A method for fine extraction of micro water bodies based on multi-source collaboration and edge perception

Abstract

The present application relates to satellite remote sensing image processing technical field, specifically to a kind of based on multi-source cooperation and edge perception's fine extraction method of micro water body, comprising the following steps: S1, obtains multi-source remote sensing image and pre-processes generation standardized data;S2, based on the data constructs multi-dimensional feature stack containing spectrum, water body index, semantic priori, texture and topography;S3, combined with semantic priori and normalized shadow index automatically constructs positive and negative sample library;S4, using the sample library and feature stack trains random forest and obtains initial water body binary graph;S5, based on water body index calculation gradient and generates edge mask;S6, to initial result is connected domain screening and historical large water body mask elimination, combined with morphological processing and edge mask reconstruction obtains final result.The present application, realizes the precision of water body recognition under complex background to improve, micro water body boundary integrity enhancement and the automation and generalization of sample construction.

Description

Technical Field

[0001] This invention relates to the field of satellite remote sensing image processing technology, and in particular to a method for refined extraction of small water bodies based on multi-source collaboration and edge perception. Background Technology

[0002] Water body extraction is a crucial foundational step in remote sensing applications, playing a vital role in refined water resource management, emergency monitoring of floods and droughts, and the treatment of complex water bodies such as those with eutrophication and high suspended solids concentrations. With the development of Earth observation technologies, the collaborative use of optical remote sensing data and synthetic aperture radar (SAR) data (such as Sentinel-1 and Sentinel-2) for large-scale water body monitoring has become the mainstream approach. However, traditional methods still struggle to achieve high-precision and stable extraction in small water bodies such as rural ponds and irrigation ditches, as well as in complex optical water body scenarios.

[0003] Existing technologies suffer from the following main shortcomings: First, threshold segmentation methods based on spectral indices suffer from the problem of identical or dissimilar objects, easily misclassifying building shadows as water bodies, and lack the ability to identify eutrophic or highly turbid water bodies. Second, traditional machine learning methods rely on manually labeled samples, making it difficult to automatically acquire high-quality, difficult-to-classify samples, and pixel-level feature-based classification ignores spatial structure information, resulting in fragmented and discontinuous extraction results for small water bodies. Third, simply overlaying SAR and optical data makes it difficult to simultaneously achieve denoising and edge preservation; speckle noise in SAR images can easily lead to excessive smoothing or even loss of the boundaries of small water bodies during the removal process, making it difficult to guarantee the integrity and continuity of small water bodies. Finally, after extracting water bodies, traditional methods often fail to automatically and accurately separate small water bodies (such as ponds and small ditches) from large, continuous water networks (such as main rivers and large lakes), and conventional area threshold screening is easily ineffective due to the connectivity or breakage of the water system. Summary of the Invention

[0004] This invention provides a refined extraction method for small water bodies based on multi-source collaboration and edge perception. It automatically constructs difficult-to-distinguish samples by introducing a normalized shadow index, enhances the identification ability of complex water bodies by combining multi-source feature fusion and an improved water body sensitivity index, and introduces historical water body priors to remove interference from large connected water networks. Finally, it preserves water body boundaries while denoising through a gradient-based edge constraint reconstruction mechanism, thereby achieving high-precision, continuous and robust extraction of small water bodies in complex backgrounds.

