample set to support the task of intelligent water body interpretation. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without creative work.

[0044] Figure 1 A flow chart showing a method for constructing a remote sensing image water body sample set according to an embodiment of the present application is shown;

[0045] Figure 2 A schematic diagram of an original satellite remote sensing image according to an embodiment of the present application is shown;

[0046] Figure 3 A flow chart showing an embodiment of the present application of preprocessing each original satellite remote sensing image to obtain an initial water sample set;

[0047] Figure 4 A schematic diagram of a process for obtaining a water body slice image according to an embodiment of the present application is shown;

[0048] Figure 5 A schematic diagram of grid slicing of satellite remote sensing images according to an embodiment of the present application is shown;

[0049] Figure 6 A schematic diagram of a water sample label according to an embodiment of the present application is shown;

[0050] Figure 7 A schematic diagram of a process for obtaining a water body denoising sample set according to an embodiment of the present application is shown;

[0051] Figure 8 A schematic diagram of a process for screening a first-round verification subset according to a semantic segmentation model to obtain an updated subset is shown in an embodiment of the present application;

[0052] Figure 9 A schematic diagram of a process for classifying a water body denoising sample set to obtain a remote sensing image water body sample set is shown in an embodiment of the present application;

[0053] Figure 10 A schematic diagram of a process for randomly sampling all water sample pixels in a water denoising sample set according to an embodiment of the present application is shown;

[0054] Figure 11 A structural schematic diagram of a device for constructing a remote sensing image water body sample set according to an embodiment of the present application is shown. DETAILED DESCRIPTION

[0055] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments.

[0056] The components of the embodiments of the present application generally described and illustrated in the drawings herein may be arranged and designed in a variety of different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed application, but rather merely represents selected embodiments of the present application. All other embodiments obtained by those skilled in the art based on the embodiments of the present application without creative effort are within the scope of protection of the present application.

[0057] Hereinafter, the terms "including", "having" and their cognates used in various embodiments of the present application are intended only to indicate specific features, numbers, steps, operations, elements, components or combinations of the aforementioned items, and should not be understood as excluding the existence of one or more other features, numbers, steps, operations, elements, components or combinations of the aforementioned items or adding the possibility of one or more features, numbers, steps, operations, elements, components or combinations of the aforementioned items. In addition, the terms "first", "second", "third" and the like are only used to distinguish descriptions and should not be understood as indicating or implying relative importance.

[0058] Unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by those skilled in the art to which the various embodiments of the present application belong. The terms (such as those defined in generally used dictionaries) will be interpreted as having the same meaning as in the context of the relevant technical field and will not be interpreted as having an idealized meaning or an overly formal meaning unless clearly defined in the various embodiments of the present application.

[0059] The following describes some embodiments of the present application in detail with reference to the accompanying drawings. In the absence of conflict, the following embodiments and features in the embodiments may be combined with each other.

[0060] Remote sensing imagery water datasets are mostly derived from satellite data such as Landsat and Sentinel. Their spatial resolution is low, and they lack detailed delineation of water morphology and color, hindering high-precision intelligent water resource extraction. Furthermore, current large-scale water sample acquisition often relies on automated or semi-automated collection of multiple remote sensing images and historical patches. The accuracy of historical patches, the temporal alignment between images, and image quality all directly impact sample quality, leading to noise, outliers, and duplicate data in the sample data, making it unsuitable for direct application in deep learning training. However, traditional approaches for automatic cleaning and denoising of sample sets, such as those for water semantic segmentation, primarily fall into two categories: rule-based or statistical methods, and machine learning or deep learning-based methods. These rule-based and statistical methods often require pre-set rules or statistical thresholds, relying heavily on manual experience and expertise, significantly impacting cleaning performance. Selecting appropriate parameters is often challenging, making them unsuitable for complex and variable sample data and prone to missing or misclassifying noisy data. With the continuous advancement of technologies such as machine learning and deep learning, it is gradually possible to automatically identify and process noise, outliers and duplicate data in data sets by automatically learning data features. However, it often requires a large amount of training data and is sensitive to the selection of hyperparameters and model structures. Therefore, after obtaining the initial water sample set by pre-processing satellite remote sensing images, this application further proposes a random noise cleaning method for water samples based on cross-validation and confidence iteration learning. On the basis of semantic segmentation deep learning iterative training, an empirical statistical threshold intervention of confidence is also added to combine deep learning with traditional methods to achieve the construction of a high-quality water sample set. Therefore, this application proposes a remote sensing image sample set construction scheme for intelligent interpretation of water bodies, including satellite remote sensing gridding, patch vectors, water body location mark sample label production, and noise cleaning based on cross-validation and confidence iteration learning methods; in addition, it also forms a high-quality and refined water sample set based on steps such as automatic sample classification based on RGB mode and HSV transformation to support intelligent interpretation tasks such as water body classification, post-classification processing, and accuracy verification.

