A large-scale fast remote sensing water body extraction method and system based on deep learning

Abstract

The application belongs to the technical field of information technology services, and discloses a large-range fast remote sensing water body extraction method and system based on deep learning, which realizes fast and large-range water body extraction by fusing deep learning and traditional machine learning. The application has the high efficiency of traditional machine learning and the deep feature extraction of deep learning, and has great innovation in data preparation, network framework optimization and image clustering post-processing. The application firstly preliminarily determines the pixel clustering range by using traditional machine learning, then further clusters the pixels in the range by using a convolutional neural network layer, and finally obtains the water body extraction result by clustering and binarization. The application has strong applicability and can be used for various practical needs including flood mapping, surface water extraction, shoreline change monitoring and the like, and can meet the requirements of fast, large-range and high-precision water body extraction.

Description

Technical Field

[0001] This invention belongs to the field of information technology service technology, and in particular relates to a method for large-scale and rapid remote sensing water body extraction based on deep learning. Background Technology

[0002] Against the backdrop of rising global sea levels and drastic changes in coastlines, coastal populations face increasingly severe flood disasters, with nearly 23% of the world's population directly exposed to once-in-a-century floods. Therefore, rapid and large-scale flood monitoring is of paramount importance, concerning human property and life safety. Surface water extraction, including flood extraction, has seen continuous development in recent years. Traditional surface water mapping methods, through manual measurement, can accurately determine the extent of surface water bodies, but their efficiency is very low, requiring significant manpower and resources, and their applicability is limited. Water extraction methods based on remote sensing and traditional machine learning can be divided into supervised and unsupervised classification. Unsupervised classification methods include common threshold classification methods such as maximum inter-class variance and minimum error segmentation. These methods have high segmentation efficiency and can quickly extract large areas of water bodies using SAR imagery. However, water extraction methods relying on a single threshold also have significant drawbacks. For areas with complex land cover types, SAR backscattering is strong, complex, and variable, resulting in poor classification details when using a single threshold. Supervised classification methods, including support vector machines and random forests, often offer improved accuracy compared to traditional threshold segmentation methods. However, they are less efficient at extracting water bodies and require significant human effort to select training samples. Traditional machine learning has very limited capabilities in handling speckle noise in SAR images, typically relying on filtering methods to reduce this noise.

[0003] In recent years, deep learning has developed rapidly and has been widely applied in various fields, including natural language recognition, object recognition, and natural image classification. Deep learning can extract deep features from images, which has significant advantages over traditional machine learning. Currently, many studies have applied deep learning to the classification of multispectral and hyperspectral images, but its application to SAR image classification is still relatively limited. Supervised deep learning often relies on massive amounts of training data for good classification results, requiring significant human and material resources to create training labels. This is a major drawback for water body extraction problems that prioritize timeliness, such as flood mapping. Unsupervised deep learning methods have enormous potential for rapid and large-scale flood monitoring and represent an important research direction for flood extraction.

[0004] Based on the above analysis, the existing technologies have the following problems and shortcomings: the original surface water mapping method is very inefficient and requires a lot of manpower and resources, and its applicability is not strong; the water body extraction method based on remote sensing and traditional machine learning has high segmentation efficiency, but it is poor in classification details for areas with complex land cover types; supervised classification methods have a certain improvement in accuracy compared with traditional threshold segmentation methods, but the water body extraction efficiency is low and requires manpower to select training samples; traditional machine learning has very limited ability to deal with speckle noise in SAR images, and generally reduces speckle noise through filtering methods. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a method for large-scale and rapid remote sensing water body extraction based on deep learning.

[0006] This invention is implemented as follows: a method for large-scale, rapid remote sensing water body extraction based on deep learning, comprising:

[0007] The preprocessed image is segmented into training images that meet the computing power requirements of the computer. In the early stage of convolutional clustering training, the pixel features of the original training image are fused together, and the segmentation labels of superpixels are combined to preserve as much important superpixel block information as possible in the clustering feature map;

[0008] When training on training images with extremely uneven proportions of water and non-water bodies, two hyperparameters are used to prevent the loss of water or non-water body pixel blocks as the number of training iterations increases.

[0009] The clustered images typically contain 34 categories. Using the pixel features of the original training images, a reference threshold is obtained using the traditional threshold segmentation method. The clustered images are then binarized to finally generate a water-non-water body classification map.

[0010] Another object of the present invention is to provide a large-scale rapid remote sensing water body extraction system based on deep learning, the large-scale rapid remote sensing water body extraction system based on deep learning comprising:

[0011] The data preprocessing module performs orbit correction and thermal noise removal on the raw input data; after radiometric calibration, the processed data is filtered and topographically corrected; the corrected data is converted to dB units for image cropping.

[0012] The convolutional network clustering module generates segmentation labels after the original image is segmented by superpixels. The original image is then processed by three convolutional layers to output feature maps. The feature maps are further processed by original image fusion, segmentation label clustering, and two hyperparameter constraint steps to generate clustering feature maps. Finally, the clustering feature maps and feature maps are used to generate a loss function, which is then used to continuously optimize the parameters of each node in the network through backpropagation.