[0005] A method for refined extraction of small water bodies based on multi-source collaboration and edge sensing includes the following steps: S1. Acquire multi-source remote sensing image data of the target area, including optical remote sensing images and synthetic aperture radar images. Perform cloud removal, time-series synthesis and normalization processing on the optical remote sensing images. Perform thermal noise removal, radiometric calibration, terrain correction and smoothing processing on the synthetic aperture radar images to generate standardized multi-source data. S2, based on standardized multi-source data, constructs a multi-dimensional feature stack, including basic spectral features, water body sensitivity index features, land cover semantic prior features, improved water pollution index, combined enhanced water body index, land cover prior probability features, SAR multi-scale texture and roughness features, and terrain-aided features. S3, based on the semantic prior features of land cover and combined with the normalized shadow index, automatically constructs a training sample set. Among them, water areas with confidence scores higher than the screening threshold are selected as positive samples, and shadow pixels in building and vegetation areas are extracted as difficult-to-distinguish negative samples, thus forming a sample library for training the random forest classifier. S4, input the sample library and multidimensional feature stack into the random forest classifier for training, and perform pixel-by-pixel classification on the whole area image to obtain the initial water body binary map; S5: Calculate the image gradient based on the water body sensitivity index feature, and extract edge pixels with significant gradient changes by using preset high gradient thresholds and low gradient thresholds to generate an edge mask representing the location of the water body boundary. S6. Connectivity filtering is performed on the initial binary water body map. A large water body skeleton mask is constructed by combining the historical surface water occurrence rate to remove large stable water bodies. Then, morphological smoothing is performed, and edge constraint reconstruction is performed by combining the edge mask. While removing noise, the true boundary of small water bodies is preserved, and the final small water body extraction result is obtained.

[0006] Optionally, S1 includes: S11. Acquire optical remote sensing images, remove cloud and cloud shadow pixels using the QA band, and perform median composite analysis on the multi-temporal images to obtain a cloud-free time-series image, represented as: ; in, For cloudless time-series images at the pixel level band Reflectivity at that location For the first Pixel values ​​corresponding to real-time optical remote sensing images For QA band marker values, 0 indicates no clouds or cloudless shadow pixels. This is for median operations; S12, normalization processing of cloudless time-series images, is represented as follows: ; in, This is the normalized optical remote sensing image. , bands Minimum and maximum values; S13, acquire synthetic aperture radar imagery, and perform thermal noise removal, radiometric calibration, and terrain correction on it, represented as: ; ; ; in, To synthesize aperture radar images at the pixel level The amplitude value at that point, The pixel values ​​of the synthetic aperture radar image after removing thermal noise. For thermal noise, Image of backscattering coefficient at pixels The value at that location, For radiation calibration coefficients, The backscattering coefficient image after terrain correction at the pixel level The value at that location, This is the local angle of incidence; S14, the backscattering coefficient image after terrain correction is smoothed, and is represented as follows: ; in, This is the smoothed backscattering coefficient image. In pixels A sliding window centered on the user; S15 combines the normalized optical remote sensing imagery with the smoothed backscattering coefficient imagery to form standardized multi-source data.

[0007] Optionally, S2 includes: S21, extract the basic spectral features from the normalized optical remote sensing image, including blue band, green band, red band, near-infrared band and short-wave infrared band; S22, based on the green band, near-infrared band, and shortwave infrared band, calculates the water body sensitivity index characteristics, including the normalized differential water body index and the improved normalized differential water body index, expressed as: ; ; in, The normalized differential water index, For green band reflectivity, For near-infrared reflectivity, For shortwave infrared reflectivity, To improve the normalized differential water index; S23, an improved water pollution index is constructed using the red, green, and near-infrared bands, and is expressed as: ; in, To improve the water pollution index, Reflectivity in the red band; S24, a combined enhanced water quality index constructed by integrating blue, green, near-infrared, and short-wave infrared bands, is expressed as: ; in, To enhance the water quality index, Blue band reflectivity; S25. Based on the normalized optical remote sensing image, the probability information of water bodies and submerged vegetation is extracted to construct the prior probability features of land cover, including the probability of water bodies and the probability of submerged vegetation, as follows: ; ; in, For pixels The probability value of belonging to the water body category. For pixels The probability value of belonging to the submerged vegetation category. For pixels Category The discrimination score, For the set of all land cover categories, It is an exponential function. Scoring for water bodies The score is used to distinguish between submerged vegetation. ; ; ; in, , , These are the weight vectors for the corresponding categories. , , These are the bias terms for the corresponding categories; S26, SAR multi-scale texture and roughness features are calculated based on the smoothed backscattering coefficient image, including contrast, entropy, and spatial roughness, expressed as: ; ; ; in, For contrast, For entropy, For spatial roughness, For elements of the gray-level co-occurrence matrix, For the sliding window area, The number of pixels within the window. The mean NDWI value within the window; S27, using a digital elevation model to calculate auxiliary terrain features, including elevation and slope, is expressed as follows: ; in, This is the elevation value. This is the slope value. , These represent the rates of change of elevation in the spatial direction.