[0061] The following describes the method for constructing the remote sensing image water body sample set with reference to some specific embodiments.

[0062] Figure 1 A flow chart of a method for constructing a remote sensing image water body sample set according to an embodiment of the present application is shown. Exemplarily, the method for constructing a remote sensing image water body sample set includes the following steps:

[0063] S100: Acquire multiple original satellite remote sensing images, and pre-process each original satellite remote sensing image to obtain an initial water body sample set.

[0064] This application constructs water samples mainly using Figure 2 The original satellite remote sensing image shown has a spatial resolution of approximately 2 meters. Exemplarily, preprocessing of the original satellite remote sensing image includes grid slicing, cropping and extracting patches, and location tagging based on water patches. The resulting initial water sample set includes each water sample, each of which includes a water slice image and a water sample label.

[0065] In one embodiment, Figure 3 As shown in the figure, each original satellite remote sensing image is preprocessed to obtain an initial water sample set, including:

[0066] S110, gridding the original satellite remote sensing image based on geographic coordinates to obtain hierarchical grids corresponding to different slice sizes.

[0067] like Figure 4 As shown in the figure, the above-mentioned gridding processing of the original satellite remote sensing image based on geographic coordinates is performed to obtain hierarchical grids corresponding to different slice sizes, including:

[0068] S111, gridding the original satellite remote sensing image, and grading all the gridded slices based on the image resolution and slice size.

[0069] In this embodiment, the satellite remote sensing image is gridded to obtain the following Figure 5 As shown in the figure, the grids of satellite remote sensing images are graded based on geographic coordinates. The first-level grid is a vector range of 180°×180° in longitude and latitude. The second-level grid is one-quarter the size of level 1, with a longitude and latitude range of 90°×90°. And so on. The longitude and latitude range of each level can be determined.

[0070] The calculation formula for the longitude and latitude range of grid level N is:

[0071] Lon N =360° / 2 N

[0072] Lat N =360° / 2 N ;

[0073] Where, Lon is longitude, Lon N is the longitude of the Nth level grid, Lat is the latitude, Lat N is the latitude of the Nth level grid.

[0074] S112, determining the row and column numbers of each level of the slicing grid according to the latitude and longitude coordinates and the slicing grid level.

[0075] In one embodiment, if 180° west longitude and 90° north latitude are used as the origin of the coordinate system, the row and column numbers of the slices of the Nth level slice grid are calculated based on the longitude and latitude coordinates Lon, Lat, and the following formula can be used:

[0076] R N =(90°-Lat)×2 N / 360°

[0077] C N =(180°+Lon)×2 N / 360°.

[0078] It can be understood that the above formula can be used to calculate the row number R of the N-level slice grid where the point is located based on the latitude and longitude coordinates of a point on the satellite image. N 、Column number C N .

[0079] S120 , performing grid positioning and cropping processing on the original satellite remote sensing image with hierarchical grids according to image resolution and water body spot vectors, and extracting a plurality of water body spot sample slice images.

[0080] First, the corresponding level of slice size is selected according to the image resolution and water body spot vector, and the slice grid position is located. Slicing processing is performed to obtain multiple water body slice images.

[0081] In this embodiment, the corresponding grid level and slice size are selected for sample construction based on the corresponding latitude and longitude resolution, and the principle of being as close as possible to the original image resolution. The sample slice size is generally 2K×2K pixels. Based on the current GPU hardware conditions and the requirements of intelligent water interpretation, k is generally ≥ 8, and 512×512 pixels (corresponding to a 15-level grid) or 1024×1024 pixels (corresponding to a 14-level grid) can be selected for implementation.

