A zero-sample pond water surface extraction method based on multi-resolution remote sensing images
By combining the characteristics of medium-resolution and high-resolution remote sensing images, using the K-Means algorithm and convolutional neural network, the precise extraction of pond water bodies is achieved, and the problem of high cost of water aggregation and manual labeling in the existing technology is solved. Improves extraction accuracy and efficiency.
Patent Information
- Application Number
- CN202411187017.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-28
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2044-08-28
AI Technical Summary
The prior art is prone to problems of water aggregation and edge mixed cell when extracting pond water bodies, and methods based on high-resolution images require a lot of manual labeling, which is costly.
Using the zero-sample method of combined multi-resolution remote sensing images, the spectral characteristics of medium-resolution images and spatial characteristics of high-resolution images are combined, and the K-Means algorithm and convolutional neural network are used to achieve accurate extraction of pond water bodies and reduce the Manual labeling cost.
It effectively solves the accuracy problem of water extraction in middle-segment images, realizes zero sample local object extraction on high-segment images, reduces manual labeling costs, and improves extraction accuracy and efficiency.
Smart Images

Figure CN118711071B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of target object extraction, and in particular to a zero-sample pond water surface extraction method combined with multi-resolution remote sensing images. Background Art
[0002] Pond water refers to a relatively closed water body with a water storage capacity of less than 100,000 cubic meters, which is naturally formed or artificially excavated. It is an important part of the national land space and the basis of fishery breeding, aquatic planting and salt industry. How to effectively and accurately extract pond water is of great significance to promoting green development and promoting the development of fishery economy. The large-scale, low-cost and real-time characteristics of ground remote sensing technology make it an effective means to obtain ground information in the fields of resource survey and environmental monitoring. At present, the methods of extracting pond water using remote sensing images are mostly based on two data sources: medium-resolution remote sensing images and high-resolution remote sensing images. The characteristic index is constructed based on spectral features on medium-resolution images to extract water bodies, but the aggregation characteristics of pond water bodies will cause the extraction results to appear that the water bodies are interconnected and there are mixed pixels at the boundaries, thus affecting the extraction accuracy of the results; high-resolution images, due to their resolution advantages, overcome the influence of water body spacing and mixed pixels on the accuracy of water body information extraction to a certain extent. High-resolution image classification based on convolutional neural networks has become a hot solution, but this method is difficult to get rid of the dilemma of high cost of obtaining samples by manual annotation.
[0003] Therefore, for the problem of accurate extraction of pond water bodies in remote sensing images, the existing pond water extraction methods have the following limitations: (1) When extracting pond water bodies on medium-resolution images, it is easy to have clustered water adhesion and edge mixed pixels, resulting in blurred edge information; (2) In the pond water extraction method based on high-resolution images using convolutional neural networks, the training and prediction of the model rely on manually labeled high-quality pixel-level samples, which requires a lot of manpower and material resources. Summary of the invention
[0004] Purpose of the invention: To propose a zero-sample pond water surface extraction method combining multi-resolution remote sensing images to solve the above-mentioned problems existing in the prior art.
[0005] The present invention proposes a zero-sample pond water surface extraction method using multi-resolution remote sensing images, the steps of which are as follows:
[0006] Step 1, collect medium-resolution remote sensing images and high-resolution remote sensing images of the area of interest and preprocess them; mark the pond water bodies on the preprocessed high-resolution remote sensing images, generate high-resolution pond water body true value labels, and resample the high-resolution pond water body true value labels to medium-resolution pond water body true value labels, where the number of rows and columns of the medium-resolution pond water body true value labels is consistent with the number of rows and columns of the original medium-resolution remote sensing images;
[0007] Step 2, calculate the normalized water index NDWI of the medium-resolution remote sensing image in the area of interest; use the predetermined band spectral reflectance set to calculate the normalized difference spectral vector NDSV; set a threshold based on the normalized water index NDWI to separate the surface water body from the background objects, binarize the normalized water index NDWI according to the threshold, and obtain the preliminary result of the medium-resolution image water body;
[0008] Step 3, marking the preliminary results of the water body in the center-divided image to obtain water body patches; inputting the normalized water body index NDWI, the normalized difference spectral vector NDSV and the near-infrared band as clustering features, using the K-Means algorithm to cluster the pixels in the water body patches according to the clustering features, and extracting the water body extraction results of the center-divided image;
[0009] Step 4, extracting a geometric feature set of water body patches from the water body extraction result of the center-divided image; constructing a classification decision tree based on the geometric feature set, and selecting the water body patches that meet the conditions as pond samples and non-pond samples; forming a training sample set with the pond samples and non-pond samples, training a classifier using the training sample set to obtain the center-divided image pond water body sample labels, and performing accuracy evaluation on the center-divided image pond water body sample labels according to the center-divided pond water body true value labels;
[0010] Step 5: Segment the high-resolution remote sensing image to obtain object units, resample the medium-resolution image pond water sample labels to high resolution; set sample cross-scale migration rules to judge by object units, migrate the medium-resolution image pond water sample to high resolution, and obtain the high-resolution image pond water sample labels;
[0011] Step 6. Synchronously crop the high-resolution image pond water sample labels and the corresponding high-resolution remote sensing images, set the label extraction rules for scene-by-scene judgment, and obtain the scene label data set; send the scene label data set to the convolutional neural network classification model for model training and parameter iterative optimization; send the high-resolution remote sensing image to the trained convolutional neural network classification model to predict the high-resolution pond water extraction results, and evaluate the accuracy of the results based on the high-resolution pond water true value labels.