[0008] Optionally, S3 includes: S31, based on the prior probability characteristics of land cover, and by comparing it with a preset water body probability screening threshold. (Set to 0.9) for comparison, when When the time comes, select the corresponding pixel and mark it as a positive sample; S32, calculate the normalized shading index, used to identify shading regions similar to the water body spectrum, expressed as: ; ; in, For pixels Normalized shadow index at the location, For near-infrared reflectivity, For color saturation components, , , These are the reflectance values ​​for the red, green, and blue bands, respectively. S33, extract building and tree category pixels based on land cover semantic data to construct a building and tree area mask. Within the constraints of the building and tree area mask, filter the area using a preset shadow screening threshold. (Set to 0.85) for comparison, when When this happens, the corresponding pixel is marked as a difficult-to-distinguish negative sample; S34, label the positive samples and the difficult-to-distinguish negative samples to form a sample library. , among which, if ,but ,like and ,but (Non-water body) For sample labels, This is a positive water sample. It is difficult to distinguish negative samples.

[0009] Optionally, S4 includes: S41, the generated sample library and multidimensional feature stack are input into a random forest classifier for training, resulting in a classification model for water body identification, represented as: ; ; in, For random forest classifiers, This is the training function for the random forest. For pixels The feature vector at that location, For each feature in the multidimensional feature stack at the pixel level The value at that location, The feature dimension; S42, the trained random forest classifier is used to classify the entire region image pixel by pixel, obtaining the classification result for each pixel, and generating an initial binary map of the water body, represented as: ; ; in, For pixels The classification results For the initial binary image of the water body in pixels The value at that location, when When, it indicates that the pixel is identified as a body of water. When the value is zero, it indicates that the pixel is determined to be a non-water body.

[0010] Optionally, S5 includes: S51, based on the water body sensitivity index feature, uses the Canny edge detection operator to calculate the gradient, obtaining the image gradient magnitude, expressed as: ; ; ; in, For pixels gradient magnitude at that point , These are the gradient components in the horizontal and vertical directions, respectively; S52, based on gradient magnitude, filters pixels with significant gradient changes by using high and low gradient thresholds, and generates a binarized edge mask.

[0011] Optionally, the binarized edge mask is represented as: ; in, For binary edge masking, For edge pixels, Non-edge pixels, For high gradient threshold, This is a low gradient threshold.

[0012] Optionally, S6 includes: S61, acquire historical surface water occurrence rate data, extract historical stable water bodies with occurrence rates greater than a set occurrence rate threshold, and perform connected component analysis to generate a large water body skeleton mask, represented as: ; ; in, Historical occurrence rate of surface water bodies As the incidence threshold, A binary diagram of historically stable water bodies. For the connectivity of historically stable water bodies, The threshold for the skeleton area of ​​a large water body. It serves as a mask for the framework of large bodies of water.

[0013] S62, Perform connected component analysis on the initial binary water body graph, calculate the area of ​​each connected component, and perform joint screening based on the large water body skeleton mask and area threshold, as shown below: ; ; in, This represents the area of ​​the connected region (number of pixels). For the first Connected components, The filtered connected components, Minimum area threshold, The maximum area threshold, ; S63, the morphological smoothing process is applied to the screened water area, represented as follows: ; in, This is the smoothed binary image. The filtered binary image. In pixels A sliding window centered on the center. These are the pixel coordinates within the window. For the mode operation; S64 integrates the morphological smoothing results with the edge mask and the initial water body binary map to generate the final water body binary map as the final result of small water body extraction.

[0014] Optionally, the final binary water body map is represented as follows: ; in, This is the final binary map of the water body.