[0013] The clustering binarization module combines the pixel intensity information in the original image with the clustering results output by the network to obtain a clustering result image that best reflects the features of the original image.

[0014] Furthermore, after preprocessing, the intensity data is converted into dB data using the following formula:

[0015] dB = 10 * log10(I);

[0016] Where I is the preprocessed pixel intensity, and dB is the decibel data generated by the conversion.

[0017] Furthermore, the preprocessed image unit is dB, and the data type is 32-bit floating-point; before the convolutional neural network clusters the images, the data type is converted to unsigned 8-bit integer; before the data type conversion, all training images are linearly stretched by 1%, and all images are stretched by the same range.

[0018] Furthermore, in the early stages of training, low-dimensional features need to be preserved by "fusion of input images." As the number of training iterations increases, some small areas of water in the images will gradually disappear as low-dimensional features, thus reducing the final clustering accuracy. The original image and the feature map output from the convolutional layer are then fused to obtain a new feature map containing both feature information and information from the original image. The fusion method involves weighting the original image and the feature map according to a certain ratio. As the number of training iterations increases, the weight of the original image gradually decreases, while the weight of the feature map gradually increases.

[0019]

[0020] Where A represents the pixel value of the "new feature map" after fusing the original image, n represents the number of training iterations, B represents the pixel value corresponding to the input image, and C represents the pixel value corresponding to the output feature map of the convolutional layer.

[0021] Furthermore, the newly generated feature map after fusion still needs to undergo segmentation label processing.

[0022] Another object of the present invention is to provide a computer device, the computer device including a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform the steps of the data synchronization optimization method between serially connected chips.

[0023] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the data synchronization optimization method between serially connected chips.

[0024] Another objective of this invention is to provide an information data processing terminal for implementing the data synchronization optimization system between the serially connected chips.

[0025] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:

[0026] First, it combines the efficiency of traditional threshold segmentation methods with the deep feature extraction capabilities of deep learning. In the early training stage, by fusing pixel information features from the original image, it retains small superpixel feature information to a certain extent, thus reducing the error of missing segments.

[0027] The deep learning framework was optimized by using two hyperparameter constraints to overcome the problem of water or non-water regions being lost during clustering due to the extremely uneven proportion of water and non-water bodies.

[0028] The final water body extraction result map is generated through a unique clustering and binarization operation. This process is basically unsupervised, and only a small amount of human supervision is required in SAR images taken under extreme weather conditions.

[0029] Second, this invention first segments the preprocessed image into training images that meet the computing power requirements. For training images with highly uneven proportions of water and non-water bodies, water or non-water body pixel blocks are prone to being lost as the number of training iterations increases. This framework solves this problem in two ways: First, in the early stage of network training, the pixel features of the original image are used, combined with superpixel segmentation labels, to constrain the features of superpixel blocks and suppress feature vanishing. Second, two hyperparameters are used to constrain the pixel value vanishing at both endpoints. The clustered image generally contains 34 classes. This invention uses the pixel features of the original training image, combined with a reference threshold generated by the traditional threshold segmentation method, to binarize the clustered image, ultimately generating a water-non-water body classification map.

[0030] Third, as supplementary evidence of the inventive step of the claims of this invention, it is also reflected in the following important aspects:

[0031] (1) The expected benefits and commercial value of the technical solution of this invention after transformation are as follows: This invention can be deployed on a cloud platform, such as AI Earth. With the help of the cloud platform, an integrated process including data download and preprocessing, convolutional network clustering, and post-clustering binarization can be realized. Based on this, it can quickly realize emergency response to floods and can also be used for long-term, large-scale monitoring of changes in surface waters.

[0032] (2) The technical solution of the present invention fills the technical gap in the industry at home and abroad: There are already articles on deep learning to extract water bodies at home and abroad, but the innovation of this method is that by optimizing the network structure, its accuracy and stability are improved, making it able to adapt to the complex surface water extraction needs.

[0033] (3) Does the technical solution of this invention overcome technical bias? With the widespread adoption of deep learning in various fields, many deep learning methods are considered to "blindly utilize deep learning without combining it with specific needs." However, this method differs significantly from traditional supervised deep learning methods. Through innovation in the structure of convolutional neural networks, it focuses on leveraging the deep feature extraction capabilities of convolutional layers, integrating traditional threshold segmentation and various constraints, aiming at rapid extraction of large-scale surface water areas. This achieves a breakthrough in the accuracy and stability of the framework. The results show that this method can effectively reduce speckle noise in SAR images. Attached Figure Description

[0034] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This is a flowchart of the Sentinel-1 preprocessing provided in an embodiment of the present invention;

[0036] Figure 2 This is a diagram of the convolutional neural network structure provided in an embodiment of the present invention;

[0037] Figure 3 This is a comparative schematic diagram of different superpixel segmentation methods provided in the embodiments of the present invention, wherein (a) is a schematic diagram of the Felzenszwalb segmentation method, (b) is a schematic diagram of the SLIC segmentation method, (c) is a schematic diagram of the Quik shift segmentation method, and (d) is a schematic diagram of the Watershed segmentation method.