[0012] In a further embodiment, in step 1, medium-resolution remote sensing images and high-resolution remote sensing images of the area of interest are collected and preprocessed, specifically including: radiation calibration, atmospheric correction, unified coordinate system, geographic registration, and image cropping.
[0013] In a further embodiment, in step 1, the high-score pond water body true value label is resampled to the medium-score pond water body true value label, satisfying the following conditions:
[0014] Adjust the marginal rows and columns of the medium-resolution pond water body true value labels to ensure that the number of rows and columns of the resampled medium-resolution pond water body true value labels is consistent with the number of rows and columns of the original medium-resolution remote sensing image.
[0015] In a further embodiment, step 2 specifically includes:
[0016] Step 2.1: Normalize the reflectance of the green band and near-infrared band of the medium-resolution remote sensing image to calculate the normalized water index NDWI:
[0017]
[0018] In the formula, is the surface reflectance of the pixel in the green light band, is the surface reflectance of the pixel in the near-infrared band;
[0019] Step 2.2: Calculate the normalized difference spectral vector NDSV using the set of spectral reflectances of the predetermined band:
[0020]
[0021]
[0022] In the formula, ; ; is the number of bands; and is the band reflectivity; For the Band reflectivity and The normalized difference in reflectance between the bands.
[0023] In a further embodiment, step 3 specifically includes:
[0024] Step 3.1, traverse each water patch in turn through the K-Means algorithm according to the clustering characteristics, cluster the pixels in each water patch, and obtain the patch clustering results;
[0025] The K-Means clustering algorithm uses the Euclidean distance between the pixel cluster feature vector and the cluster center feature vector as a similarity measure, and the squared error SSE of the patch as the objective function to measure the clustering quality. It seeks the best clustering result by finding the minimum objective function.
[0026] Step 3.2: Select the Calinski-Harabasz index to evaluate the clustering results and select the optimal number of clusters block by block. The calculation formula is as follows:
[0027]
[0028] In the formula, is the trace of the matrix, is the number of all pixels in the patch, is the number of classification categories; is the between-class covariance matrix, is the intra-class data covariance matrix;
[0029] Step 3.3: Set the threshold according to the area of each water patch and the average value of the normalized water index NDWI of the water patch to eliminate the classes that do not meet the conditions.
[0030] In a further embodiment, in step 3.1, the Euclidean distance calculation formula between the pixel object in the patch and the cluster center is as follows:
[0031]
[0032] In the formula, is the current pixel object, is the category number of this class, For the Cluster centers, is the pixel cluster feature number, is the number of clustering features of the pixel, for No. Clustering features, for No. Clustering features;
[0033] The calculation formula of the error sum of squares SSE of the entire patch is as follows:
[0034]
[0035] In the formula, is the category number of this class, Indicates the preset number of clusters, is the current pixel object, For the Cluster center.
[0036] In a further embodiment, the between-class covariance matrix and the intra-class data covariance matrix The calculation formula is as follows:
[0037]
[0038] In the formula, is the category number of this class, Indicates the preset number of clusters, is the number of all pixels in the patch, Indicates The mean vector of clustering features of class pixels, is the cluster feature mean of all pixels in the patch, is the transpose of the matrix, is the current pixel clustering feature vector.
[0039] In a further embodiment, in step 4, a geometric feature set of water patches is extracted from the water body extraction result of the center-divided image, and the geometric features included in the geometric feature set are: area, perimeter, minimum circumscribed rectangle area / patch area, minimum circumscribed rectangle length / minimum circumscribed rectangle width, compactness, area / perimeter.
[0040] In a further embodiment, step 5 specifically includes:
[0041] Step 5.1, extract high-resolution remote sensing image object units using an image multi-scale segmentation algorithm, adjust iterative shape factor and compactness factor parameters, and obtain the optimal segmented object units;
[0042] Step 5.2, resample the pond water sample labels in the medium-resolution image to high resolution, use the nearest neighbor assignment method to obtain the pond water sample labels in the high-resolution image, adjust the edge rows and columns of the pond water sample labels in the high-resolution image, and ensure that the number of rows and columns of the pond water sample labels in the resampled high-resolution image is consistent with the number of rows and columns in the medium-resolution remote sensing image;
[0043] Step 5.3: Separate pond water sample labels and non-pond water sample labels based on the pond water pixel ratio parameter of the object unit to obtain high-resolution image pond water sample labels.