[0015] The beneficial effects of this invention are: This invention introduces an automatic mining mechanism for difficult-to-distinguish negative samples guided by an improved water pollution index (NDPI) and normalized shadow index (NSI) into the multidimensional feature stack. This enables the model to effectively distinguish between water bodies and spectrally similar features such as shadows, and enhances its response to complex water bodies such as eutrophication and high turbidity. As a result, it significantly reduces the false alarm rate and false negative rate, and improves the accuracy of water body identification under complex backgrounds and poor water quality conditions.

[0016] This invention constructs an edge mask based on the water body sensitivity index gradient and introduces an edge constraint reconstruction mechanism in the post-processing stage to fuse the morphological smoothing results with the initial classification results. While removing noise, it performs constraint restoration on the edge region, thereby effectively avoiding the problem of small water body boundary breakage caused by traditional morphological processing and significantly improving the boundary integrity and spatial continuity of small water bodies.

[0017] This invention, by combining land cover semantic information and physical index features, realizes the automated construction process of training samples without relying on manual annotation or specific data sources. It can quickly generate high-quality sample libraries in large-scale, multi-temporal remote sensing scenarios, thereby improving the versatility and scalability of the method and enhancing the system's application capabilities under different regional and multi-source data conditions.

[0018] This invention, by introducing prior knowledge of the occurrence of historical surface water bodies, constructs a large water body skeleton mask and jointly eliminates interference from large, stable, and interconnected water networks during the connectivity domain screening stage. This effectively solves the problem of fragmented blocks at the edges of large water bodies being misjudged as small water bodies in complex water system scenarios, thereby achieving high-precision and high-purity extraction of real small water bodies such as rural ponds and small ditches. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the extraction method according to an embodiment of the present invention. Detailed Implementation

[0021] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. Those skilled in the art may employ other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0022] like Figure 1 As shown, a method for refined extraction of small water bodies based on multi-source collaboration and edge perception includes the following steps: S1. Acquire multi-source remote sensing image data of the target area, including optical remote sensing images and synthetic aperture radar images. Perform cloud removal, time-series synthesis and normalization processing on the optical remote sensing images. Perform thermal noise removal, radiometric calibration, terrain correction and smoothing processing on the synthetic aperture radar images to generate standardized multi-source data. S2, based on standardized multi-source data, constructs a multi-dimensional feature stack, including basic spectral features, water body sensitivity index features, land cover semantic prior features, improved water pollution index, combined enhanced water body index, land cover prior probability features, SAR multi-scale texture and roughness features, and terrain auxiliary features, to enhance the ability to distinguish complex water bodies and interfering land features. S3, based on the semantic prior features of land cover and combined with the normalized shadow index, automatically constructs a training sample set. Among them, water areas with confidence scores higher than the screening threshold are selected as positive samples, and shadow pixels in building and vegetation areas are extracted as difficult-to-distinguish negative samples, thus forming a sample library for training the random forest classifier. S4. Input the sample library and multidimensional feature stack into the random forest classifier for training, and perform pixel-by-pixel classification on the whole area image to obtain the initial water body binary map. S5 calculates the image gradient based on the water body sensitivity index feature. Through preset high gradient threshold and low gradient threshold, it extracts edge pixels with significant gradient changes and generates an edge mask that represents the water body boundary position. This mask is used to mark the precise boundary position between water and land, and between water and buildings. S6. Connectivity filtering is performed on the initial binary water body map. A large water body skeleton mask is constructed by combining the historical surface water occurrence rate to remove large stable water bodies. Then, morphological smoothing is performed, and edge constraint reconstruction is performed by combining the edge mask. While removing noise, the true boundary of small water bodies is preserved, and the final small water body extraction result is obtained.