[0082] In one embodiment, the relationship between the latitude and longitude resolution Res, the slice size Tsize, and the grid level N can be described as follows:

[0083] Res=360° / (Tsize×2 N )

[0084] In this embodiment, the corresponding sample slice images are cropped based on historical data such as geographic monitoring and land change surveys, or manually drawn water body patch vectors. First, the latitude and longitude ranges of each water body patch element in the vector layer are read. Based on the row and column numbers of the grid slices, the grids within the latitude and longitude range corresponding to the water body patch vectors are calculated. Overlapping grids of adjacent water body patches are deduplicated, retaining a single target grid. The water body slice image is then cropped based on the target grid, resulting in the extraction of multiple sample water body patch slice images.

[0085] It should be noted here that: for the same water body patch, the original satellite remote sensing images of multiple seasons and phases can be cropped to obtain different water body samples corresponding to the original satellite remote sensing images of the same water body in different seasons; and there will be multiple grid slices after gridding. For generally larger water body patches, a water body patch sample slice image may include the original satellite remote sensing image corresponding to one slice grid, or may include the original satellite remote sensing images corresponding to multiple grid slices.

[0086] S130 , performing position marking processing on each water body patch sample slice image to obtain a water body label image.

[0087] For example, based on the water body spot vector and the corresponding geographic grid, a corresponding binary image sample label with a geographic location mark is produced, that is, a water body sample label is obtained, such as Figure 6 The geographical location has an absolute location and a relative location. The marking according to the longitude and latitude is to mark the absolute location of the geographical location.

[0088] S140: Construct an initial water body sample set based on the water body patch sample slice image and the water body label image.

[0089] In this embodiment, the water body initial sample set includes multiple water body samples, and each water body sample includes a water body patch sample slice image and a corresponding water body label image.

[0090] It can be understood that the initial water body sample set is constructed by gridding and slicing the satellite images based on geographic coordinates (i.e., longitude and latitude information), and slicing and labeling the multi-phase satellite remote sensing images corresponding to historical patches and grids.

[0091] S200, based on the cross-validation and confidence iteration learning method, the noise of the initial water sample set is cleaned to obtain the water denoising sample set.

[0092] For example, the process of performing noise cleaning on the initial water sample set based on the cross-validation and confidence iteration learning method to obtain the water denoising sample set is as follows: Figure 7 Shown, including:

[0093] S210, dividing the water body samples in the initial water body sample set into multiple parts; and selecting one of them as a validation subset and the remaining parts as training subsets to determine the first round of training subsets and validation subsets.

[0094] In one embodiment, the initial water sample set is divided into K parts. During each training run, one of the parts is selected as the validation subset, and the remaining K-1 parts are selected as the training subsets. For example, suppose the initial water sample set is evenly divided into four parts, labeled A, B, C, and D. During the first round of training, D is used as the validation subset, and the remaining A, B, and C are used as the training subsets.

[0095] S220: Use the first-round training subset to train the semantic segmentation model, and screen the first-round verification subset according to the semantic segmentation model to obtain an updated subset.

[0096] In one embodiment, the DeeplabV3+ semantic segmentation model can be used for model training. The DeeplabV3+ semantic segmentation model includes an encoder module and a decoder module; the encoder module is used to gradually reduce feature maps and extract higher semantic information, and the decoder module is used to gradually restore spatial information.

[0097] In some other implementations, a semantic segmentation model such as UNet++ can also be used for model training. The UNet++ semantic segmentation model includes a compression channel and an expansion channel. The compression channel is used to extract image features layer by layer, and the expansion channel is used to restore the image's position information. It can also be understood that the compression channel is an encoder, and the expansion channel is a decoder.

[0098] Still taking the above A, B, C and D as an example, after the first round of training, the first round of verification subset D is screened according to the semantic segmentation model to obtain the updated subset D`.