[0044] In a further embodiment, step 6 specifically includes:
[0045] Step 6.1, set the size of the cropped image, set the label extraction rules according to the parameters of the pond water pixel ratio in the scene, assign labels to the cropped scene according to the rules, and obtain the scene label data set;
[0046] Step 6.2: Based on the true value labels of high-resolution pond water bodies, the predicted high-resolution pond water body extraction results are evaluated for accuracy using five indicators, including accuracy, precision, recall, balanced F-score, and intersection-over-union ratio.
[0047] Compared with the prior art, the present invention has the following beneficial effects:
[0048] (1) The present invention proposes a zero-sample pond water surface extraction method using combined multi-resolution remote sensing images. This method combines the advantages of the spectral features of medium-resolution images and the fine spatial features of high-resolution images to effectively solve the problem of fine extraction of medium-resolution pond water bodies. At the same time, it achieves zero-sample extraction of pond water bodies using high-resolution images, greatly reducing the cost of manual labeling.
[0049] (2) The water body extraction strategy of the center-resolution image is to fully exploit the spectral characteristics of the center-resolution image to form clustering results without relying on external data, evaluate and select the optimized water body extraction results category by category, reduce the adhesion phenomenon between water bodies, and improve the water body extraction accuracy of the center-resolution image.
[0050] (3) Zero-sample extraction of pond water bodies from high-resolution images. The pond water body sample labels extracted from medium-resolution images are migrated across scales to a high-resolution image pond water body label dataset, which is used as training data for training the convolutional neural network model. This achieves unlabeled zero-sample extraction of ground objects on high-resolution images while ensuring accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0051] Figure 1 This is a flow chart of the zero-sample pond water surface extraction method using combined multi-resolution remote sensing images proposed in the present invention.
[0052] Figure 2 The figure is a schematic diagram of constructing a decision tree in an embodiment of the present invention.
[0053] Figure 3 Schematic diagram of sample cross-scale migration rules in an embodiment of the present invention.
[0054] Figure 4 This is a rule graph for extracting a label data set in an embodiment of the present invention.
[0055] Figure 5 This is the final result diagram of the method of the present invention. DETAILED DESCRIPTION
[0056] In the following description, a large number of specific details are provided to provide a more thorough understanding of the present invention. However, it is apparent to those skilled in the art that the present invention can be implemented without one or more of these details. In other examples, in order to avoid confusion with the present invention, some technical features known in the art are not described.
[0057] Embodiment 1:
[0058] This embodiment provides a zero-sample pond water surface extraction method using multi-resolution remote sensing images, including the following technical solutions:
[0059] Step 1: Data preparation and preprocessing. Collect medium-resolution and high-resolution remote sensing images of the area of interest, preprocess the remote sensing images; annotate the pond water bodies on the high-resolution images through human-computer interaction, generate high-resolution pond water body true value labels, and resample to medium resolution to generate medium-resolution pond water body true value labels;
[0060] Step 2: Extract spectral feature set and obtain preliminary results of water bodies in the center-segmented image. Calculate the normalized water index NDWI of the area of interest; use the spectral reflectance set of a specific band to calculate the normalized difference spectral vector NDSV; set a threshold based on the normalized water index NDWI to separate surface water bodies from background objects, binarize NDWI according to the threshold, and obtain preliminary results of water bodies in the center-segmented image;
[0061] Step 3: Fine extraction of water bodies in the median-resolution image. Use the four-connected region method to mark the preliminary results of water bodies in the median-resolution image and obtain water patches; input the normalized water index NDWI, normalized difference spectral vector NDSV and near-infrared band as clustering features, use the K-Means algorithm to cluster the pixels in the water patches according to the clustering features, set the maximum number of clustering categories, and evaluate the best clustering results for each patch based on the Calinski-Harabasz index; set rules to evaluate and select the patch clustering results for each category to achieve more accurate water body extraction results from the median-resolution image;
[0062] Step 4: Generate pond water surface samples from the center-resolution image. The object-oriented method extracts the geometric feature set of water patches from the water body extraction results of the center-resolution image; a classification decision tree is constructed based on the feature set to select the qualified water patches as pond samples and non-pond samples. The pond samples and non-pond samples are combined into a training sample set, and the classifier is trained using the training sample set to obtain the center-resolution image pond water body sample labels, and the center-resolution image pond water body sample labels are evaluated for accuracy based on the center-resolution pond water body true value labels;
[0063] Step 5: Sample cross-scale migration. The multi-scale segmentation algorithm segments the high-resolution remote sensing image to obtain object units, and resamples the labels of pond water samples in the medium-resolution image to high resolution; set the sample cross-scale migration rule to judge each object unit, migrate the pond water samples in the medium-resolution image to high resolution, and obtain the labels of pond water samples in the high-resolution image;
[0064] Step 6: Convolutional neural network model training. Synchronously crop the high-resolution image pond water sample labels and the corresponding high-resolution remote sensing images, set the label extraction rules for scene-by-scene judgment, and obtain the scene label data set. Send it to the convolutional neural network classification model for model training and parameter iterative optimization. After the training is completed, the high-resolution remote sensing image is sent to the trained model to predict the high-resolution pond water extraction results, and the results are evaluated for accuracy based on the high-resolution pond water true value labels.