[0023] S1 includes: S11. Acquire optical remote sensing images, remove cloud and cloud shadow pixels using the QA band, and perform median composite analysis on the multi-temporal images to obtain a cloud-free time-series image, represented as: ; in, For cloudless time-series images at the pixel level band Reflectivity at that location For the first Pixel values ​​corresponding to real-time optical remote sensing images For QA band marker values, 0 indicates no clouds or cloudless shadow pixels. This is for median operations; S12, normalization processing is performed on the cloudless time-series image to eliminate radiometric differences between different bands and different times, as shown below: ; in, This is the normalized optical remote sensing image. , bands Minimum and maximum values; S13, acquire synthetic aperture radar imagery, and perform thermal noise removal, radiometric calibration, and terrain correction on it, represented as: ; ; ; in, To synthesize aperture radar images at the pixel level The amplitude value at that point, The pixel values ​​of the synthetic aperture radar image after removing thermal noise. For thermal noise, Image of backscattering coefficient at pixels The value at that location, For radiation calibration coefficients, The backscattering coefficient image after terrain correction at the pixel level The value at that location, This is the local angle of incidence; S14, the backscattering coefficient image after terrain correction is smoothed to reduce the impact of speckle noise on subsequent analysis, as shown below: ; in, This is the smoothed backscattering coefficient image. In pixels A sliding window centered on the user; S15 combines the normalized optical remote sensing imagery with the smoothed backscattering coefficient imagery to form standardized multi-source data.

[0024] S2 includes: S21, extract the basic spectral features from the normalized optical remote sensing image, including blue band, green band, red band, near-infrared band and short-wave infrared band; S22, based on the green band, near-infrared band, and short-wave infrared band, calculates water body sensitivity index characteristics, including the normalized differential water body index and the improved normalized differential water body index, to enhance water body identification capabilities, expressed as: ; ; in, The normalized differential water index, For green band reflectivity, For near-infrared reflectivity, For shortwave infrared reflectivity, To improve the normalized differential water index; S23, an improved water pollution index is constructed using red, green, and near-infrared bands to enhance the response to eutrophic and highly turbid water bodies, and is expressed as: ; in, To improve the water pollution index, Reflectivity in the red band; S24, combining blue, green, near-infrared, and short-wave infrared bands to construct a combined enhanced water index, aims to suppress the interference of mixed pixels on water body extraction, and is represented as: ; in, To enhance the water quality index, Blue band reflectivity; S25. Based on the normalized optical remote sensing image, the probability information of water bodies and submerged vegetation is extracted to construct the prior probability features of land cover, including the probability of water bodies and the probability of submerged vegetation, as follows: ; ; in, For pixels The probability value of belonging to the water body category. For pixels The probability value of belonging to the submerged vegetation category. For pixels Category The discrimination score, For the set of all land cover categories, It is an exponential function. Scoring for water bodies The score is used to distinguish between submerged vegetation. ; ; ; in, , , These are the weight vectors for the corresponding categories. , , These are the bias terms for the corresponding categories; S26, SAR multi-scale texture and roughness features are calculated based on the smoothed backscattering coefficient image, including contrast, entropy, and spatial roughness, to distinguish water bodies and interfering targets such as shadows, as shown below: ; ; ; in, For contrast, For entropy, For spatial roughness, For elements of the gray-level co-occurrence matrix, For the sliding window area, The number of pixels within the window. The mean NDWI value within the window; S27, using a digital elevation model to calculate auxiliary terrain features, including elevation and slope, to help eliminate the interference of mountain shadows caused by terrain undulations, is represented as: ; in, This is the elevation value. This is the slope value. , These represent the rates of change of elevation in the spatial direction.

[0025] S3 includes: S31, based on the prior probability characteristics of land cover, and by comparing it with a preset water body probability screening threshold. (Set to 0.9) for comparison, when When the time comes, select the corresponding pixel and mark it as a positive sample; S32, calculate the normalized shading index to identify shading regions similar to the water body spectrum, expressed as: ; ; in, For pixels Normalized shadow index at the location, For near-infrared reflectivity, For color saturation components, , , These are the reflectance values ​​for the red, green, and blue bands, respectively. S33: Based on the semantic data of land cover, extract building and tree category pixels to construct building and tree region masks. Within the constraints of the building and tree region masks, and by comparing with a preset shadow filtering threshold... (Set to 0.85) for comparison, when When this happens, the corresponding pixel is marked as a difficult-to-distinguish negative sample; ; in, To mask the areas with buildings and trees, For pixels in land cover semantic data Corresponding category tags, For architecture, Trees; S34, label the positive samples and the difficult-to-distinguish negative samples to form a sample library. , among which, if ,but ,like and ,but (Non-water body) For sample labels, This is a positive water sample. It is difficult to distinguish negative samples.