[0099] like Figure 8 As shown, this application screens the first round of verification subsets based on the semantic segmentation model to obtain an updated subset, including:

[0100] S221, input the validation subset into the semantic segmentation model to obtain the pixel probability of each pixel point in each water body sample in the validation subset.

[0101] After the verification subset D is input into the semantic segmentation model, the semantic segmentation model will directly output the pixel probability of each pixel point in the water sample of the verification subset D.

[0102] S222: Calculate the confidence level of each water body patch in the water body label image based on the pixel probability of each pixel point in the water body sample.

[0103] This application removes the 5% of pixel values ​​with the lowest probability values ​​from the probability values ​​of all pixels in the water body map, and takes the average value of the remaining values ​​as the confidence level of the water body map.

[0104] S223 , calculating the confidence level of the water body label image according to the confidence levels of all water body patches in the water body label image.

[0105] A water body label image of the present application may contain multiple water body patches. For example, the average confidence value of all water body patches in the water body label image can be calculated as the confidence value of the entire water body label image.

[0106] S224, comparing the confidence of each water body label image with a set threshold, and taking the water body samples corresponding to the confidence greater than the set threshold as the updated subset after screening.

[0107] For example, it is assumed that the threshold value is set to P thresh , then the confidence of each water body label image can be compared with the set threshold P thresh When the confidence of the water body label image is less than the set threshold P thresh When , it indicates that the confidence of the water body label image is low, then the water body sample corresponding to the water body label image may be an erroneous water body sample, and the water body sample corresponding to the water body label image is deleted from the verification subset. When the confidence of the water body label image is greater than the set threshold P thresh When , it indicates that the confidence of the water body label image is relatively high, the water body sample corresponding to the water body label image is retained, and the verification subset is screened based on the retained water body sample corresponding to the water body label image.

[0108] S230, using the updated subset as a training subset for the next round, and selecting a subset from the training subset of the first round as a verification subset for the next round, so as to determine the training subset and verification subset for the next round and perform iterative training until each water sample in the initial water sample set has been screened once.

[0109] Still taking the above A, B, C and D as an example, after the first round of training, the verification subset D is screened to D`, and the screened D` is used as the update subset; then, in the second iteration, the first round update subset D` is used as a training subset for the second round, and a piece B is selected from the first round training subset (A, B, C) as the verification subset for the second round. Other training subsets for the second round also include the remaining pieces A and C in the first round training subset that are not used as verification subsets. The final training subset for the second round is A, C, D`, and the verification subset B is screened to be the update subset B`; similarly, in the third iteration, the verification subset is C, and the corresponding training subsets are A, B`, D`. The verification subset C is updated to be the update subset C`; in the fourth iteration, the verification subset is A, and the corresponding training subsets are B`, C`, D`, and the verification subset A is the update subset A`. When ABCD are all used as validation subsets and screened once, the cross-validation-based confidence iteration learning is completed.

[0110] S240, all updated subsets after screening are used as water body denoising sample sets.

[0111] This application uses all updated subsets A`, B`, C`, and D` after screening as water body denoising sample sets.

[0112] It can be understood that this application adopts a noise cleaning method based on verification sample confidence screening and multiple rounds of cross-validation, that is, iterative calculations are carried out through training, verification, and denoising screening links to achieve automatic denoising and cleaning of the initial water sample set.

[0113] S300, based on the water samples in RGB mode and HSV color space, classifies the water denoising sample set to obtain the remote sensing image water sample set.

[0114] Due to the influence of sediment content, light, microorganisms and human activities, water bodies such as rivers and lakes often appear in different colors in satellite remote sensing images, such as yellow, green, black, etc. The diversity of water colors also increases the difficulty of deep learning methods. In order to further subdivide the water body denoising sample set, this application creatively proposes an automatic classification method for water body samples based on RGB mode and HSV color space, which can realize automatic and efficient classification of samples.

[0115] like Figure 9 As shown in FIG, the above process of classifying the water body denoising sample set to obtain the remote sensing image water body sample set includes:

[0116] S310: Randomly sample all water sample pixels in the water denoising sample set.

[0117] In one embodiment, the process of randomly sampling all water sample pixels in the water denoising sample set is as follows: Figure 10As shown, including:

[0118] S311, calculating the proportion of each water body spot in the water body spot sample slice image according to each water body label image in the water body noise sample set, and determining the sampling number of water body sample pixels in the corresponding water body spot sample slice image according to the proportion value.