[0065] Embodiment 2:
[0066] Based on the aforementioned embodiment 1, this embodiment optimizes the above technical solution and discloses some specific measures.
[0067] Step 1 includes the following sub-steps:
[0068] Step 1.1: Image preprocessing steps include radiometric calibration, atmospheric correction, unified coordinate system, geo-registration, and image cropping.
[0069] Step 1.2: Resample the true value labels from high resolution to medium resolution, use the nearest neighbor assignment method to obtain the true value labels of the medium resolution pond water bodies, and adjust the image edge rows and columns to ensure that the number of rows and columns after label resampling is consistent with the number of rows and columns of the medium resolution remote sensing image.
[0070] Step 2 includes the following sub-steps:
[0071] Step 2.1: Normalize the reflectance of the green band and near-infrared band of the medium-resolution remote sensing image to calculate the normalized water index NDWI. The calculation method of the pixel normalized water index NDWI is as follows:
[0072]
[0073] In the formula, is the surface reflectance of the pixel in the green light band, is the surface reflectance of the pixel in the near-infrared band.
[0074] The specific band spectral reflectance set refers to a set of specific band spectral reflectances with consistent spatial resolution that is selected to meet experimental requirements based on the resolution of each band of the medium-resolution image.
[0075] Step 2.2: Normalized difference spectral vector NDSV, the specific calculation formula is as follows:
[0076]
[0077]
[0078] In the formula, ; ; is the number of bands; and is the band reflectivity. For the Band reflectivity and The normalized difference in reflectance between the bands.
[0079] The spectral feature set refers to the set of normalized water index NDWI and normalized difference spectral vector NDSV calculated based on the spectral reflectance of medium-resolution remote sensing bands.
[0080] Step 3 includes the following sub-steps:
[0081] Step 3.1: Based on the clustering characteristics, the K-Means algorithm is used to traverse each water patch in turn, cluster the pixels in each water patch, and obtain the patch clustering results. The K-Means clustering algorithm uses the Euclidean distance between the pixel cluster feature vector and the cluster center feature vector as a similarity measure, and the patch square error sum SSE as the objective function to measure the clustering quality, and seeks the best clustering result by finding the minimum objective function. The Euclidean distance calculation formula between the pixel object in the patch and the cluster center is as follows:
[0082]
[0083] In the formula, is the current pixel object, is the category number of this class, For the Cluster centers, is the pixel cluster feature number, is the number of clustering features of the pixel, for No. Clustering features, for No. Clustering features;
[0084] The calculation formula of the error sum of squares SSE of the entire patch is as follows:
[0085]
[0086] In the formula, is the category number of this class, Indicates the preset number of clusters, is the current pixel object, For the Cluster center.
[0087] Step 3.2: Select the Calinski-Harabasz index to evaluate the clustering results and select the optimal number of clusters block by block.
[0088] The Calinski-Harabasz index is used as an evaluation indicator to measure the clustering results. The specific calculation formula is as follows:
[0089]
[0090] In the formula, is the trace of the matrix, is the number of all pixels in the patch, is the number of classification categories; is the between-class covariance matrix, is the intra-class data covariance matrix, and the specific calculation formula is as follows:
[0091]
[0092] In the formula, is the category number of this class, Indicates the preset number of clusters, is the number of all pixels in the patch, Indicates The mean vector of clustering features of class pixels, is the cluster feature mean of all pixels in the patch, is the transpose of the matrix, is the current pixel clustering feature vector; the Calinski-Harabasz index essentially calculates the ratio of the distance between clustering features between classes and within classes. The larger the value, the tighter the class itself is and the more dispersed the classes are, which means a better clustering result.
[0093] Step 3.3: Set rules to evaluate and recommend by category. Set thresholds based on the area of each water patch and the average NDWI of the water patch to eliminate classes that do not meet the conditions.
[0094] Step 4 includes the following sub-steps:
[0095] Step 4.1: The geometric features included in the geometric feature set are: area, perimeter, minimum enclosing rectangle area / pattern area (B1), minimum enclosing rectangle length / minimum enclosing rectangle width (B2), compactness, area / perimeter.