[0026] S4 includes: S41, the generated sample library and multidimensional feature stack are input into a random forest classifier for training, resulting in a classification model for water body identification, represented as: ; ; in, For random forest classifiers, This is the training function for the random forest. For pixels The feature vector at that location, For each feature in the multidimensional feature stack at the pixel level The value at that location, The feature dimension; S42, the trained random forest classifier is used to classify the entire region image pixel by pixel, obtaining the classification result for each pixel, and generating an initial binary map of the water body, represented as: ; ; in, For pixels The classification results For the initial binary image of the water body in pixels The value at that location, when When, it indicates that the pixel is identified as a body of water. When the value is zero, it indicates that the pixel is determined to be a non-water body.

[0027] S5 includes: S51, based on the water body sensitivity index feature, uses the Canny edge detection operator to calculate the gradient, obtaining the image gradient magnitude, expressed as: ; ; ; in, For pixels gradient magnitude at that point , These are the gradient components in the horizontal and vertical directions, respectively; S52, based on gradient magnitude, filters pixels with significant gradient changes by using high and low gradient thresholds, and generates a binarized edge mask.

[0028] The binary edge mask is represented as follows: ; in, For binary edge masking, For edge pixels, Non-edge pixels, For high gradient threshold, This is a low gradient threshold.

[0029] S6 includes: S61, acquire historical surface water occurrence rate data, extract historical stable water bodies with occurrence rates greater than a set occurrence rate threshold, and perform connected component analysis to generate a large water body skeleton mask, represented as: ; ; in, Historical occurrence rate of surface water bodies As the incidence threshold, A binary diagram of historically stable water bodies. For the connectivity of historically stable water bodies, The threshold for the skeleton area of ​​a large water body. It serves as a mask for the framework of large bodies of water.

[0030] S62, Perform connected component analysis on the initial binary water body graph, calculate the area of ​​each connected component, and perform joint screening based on the large water body skeleton mask and area threshold, as shown below: ; ; in, This represents the area of ​​the connected region (number of pixels). For the first Connected components, The filtered connected components, Minimum area threshold, The maximum area threshold, ; S63, morphological smoothing is performed on the screened water area to eliminate boundary jaggedness and fill in internal micro-voids, represented as: ; in, This is the smoothed binary image. The filtered binary image. In pixels A sliding window centered on the center. These are the pixel coordinates within the window. For the mode operation; S64 integrates the morphological smoothing results with the edge mask and the initial water body binary map to achieve edge fidelity and generate the final water body binary map as the final result of small water body extraction.

[0031] The final binary water body diagram is represented as follows: ; in, This is the final binary map of the water body.

[0032] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0033] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for refined extraction of small water bodies based on multi-source collaboration and edge sensing, characterized in that, Includes the following steps: S1. Acquire multi-source remote sensing image data of the target area, including optical remote sensing images and synthetic aperture radar images. Perform cloud removal, time-series synthesis and normalization processing on the optical remote sensing images. Perform thermal noise removal, radiometric calibration, terrain correction and smoothing processing on the synthetic aperture radar images to generate standardized multi-source data. S2, based on standardized multi-source data, constructs a multi-dimensional feature stack, including basic spectral features, water body sensitivity index features, land cover semantic prior features, improved water pollution index, combined enhanced water body index, land cover prior probability features, SAR multi-scale texture and roughness features, and terrain-aided features. S3, based on the semantic prior features of land cover and combined with the normalized shadow index, automatically constructs a training sample set. Among them, water areas with confidence scores higher than the screening threshold are selected as positive samples, and shadow pixels in building and vegetation areas are extracted as difficult-to-distinguish negative samples, thus forming a sample library for training the random forest classifier. S4, input the sample library and multidimensional feature stack into the random forest classifier for training, and perform pixel-by-pixel classification on the whole area image to obtain the initial water body binary map; S5: Calculate the image gradient based on the water body sensitivity index feature, and extract edge pixels with significant gradient changes by using preset high gradient thresholds and low gradient thresholds to generate an edge mask representing the location of the water body boundary. S6. Connectivity filtering is performed on the initial binary water body map. A large water body skeleton mask is constructed by combining the historical occurrence rate of surface water bodies to remove interference from large water bodies. Morphological smoothing is then performed. Edge constraint reconstruction is then performed by combining the edge mask to remove noise while preserving the true boundaries of small water bodies, thus obtaining the final small water body extraction result.