[0119] The percentage value and the number of pixel sampling N in this application are:

[0120]

[0121] Where, P is the proportion; N T is the total number of water sample pixels in a water body label image, N is the number of pixels to be sampled, m1 and m2 are sampling ratio coefficients, and the value range is 0.1 to 10.

[0122] S312, when randomly sampling the water sample pixel points in the water patch sample slice image, ensure that the sampling numbers of small water samples and large water samples in the water denoising sample set meet the set relationship.

[0123] When P < 1%, the water sample is small and is considered a small water sample; when P > 1%, the water sample is large and is considered a large water sample; to prevent excessive sampling of large water samples and too few sampling points of small water samples, it is required to ensure that the number of pixel points sampled for small water samples and large water samples is relatively balanced, and that large water samples have relatively more pixel points sampled. Generally, m1 is 2 to 5 times of m2, for example, m1 = 10, m2 = 4. Therefore, after determining the number of samples to be sampled for each water patch, this application adopts random sampling and satisfies the sampling balance. N water sample pixel points are randomly selected from the water sample pixel points in each water patch for subsequent classification calculations.

[0124] S320 , finding the mode of the RGB values ​​of all sampled water body pixels to obtain the RGB value of the water body.

[0125] This application is based on randomly sampled water sample pixels, statistics the R, G, and B values ​​corresponding to all water sample pixels, and determines the pixel R, G, and B values ​​that best represent the color of the water body by finding the majority method. RGB can reflect the color of the water body.

[0126] S330 , converting the RGB values ​​into HSV values, and classifying the water body denoising sample set based on the color classification rule of the HSV values ​​to obtain a remote sensing image water body sample set.

[0127] Considering that although RGB values ​​can reflect the color of water bodies, they cannot automatically classify the color gamut, this application introduces the HSV color space, converts the representative RGB value of each water object into HSV, and then uses the HSV color classification rules for automatic classification.

[0128] Among them, the RGB value is converted to HSV:

[0129]

[0130] Where R represents the red value, G represents the green value, and B represents the blue value; R′ represents the conversion of the red value to 0~1, G′ represents the conversion of the green value to 0~1, and B′ represents the conversion of the blue value to 0~1.

[0131]

[0132] Where C max Indicates taking the maximum value from R′, G′, B′, C min Indicates taking the minimum value from R′, G′, B′, Δ indicates C max with C min The difference.

[0133]

[0134] In the formula, H represents hue, S represents saturation, and B represents brightness.

[0135] Furthermore, the water color is automatically determined according to the interval range of the HSV color value comparison table in Table 1, and the water samples are divided into several color categories such as black, yellow, green, blue, and others.

[0136] Table 1

[0137]

[0138] In response to the demand for constructing high-quality water sample sets for business applications in large-scale intelligent interpretation of remote sensing of water bodies, this application further proposes a method for random noise cleaning of water samples based on cross-validation and confidence iteration learning, and a sample set construction method based on water body RGB mode and HSV color space classification, based on gridding and geographic location annotation. This method realizes automatic cleaning and automatic fine classification of water samples, and can construct a high-quality remote sensing water sample set, thereby effectively solving the problems of noise, outliers and insufficient sample classification in large-scale water sample sets obtained by automatic / semi-automatic methods such as remote sensing images and historical maps, and improving the accuracy of intelligent interpretation of water bodies to meet the needs of water resources management, environmental protection, disaster monitoring and other fields.

[0139] like Figure 11As shown, the embodiment of the present application further proposes a device for constructing a remote sensing image water body sample set, comprising:

[0140] The preprocessing module 10 is used to obtain a plurality of original satellite remote sensing images and preprocess each original satellite remote sensing image to obtain an initial water body sample set;

[0141] A denoising module 20 is used to perform noise cleaning on the initial water sample set based on a cross-validation and confidence iteration learning method to obtain a water denoising sample set;

[0142] The classification module 30 is used to classify the water body denoising sample set based on the water body samples in the RGB mode and HSV color space to obtain a remote sensing image water body sample set.