[0096] Step 4.2: The SVM classifier constructs an optimal hyperplane to maximize the distance between any categories in the sample space, thereby achieving the distinction between different categories. In SVM, the optimal hyperplane can be achieved by setting different kernel functions, selecting kernel scaling size, regularization parameters, etc. to achieve classification and obtain classification results.
[0097] Step 4.3: Based on the true value labels of the middle-resolution pond water bodies, the accuracy of the middle-resolution image pond water body sample labels is evaluated using four indicators: accuracy, precision, recall, and Kappa coefficient.
[0098] Step 5 includes the following sub-steps:
[0099] Step 5.1: Use the image multi-scale segmentation algorithm to extract high-resolution remote sensing image object units, adjust the iterative shape factor and compactness factor parameters, and obtain the optimal segmented object units;
[0100] Step 5.2: Resample the sample labels of pond water bodies in the medium-resolution image to high resolution, use the nearest neighbor assignment method to obtain the sample labels of pond water bodies in the high-resolution image, and adjust the rows and columns of the image edge to ensure that the number of rows and columns after label resampling is consistent with the number of rows and columns in the medium-resolution remote sensing image;
[0101] Step 5.3: Separate pond water sample labels and non-pond water sample labels based on the pond water pixel ratio parameter of the object unit, and obtain the pond water sample labels of the high-resolution image.
[0102] Step 6 includes the following sub-steps:
[0103] Step 6.1: Set the size of the cropped image, set the label extraction rules according to the parameters of the pond water pixel ratio in the scene, label the cropped scene according to the rules, and obtain the scene label dataset.
[0104] Step 6.2: Based on the true value labels of high-resolution pond water bodies, the predicted high-resolution pond water body extraction results are evaluated for accuracy using five indicators: accuracy, precision, recall, balanced F-score, and intersection-over-union ratio.
[0105] Embodiment 3:
[0106] In combination with the above embodiment and embodiment 2, the study area of this embodiment is a coastal area in Yancheng City, Jiangsu Province. The medium-resolution remote sensing image data selects Sentinel-2 satellite images, and the high-resolution remote sensing image data selects 2-meter resolution image data provided by a domestic satellite. The images are acquired in September 2020, and the image geographic coordinate systems are CGCS2000_3_Degree_GK_CM_120E.
[0107] This embodiment is based on a zero-sample pond water surface extraction method based on a combined multi-resolution remote sensing image. The overall flow chart of the method is as follows: Figure 1 As shown, the following steps are included:
[0108] Step 1: Use Google Earth Engine (GEE) to obtain the atmospherically and orthorectified reflectance data (Level-2A) of Sentinel-2. The directory of this product in GEE is “COPERNICUS / S2_SR”. Perform preprocessing such as cropping on Level-2A products and domestic satellite high-resolution images, annotate pond water bodies through human-computer interaction, generate high-resolution pond water body true value labels, and resample to medium resolution to generate medium-resolution pond water body true value labels.
[0109] Step 1.1: Use the "geo-referencing" tool provided by ArcMap to correct the medium-resolution remote sensing image to the high-resolution remote sensing image to ensure that the objects in the two images overlap. Generate a clipping range vector file in ArcMap and clip the image to obtain multi-resolution remote sensing image data of the study area.
[0110] Step 1.2: Use the ArcMap editor function to manually annotate the pond water body features in the high-resolution remote sensing image. Ensure the edge accuracy as much as possible during the annotation process, generate the SHP file of the real label surface element of the pond water body, and save it as a TIFF file.
[0111] Step 1.3: In the "raster resampling" function provided by ArcMap software, select the nearest neighbor assignment method to resample the true value label from high resolution to medium resolution; and program in Matlab to adjust the rows and columns of the image edge to ensure that the two labels completely overlap and the number of rows and columns after label resampling is consistent with the number of rows and columns of the medium-resolution remote sensing image.
[0112] Step 2: Calculate the normalized water index NDWI in ENVI software, and program the calculation of the normalized difference spectral feature vector NDSV in Matlab; binarize the normalized water index NDWI to obtain the preliminary results of the water body in the center-divided image.
[0113] Step 2.1: Use the "Band Math" function provided in ENVI to calculate the normalized water index NDWI and enter the calculation formula:
[0114]
[0115] In the formula, is the green band reflectivity (the third band) of the Sentinel-2 image. is the near-infrared reflectivity of Sentinel-2 images (band 8); Indicates keeping the band data format as floating point.
[0116] Step 2.2: Use Matlab programming to calculate the normalized difference spectral vector NDSV. Select the 10-meter resolution bands of the Sentinel-2 image, namely the second band, the third band, the fourth band, and the eighth band. Calculate the normalized difference between the selected bands to form the normalized difference spectral vector NDSV.
[0117] Step 2.3: In the "New Raster Color Slice" function provided by ENVI, adjust the threshold to separate the water objects and other objects to achieve the best separation effect, and generate the preliminary result of the water body in the center-divided image.