2. The method for refined extraction of small water bodies based on multi-source collaboration and edge perception according to claim 1, characterized in that, S1 includes: S11. Acquire optical remote sensing images, remove cloud and cloud shadow pixels using the QA band, and perform median composite analysis on the multi-temporal images to obtain a cloud-free time-series image, represented as: ； in, For cloudless time-series images at pixel band Reflectivity at that location For the first Pixel values ​​corresponding to real-time optical remote sensing images For QA band marker values, 0 indicates no clouds or cloudless shadow pixels. This is for median operations; S12, normalization processing of cloudless time-series images, is represented as follows: ； in, This is the normalized optical remote sensing image. , bands Minimum and maximum values; S13, acquire synthetic aperture radar imagery, and perform thermal noise removal, radiometric calibration, and terrain correction on it, represented as: ； ； ； in, To synthesize aperture radar images at the pixel level The amplitude value at that point, The pixel values ​​of the synthetic aperture radar image after removing thermal noise. For thermal noise, Image of backscattering coefficient at pixels The value at that location, For radiation calibration coefficients, The backscattering coefficient image after terrain correction at the pixel level The value at that location, This is the local angle of incidence; S14, the backscattering coefficient image after terrain correction is smoothed, and is represented as follows: ； in, This is the smoothed backscattering coefficient image. In pixels A sliding window centered on the user; S15 combines the normalized optical remote sensing imagery with the smoothed backscattering coefficient imagery to form standardized multi-source data.

3. The method for refined extraction of small water bodies based on multi-source collaboration and edge perception according to claim 2, characterized in that, S2 includes: S21, extract the basic spectral features from the normalized optical remote sensing image, including blue band, green band, red band, near-infrared band and short-wave infrared band; S22, based on the green band, near-infrared band, and shortwave infrared band, calculates the water body sensitivity index characteristics, including the normalized differential water body index and the improved normalized differential water body index, expressed as: ； ； in, The normalized differential water index, For green band reflectivity, For near-infrared reflectivity, For shortwave infrared reflectivity, To improve the normalized differential water index; S23, an improved water pollution index is constructed using the red, green, and near-infrared bands, and is expressed as: ； in, To improve the water pollution index, Reflectivity in the red band; S24, a combined enhanced water quality index constructed by integrating blue, green, near-infrared, and short-wave infrared bands, is expressed as: ； in, To enhance the water quality index, Blue band reflectivity; S25. Based on the normalized optical remote sensing image, the probability information of water bodies and submerged vegetation is extracted to construct the prior probability features of land cover, including the probability of water bodies and the probability of submerged vegetation, as follows: ； ； in, For pixels The probability value of belonging to the water body category. For pixels The probability value of belonging to the submerged vegetation category. For pixels Category The discrimination score, For the set of all land cover categories, It is an exponential function. Scoring for water bodies The score is used to distinguish between submerged vegetation. ； ； ； in, , , These are the weight vectors for the corresponding categories. , , These are the bias terms for the corresponding categories; S26, SAR multi-scale texture and roughness features are calculated based on the smoothed backscattering coefficient image, including contrast, entropy, and spatial roughness, expressed as: ； ； ； in, For contrast, For entropy, For spatial roughness, For elements of the gray-level co-occurrence matrix, For the sliding window area, The number of pixels within the window. The mean NDWI value within the window; S27, using a digital elevation model to calculate auxiliary terrain features, including elevation and slope, is expressed as follows: ； in, This is the elevation value. This is the slope value. , These represent the rates of change of elevation in the spatial direction.