[0143] It can be understood that the device of this embodiment corresponds to the method for constructing the remote sensing image water body sample set of the above embodiment. The options in the above embodiment are also applicable to this embodiment, so they will not be repeated here.

[0144] The present application also provides a computer terminal. Exemplarily, the computer terminal includes a processor and a memory, wherein the memory stores a computer program, and the processor runs the computer program to enable the computer terminal to execute the functions of each module in the above-mentioned method for constructing a remote sensing image water body sample set or the above-mentioned device for constructing a remote sensing image water body sample set.

[0145] Among them, the processor can be an integrated circuit chip with signal processing capabilities. The processor can be a general-purpose processor, including a central processing unit (CPU), a graphics processing unit (GPU) and a network processor (NP), a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA) or at least one of other programmable logic devices, discrete gate or transistor logic devices, and discrete hardware components. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc., which can implement or execute the disclosed methods, steps and logic block diagrams in the embodiments of the present application.

[0146] The memory may be, but is not limited to, a random access memory (RAM), a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), etc. The memory is used to store a computer program, and the processor may execute the computer program accordingly after receiving an execution instruction.

[0147] The present application also provides a computer-readable storage medium for storing the computer program used in the above-mentioned computer terminal. For example, the computer-readable storage medium may include, but is not limited to, various media capable of storing program code, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0148] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can also be implemented in other ways. The device embodiments described above are merely schematic. For example, the flowcharts and structure diagrams in the accompanying drawings show the possible architectures, functions and operations of the devices, methods and computer program products according to the multiple embodiments of the present application. In this regard, each box in the flowchart or block diagram can represent a module, a program segment or a part of the code, and the module, program segment or a part of the code contains one or more executable instructions for implementing the specified logical functions. It should also be noted that in an alternative implementation, the functions marked in the box can also occur in an order different from that marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the structure diagram and / or flowchart, and the combination of boxes in the structure diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or action, or can be implemented using a combination of dedicated hardware and computer instructions.

[0149] In addition, the functional modules or units in the various embodiments of the present application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0150] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for enabling a computer device (which can be a smart phone, personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application.

[0151] The above is only a specific implementation method of the present application, but the scope of protection of the present application is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed in this application, which should be covered by the scope of protection of the present application.

Claims

1. A method for constructing a remote sensing image water body sample set, characterized in that: include: Acquire multiple original satellite remote sensing images, and preprocess each of the original satellite remote sensing images to obtain an initial water body sample set; The water body initial sample set is subjected to noise cleaning based on a cross-validation and confidence iteration learning method to obtain a water body denoised sample set; Based on the water samples in RGB mode and HSV color space, the water denoising sample set is classified to obtain a remote sensing image water sample set; Among them, the method based on cross-validation and confidence iteration learning performs noise cleaning on the water body initial sample set to obtain a water body denoising sample set, including: dividing the water body samples in the water body initial sample set into multiple parts, selecting one of them as a verification subset, and the remaining parts as training subsets to determine the first round of training subsets and verification subsets; using the first round of training subsets to train the semantic segmentation model, and screening the first round of verification subsets according to the semantic segmentation model to obtain an updated subset; using the updated subset as a training subset for the next round, and selecting a part from the first round of training subsets as a verification subset for the next round to determine the next round of training subsets and verification subsets and perform iterative training until each water body sample in the water body initial sample set has been screened once; using all the updated subsets after screening as the water body denoising sample set; Among them, the water body samples based on RGB mode and HSV color space are classified on the water body denoising sample set to obtain the remote sensing image water body sample set, including: randomly sampling all water body sample pixels in the water body denoising sample set; finding the mode of the RGB values ​​of all sampled water body sample pixels to obtain the RGB value of the water body sample; converting the RGB value into HSV value, and classifying the water body denoising sample set based on the color classification rule of the HSV value to obtain the remote sensing image water body sample set.