[0118] Step 3: Use the four-connected region method to mark the preliminary results of the water body in the median image and obtain water patches; input the normalized water index NDWI, the normalized difference spectral vector NDSV and the eighth band of the Sentinel-2 image as clustering features, use the K-Means algorithm to cluster the pixels in the water patches according to the above three clustering features, set rules to evaluate and select the patch clustering results one by one, and achieve more accurate water body extraction results in the median image;
[0119] Step 3.1: Call the bwlabel function in Matlab to obtain preliminary results of the four-connected component labeling of water bodies, set the maximum distance category number to 5, and call the K-Means function in a loop to cluster each water patch according to the category number 1-5; call the evalclusters function and set "Calinski-Harabasz" as the clustering evaluation criterion, and evaluate each patch to obtain the optimal number of clusters.
[0120] Step 3.2: This implementation plan sets the category-by-category evaluation and selection rules as follows: Calculate the area threshold of water patches and the normalized difference water index NDWI threshold. The area threshold calculation formula is as follows:
[0121]
[0122] In the formula, is the area threshold of the water patch, is the area of the water patch;
[0123] Secondly, the normalized water index NDWI threshold is calculated. The specific calculation formula is as follows:
[0124]
[0125] In the formula, is the normalized water index NDWI threshold of this block, is the average value of the normalized water index NDWI of all pixels in the water patch;
[0126] The average values of the area and the normalized water index NDWI of each pixel in the patch are calculated one by one. The rule is expressed as follows:
[0127]
[0128] In the formula, For this type of area, is the average value of the normalized water index NDWI of this class; if the above two equations are both true, then this class is retained, otherwise the pixels of this class are eliminated.
[0129] Step 4: Mark the water body extraction results of the median-resolution image with four-connected regions to obtain water body patches; extract the patch geometric feature set, and build a classification decision tree based on the feature set to obtain training samples; classify the median-resolution image pond water body sample labels using the SVM classifier, and evaluate the accuracy of the results.
[0130] Step 4.1: In this implementation, the geometric features selected are: area, perimeter, minimum enclosing rectangle area feature (B1), minimum enclosing rectangle aspect ratio (B2), compactness, and area / perimeter. The area and perimeter attributes can be obtained by calling the regionprops function in Matlab, the B1 and B2 attributes are implemented by programming in Matlab, and the compactness attribute and area / perimeter attribute are obtained by calling the area and perimeter attribute numerical calculations. The specific formulas are as follows:
[0131]
[0132]
[0133]
[0134]
[0135] In the formula, is the minimum circumscribed rectangle area feature, is the minimum enclosing rectangle aspect ratio, is the compactness value, is the area / perimeter value, is the minimum circumscribed rectangle area, is the area of the patch, is the length of the longer side of the minimum circumscribed rectangle, is the length of the shorter side of the minimum circumscribed rectangle, is the perimeter attribute of the patch;
[0136] Step 4.2: Based on the geometric feature set proposed in step 4.1, construct a decision tree. The specific construction framework is shown in Figure 2 .
[0137] Step 4.3: This experimental scheme selects the SVM classifier to train the training sample data set obtained in step 4.2. Use the templateSVM function in Matlab to create a support vector machine model, select the Gaussian kernel function as the classification kernel function, select the kernel function scaling size 0.1, and the regularization parameter 100; call the fitcecoc function to train the SVM learner, and call the predict function to predict all samples to obtain the labels of the pond water samples in the middle-resolution image.
[0138] Step 4.4: Based on the true value labels of the middle-resolution pond water bodies, the middle-resolution image pond water body sample labels obtained in step 4.3 are evaluated for accuracy using four indicators: accuracy, precision, recall, and Kappa coefficient.
[0139] Step 5: Use the multi-scale segmentation method to segment the high-resolution remote sensing image, obtain the object unit, and resample the sample labels of the pond water body in the medium-resolution image in ArcMap. The operation is consistent with step 1.3 to ensure that the two images completely overlap; set the sample cross-scale migration rules and judge by object unit to obtain the sample labels of the pond water body in the high-resolution image;
[0140] Step 5.1: Use the ESP2 plug-in of the eCognition software to complete the multi-scale segmentation of the high-resolution remote sensing image. The EPS2 plug-in parameters are set to: shape factor is 0.1, compactness factor is 0.5, and the optimal segmentation result can be obtained. Export the segmentation result with the best distinction between pond water and edge information. In this experiment, the Level 3 object unit is selected and saved as a grayscale TIFF image;
[0141] Step 5.2: Resample the medium-resolution image pond water sample labels to obtain high-resolution image pond water sample labels, which is consistent with the method in step 1.3;
[0142] Step 5.3: The sample cross-scale migration rules designed in this implementation plan are shown in Figure 3 , the rules are now explained in detail: each object unit is compared with the sample label of pond water body in the high-resolution image. If the number of pond water body pixels in the object unit is greater than 40% of the total number of pixels, the object unit is judged as a pond water body and assigned a pond water body label; otherwise, it is judged as a non-pond and assigned a non-pond water body label.