4. The method for refined extraction of small water bodies based on multi-source collaboration and edge perception according to claim 3, characterized in that, S3 includes: S31, based on the prior probability characteristics of land cover, and by comparing it with a preset water body probability screening threshold. When comparing, When the time comes, select the corresponding pixel and mark it as a positive sample; S32, calculate the normalized shading index, used to identify shading regions similar to the water body spectrum, expressed as: ； ； in, For pixels Normalized shadow index at the location, For near-infrared reflectivity, For color saturation components, , , These are the reflectance values ​​for the red, green, and blue bands, respectively. S33, extract building and tree category pixels based on land cover semantic data to construct a building and tree area mask. Within the constraints of the building and tree area mask, filter the area using a preset shadow screening threshold. When comparing, When this happens, the corresponding pixel is marked as a difficult-to-distinguish negative sample; S34, label the positive samples and the difficult-to-distinguish negative samples to form a sample library. , among which, if ,but ,like and ,but , For sample labels, This is a positive water sample. It is difficult to distinguish negative samples.

5. The method for refined extraction of small water bodies based on multi-source collaboration and edge perception according to claim 4, characterized in that, S4 includes: S41, the generated sample library and multidimensional feature stack are input into a random forest classifier for training, resulting in a classification model for water body identification, represented as: ； ； in, For random forest classifiers, This is the training function for the random forest. For pixels The feature vector at that location, For each feature in the multidimensional feature stack at the pixel level The value at that location, The feature dimension; S42, the trained random forest classifier is used to classify the entire region image pixel by pixel, obtaining the classification result for each pixel, and generating an initial binary map of the water body, represented as: ； ； in, For pixels The classification results For the initial binary image of the water body in pixels The value at that location, when When, it indicates that the pixel is identified as a body of water. When the value is zero, it indicates that the pixel is determined to be a non-water body.

6. The method for refined extraction of small water bodies based on multi-source collaboration and edge perception according to claim 5, characterized in that, S5 includes: S51, based on the water body sensitivity index feature, uses the Canny edge detection operator to calculate the gradient, obtaining the image gradient magnitude, expressed as: ； ； ； in, For pixels gradient magnitude at that point , These are the gradient components in the horizontal and vertical directions, respectively; S52, based on gradient magnitude, filters pixels with significant gradient changes by using high and low gradient thresholds, and generates a binarized edge mask.

7. The method for refined extraction of small water bodies based on multi-source collaboration and edge perception according to claim 6, characterized in that, The binarized edge mask is represented as follows: ； in, For binary edge masking, For edge pixels, Non-edge pixels, For high gradient threshold, This is a low gradient threshold.

8. The method for refined extraction of small water bodies based on multi-source collaboration and edge perception according to claim 7, characterized in that, S6 includes: S61, acquire historical surface water occurrence rate data, extract historical stable water bodies with occurrence rates greater than a set occurrence rate threshold, and perform connected component analysis to generate a large water body skeleton mask, represented as: ； ； in, Historical occurrence rate of surface water bodies As the incidence threshold, A binary map of historically stable water bodies. For the connectivity of historically stable water bodies, The threshold for the skeleton area of ​​a large water body. A mask for the framework of large water bodies; S62, Perform connected component analysis on the initial binary water body graph, calculate the area of ​​each connected component, and perform joint screening based on the large water body skeleton mask and area threshold, as shown below: ； ； in, The area of ​​the connected region. For the first Connected components, The filtered connected components, Minimum area threshold, The maximum area threshold, ; S63, the morphological smoothing process is applied to the screened water area, represented as follows: ； in, The smoothed binary image. The filtered binary image. In pixels A sliding window centered on the center. These are the pixel coordinates within the window. For the mode operation; S64 integrates the morphological smoothing results with the edge mask and the initial water body binary map to generate the final water body binary map as the final result of small water body extraction.

9. The method for refined extraction of small water bodies based on multi-source collaboration and edge perception according to claim 8, characterized in that, The final binary water body diagram is represented as follows: ； in, This is the final binary map of the water body.

Images