2. The method for constructing a remote sensing image water body sample set according to claim 1, characterized in that: The preprocessing of each of the original satellite remote sensing images to obtain an initial water body sample set includes: Gridding the original satellite remote sensing image based on geographic coordinates to obtain hierarchical grids corresponding to different slice sizes; Performing grid positioning and cropping processing on the original satellite remote sensing image with a hierarchical grid based on image resolution and water body spot vectors to extract a plurality of water body spot sample slice images; Performing position marking processing on each of the water body spot sample slice images to obtain a corresponding water body label image; A water body initial sample set is constructed based on all the water body spot sample slice images and the water body label images; the water body initial sample set includes multiple water body samples, and each water body sample includes a water body spot sample slice image and a corresponding water body label image.

3. The method for constructing a remote sensing image water body sample set according to claim 2, characterized in that: The gridding process of the original satellite remote sensing image based on geographic coordinates includes: Gridding the original satellite remote sensing image, and grading all grid slices after the gridding based on image resolution and slice size; The row and column numbers of each level of the slicing grid are determined according to the latitude and longitude coordinates and the slicing grid level.

4. The method for constructing a remote sensing image water body sample set according to claim 2, characterized in that: The first round of validation subset is screened according to the semantic segmentation model to obtain an updated subset, including: Inputting the validation subset into the semantic segmentation model to obtain the pixel probability of each pixel in each water body sample in the validation subset; Calculating the confidence of each water body spot sample in the water body label image according to the pixel probability of each pixel point in the water body sample; Calculating the confidence of the water body label image according to the confidence of all the water body spot samples in the water body label image; The confidence level of each water body label image is compared with a set threshold, and water body samples corresponding to the confidence level being greater than the set threshold are taken as the updated subset.

5. The method for constructing a remote sensing image water body sample set according to claim 2, characterized in that: The randomly sampling all pixel points in the water body denoising sample set includes: Calculate the proportion of each water body spot in the water body spot sample slice image according to each water body label image in the water body denoising sample set; Determine the sampling number of water body sample pixel points in the corresponding water body spot sample slice image according to the proportion value; When randomly sampling the water body sample pixel points in the water body patch sample slice image, the sampling number of small water body samples and large water body samples in the water body denoising sample set satisfies the set relationship; wherein, the water body samples with a proportion value less than 1% are defined as small water body samples, and the water body samples with a proportion value greater than 1% are defined as large water body samples.

6. A device for constructing a remote sensing image water body sample set, characterized in that: include: A preprocessing module is used to obtain a plurality of original satellite remote sensing images and preprocess each of the original satellite remote sensing images to obtain an initial water body sample set; A denoising module is used to perform noise cleaning on the initial water sample set based on a cross-validation and confidence iteration learning method to obtain a water denoising sample set; A classification module is used to classify the water body denoising sample set based on the water body samples in RGB mode and HSV color space to obtain the remote sensing image water body sample set; Among them, the method based on cross-validation and confidence iteration learning performs noise cleaning on the water body initial sample set to obtain a water body denoising sample set, including: dividing the water body samples in the water body initial sample set into multiple parts, selecting one of them as a verification subset, and the remaining parts as training subsets to determine the first round of training subsets and verification subsets; using the first round of training subsets to train the semantic segmentation model, and screening the first round of verification subsets according to the semantic segmentation model to obtain an updated subset; using the updated subset as a training subset for the next round, and selecting a part from the first round of training subsets as a verification subset for the next round to determine the next round of training subsets and verification subsets and perform iterative training until each water body sample in the water body initial sample set has been screened once; using all the updated subsets after screening as the water body denoising sample set; Among them, the water body samples based on RGB mode and HSV color space are classified on the water body denoising sample set to obtain the remote sensing image water body sample set, including: randomly sampling all water body sample pixels in the water body denoising sample set; finding the mode of the RGB values ​​of all sampled water body sample pixels to obtain the RGB value of the water body sample; converting the RGB value into HSV value, and classifying the water body denoising sample set based on the color classification rule of the HSV value to obtain the remote sensing image water body sample set.

7. A computer terminal, characterized in that: The computer terminal includes a processor and a memory, the memory stores a computer program, and the processor is used to execute the computer program to implement the method for constructing a remote sensing image water body sample set according to any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that It stores a computer program, which, when executed on a processor, implements the method for constructing a remote sensing image water body sample set according to any one of claims 1-5.

Images