[0143] Step 6: Crop the high-resolution image pond water sample labels and high-resolution remote sensing images according to the cropping size of 128 pixels × 128 pixels, set the label extraction rules to determine the labels scene by scene, and obtain the scene label data set. Send it to the SegNet training network model for model training and parameter iterative optimization. After the training is completed, the high-resolution remote sensing image is sent to the trained model to predict the pixel-level classification results, and the results are evaluated for accuracy based on the high-resolution pond water body true value labels.
[0144] Step 6.1: Label the cropped scene according to the label extraction rules to obtain the scene label dataset. The rule settings are shown in Figure 4The specific rules of the implementation plan are set as follows: if the number of pond water pixels in the scene accounts for 10% or more of the total number of pixels in the scene, the scene is assigned a label of 1, that is, a pond water sample label; if the number of pond water pixels in the scene accounts for 0% of the total number of pixels in the scene, the scene is assigned a label of 0, that is, a non-pond water sample label; if the number of pond water pixels in the scene accounts for 0%~10% of the total number of pixels in the scene (excluding 0% and 10%), the scene is eliminated.
[0145] Step 6.2: Based on the high-resolution true value labels of pond water bodies, the predicted pixel-level classification results are evaluated for accuracy using five indicators: accuracy, precision, recall, balanced F-score, and intersection-over-union ratio.
[0146] Figure 5 What is shown are examples of images and label diagrams at each stage using the technical means disclosed in the above embodiments. Figure 5 (a) shows a high-resolution remote sensing image; Figure 5 (b) shows the high-resolution ground truth label map of pond water bodies; Figure 5 (c) shows the label of the middle-divided pond water sample; Figure 5 (d) in the figure shows the high-resolution pond water extraction results.
[0147] As described above, although the present invention has been shown and described with reference to specific preferred embodiments, it should not be construed as limiting the present invention itself. Various changes may be made to it in form and detail without departing from the spirit and scope of the present invention as defined in the appended claims.
Claims
1. A zero-sample pond water surface extraction method using multi-resolution remote sensing images, characterized in that: The steps include: Step 1: collect medium-resolution remote sensing images and high-resolution remote sensing images of the area of interest and preprocess them; mark the pond water bodies on the preprocessed high-resolution remote sensing images, generate high-resolution pond water body true value labels, and resample the high-resolution pond water body true value labels to medium-resolution pond water body true value labels; Step 2, calculate the normalized water index NDWI of the medium-resolution remote sensing image in the area of interest; use the predetermined band spectral reflectance set to calculate the normalized difference spectral vector NDSV; set a threshold based on the normalized water index NDWI to separate the surface water body from the background objects, binarize the normalized water index NDWI according to the threshold, and obtain the preliminary result of the medium-resolution image water body; Step 3, marking the preliminary results of the water body in the center-divided image to obtain water body patches; inputting the normalized water body index NDWI, the normalized difference spectral vector NDSV and the near-infrared band as clustering features, using the K-Means algorithm to cluster the pixels in the water body patches according to the clustering features, and extracting the water body extraction results of the center-divided image; Step 4, extracting a geometric feature set of water body patches from the water body extraction result of the center-divided image; constructing a classification decision tree based on the geometric feature set, and selecting the water body patches that meet the conditions as pond samples and non-pond samples; forming a training sample set with the pond samples and non-pond samples, training a classifier using the training sample set to obtain the center-divided image pond water body sample labels, and performing accuracy evaluation on the center-divided image pond water body sample labels according to the center-divided pond water body true value labels; Step 5: Segment the high-resolution remote sensing image to obtain object units, resample the medium-resolution image pond water sample labels to high resolution; set sample cross-scale migration rules to judge by object units, migrate the medium-resolution image pond water sample to high resolution, and obtain the high-resolution image pond water sample labels; Step 5.1, extract high-resolution remote sensing image object units using an image multi-scale segmentation algorithm, adjust iterative shape factor and compactness factor parameters, and obtain the optimal segmented object units; Step 5.2, resample the pond water sample labels in the medium-resolution image to high resolution, use the nearest neighbor assignment method to obtain the pond water sample labels in the high-resolution image, adjust the edge rows and columns of the pond water sample labels in the high-resolution image, and ensure that the number of rows and columns of the pond water sample labels in the resampled high-resolution image is consistent with the number of rows and columns in the medium-resolution remote sensing image; Step 5.3, according to the object unit pond water pixel ratio parameter, separate the pond water sample label and the non-pond water sample label, and obtain the high-resolution image pond water sample label; Step 6. Synchronously crop the high-resolution image pond water sample labels and the corresponding high-resolution remote sensing images, set the label extraction rules for scene-by-scene judgment, and obtain the scene label data set; send the scene label data set to the convolutional neural network classification model for model training and parameter iterative optimization; send the high-resolution remote sensing image to the trained convolutional neural network classification model to predict the high-resolution pond water extraction results, and evaluate the accuracy of the results based on the high-resolution pond water true value labels.
2. The zero-sample pond water surface extraction method based on multi-resolution remote sensing images according to claim 1 is characterized in that: In step 1, medium-resolution and high-resolution remote sensing images of the area of interest are collected and preprocessed, including: radiation calibration, atmospheric correction, unified coordinate system, geographic registration, and image cropping.
3. The zero-sample pond water surface extraction method based on multi-resolution remote sensing images according to claim 1 is characterized in that: In step 1, the high-resolution pond water body true value label is resampled to the medium-resolution pond water body true value label, which meets the following conditions: Adjust the marginal rows and columns of the medium-resolution pond water body true value labels to ensure that the number of rows and columns of the resampled medium-resolution pond water body true value labels is consistent with the number of rows and columns of the original medium-resolution remote sensing image.
4. The zero-sample pond water surface extraction method based on joint multi-resolution remote sensing images according to claim 1 is characterized in that: Step 2 specifically includes: Step 2.1: Normalize the reflectance of the green band and near-infrared band of the medium-resolution remote sensing image to calculate the normalized water index NDWI: ; In the formula, is the surface reflectance of the pixel in the green light band, is the surface reflectance of the pixel in the near-infrared band; Step 2.2: Calculate the normalized difference spectral vector NDSV using the set of spectral reflectances of the predetermined band: ; ; In the formula, ; ; is the number of bands; and is the band reflectivity; For the Band reflectivity and The normalized difference in reflectance between the bands.
5. The zero-sample pond water surface extraction method of joint multi-resolution remote sensing images according to claim 1 is characterized in that: Step 3 specifically includes: Step 3.1, traverse each water patch in turn through the K-Means algorithm according to the clustering characteristics, cluster the pixels in each water patch, and obtain the patch clustering results; The K-Means clustering algorithm uses the Euclidean distance between the pixel cluster feature vector and the cluster center feature vector as a similarity measure, and the squared error SSE of the patch as the objective function to measure the clustering quality. It seeks the best clustering result by finding the minimum objective function. Step 3.2, select the Calinski-Harabasz index to evaluate the clustering results, and select the optimal number of clusters block by block; the Calinski-Harabasz index calculation formula is as follows: ; In the formula, is the trace of the matrix, is the number of all pixels in the patch, is the number of classification categories; is the between-class covariance matrix, is the intra-class data covariance matrix; Step 3.3: Set the threshold according to the area of each water patch and the average value of the normalized water index NDWI of the water patch to eliminate the classes that do not meet the conditions.
6. The zero-sample pond water surface extraction method based on multi-resolution remote sensing images according to claim 5 is characterized in that: In step 3.1, the Euclidean distance calculation formula between the pixel object in the patch and the cluster center is as follows: ; In the formula, is the current pixel object, is the category number of this class, For the Cluster centers, is the pixel cluster feature number, is the number of clustering features of the pixel, for No. Clustering features, for No. Clustering features; The formula for calculating the sum of squared errors (SSE) of the entire patch is as follows: ; In the formula, is the category number of this class, Indicates the preset number of clusters, is the current pixel object, For the Cluster center.
7. The zero-sample pond water surface extraction method based on multi-resolution remote sensing images according to claim 5 is characterized in that: The between-class covariance matrix and the intra-class data covariance matrix The calculation formula is as follows: ; In the formula, is the category number of this class, Indicates the preset number of clusters, is the number of all pixels in the patch, Indicates The mean vector of clustering features of class pixels, is the cluster feature mean of all pixels in the patch, is the transpose of the matrix, is the current pixel clustering feature vector.
8. The zero-sample pond water surface extraction method based on multi-resolution remote sensing images according to claim 1, characterized in that: In step 4, a geometric feature set of water body patches is extracted from the water body extraction result of the center-divided image, and the geometric features included in the geometric feature set are: area, perimeter, minimum enclosing rectangle area / patch area, minimum enclosing rectangle length / minimum enclosing rectangle width, compactness, and area / perimeter.
9. The zero-sample pond water surface extraction method of joint multi-resolution remote sensing images according to claim 1 is characterized in that: Step 6 specifically includes: Step 6.1, set the size of the cropped image, set the label extraction rules according to the parameters of the pond water pixel ratio in the scene, assign labels to the cropped scene according to the rules, and obtain the scene label data set; Step 6.2: Based on the true value labels of high-resolution pond water bodies, the predicted high-resolution pond water body extraction results are evaluated for accuracy using five indicators, including accuracy, precision, recall, balanced F-score, and intersection-over-union ratio.