[0038] Figure 4 This is a schematic diagram of segmentation label processing provided in an embodiment of the present invention, wherein (A) is a feature map, (B) is a segmentation label, (C) is the mean method, and (D) is the highest frequency method;

[0039] Figure 5 This is a schematic diagram of hyperparameter constraints provided in an embodiment of the present invention;

[0040] Figure 6 This is a backscattering intensity frequency statistics diagram provided in an embodiment of the present invention;

[0041] Figure 7 This is a diagram showing the verification area and sample point distribution provided in an embodiment of the present invention;

[0042] Figure 8 This is an error point distribution diagram provided in an embodiment of the present invention;

[0043] Figure 9 This is an SVM error point distribution diagram provided in an embodiment of the present invention;

[0044] Figure 10 This is an OTSU error point distribution diagram provided in an embodiment of the present invention;

[0045] Figure 11 This is a distribution diagram of MET error points provided in an embodiment of the present invention;

[0046] Figure 12 This is a distribution diagram of NN error points provided in an embodiment of the present invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0048] To address the problems existing in the prior art, the present invention provides a method, system, device, and terminal for optimizing data synchronization between serially connected chips. The present invention will be described in detail below with reference to the accompanying drawings.

[0049] This invention first segments the preprocessed image into training images that meet the computing power requirements. Based on the minimum error segmentation method, a clustering range for the deep learning framework is defined near the segmentation threshold, and the deep learning framework is used only to cluster pixels within this range. For training images with highly uneven proportions of water and non-water bodies, water or non-water body pixel blocks are easily lost as the number of training iterations increases. This framework addresses this problem through two hyperparameters. The clustered image typically contains 34 classes. This invention uses the pixel features of the original training image to binarize the clustered image, ultimately generating a water-non-water body classification map.

[0050] This paper uses Sentinel-1SAR as an example to illustrate convolutional neural networks. The proposed water extraction framework consists of three parts: data preprocessing, convolutional network clustering, and post-clustering binarization.

[0051] (1) Data preprocessing.

[0052] To extract non-water body pixels with low reflectivity, the network input is a non-linearly stretched dB-unit image. Since image quality is highly sensitive to thermal noise, thermal noise removal is a crucial step in data preprocessing. Extensive experiments have shown that VV-polarized Sentinel-1 data is far less affected by thermal noise than VH-polarized data; therefore, this paper uses Sentinel-1 VV-polarized data for water body extraction. The preprocessing steps are as follows: Figure 1 .

[0053] After preprocessing, the intensity data is converted into dB data, as shown in the formula below:

[0054] dB=10*log10(I)#(1)

[0055] Where I is the preprocessed pixel intensity, and dB is the decibel data generated by the conversion.

[0056] The preprocessed image unit is dB, and the data type is 32-bit floating-point. To facilitate image clustering by the convolutional neural network, the data type needs to be converted to unsigned 8-bit integer (Uint8). Before the data type conversion, all training images need to be linearly stretched by 1%. To ensure that the same pixel value in different images represents the same or similar land cover type, the stretching range of all images is the same. Taking this invention as an example, the image stretching range is mainly distributed in [-28.5dB, -6.7dB].

[0057] (2) Deep learning network training. The network structure is as follows: Figure 2 .

[0058] First, the original image (i.e., the input image) is segmented by superpixelation to generate segmentation labels. The original image then undergoes three convolutional layers to output feature maps. These feature maps are further processed through original image fusion, segmentation label clustering, and two hyperparameter constraints (i.e., limitations) to generate clustered feature maps. Finally, a loss function is generated using the clustered feature maps and the original feature maps, and the parameters of each node in the network are continuously optimized through backpropagation. The specific steps of the water extraction framework are described in detail below.

[0059] ① Early training requires preserving low-dimensional features through "input image fusion". As the number of training iterations increases, small areas of water in the images gradually disappear as low-dimensional features, thus reducing the final clustering accuracy. To preserve low-dimensional feature information in early training, this paper fuses the original image and the feature map output by the convolutional layer to obtain a "new feature map" containing both feature information and original image information. The fusion method involves weighting and summing the original image and the feature map according to a certain ratio. As the number of training iterations increases, the weight of the original image gradually decreases, while the weight of the feature map gradually increases. This paper performs fusion of the two types of images in the first 50 training iterations, as shown in Equation 2.

[0060]

[0061] Where A represents the pixel value of the "new feature map" after fusing the original image, n represents the number of training iterations, B represents the pixel value corresponding to the input image, and C represents the pixel value corresponding to the output feature map of the convolutional layer.

[0062] ②The "new feature map" generated after fusion still needs to be processed by "segmentation label".

[0063] Superpixels are the result of grouping several pixels with the same or similar attributes; they can also be understood as the result of image oversegmentation. Superpixels carry more information than pixels and can better align image edges. The purpose of generating superpixels is to effectively reduce data redundancy and lower the complexity of recognition tasks. Using superpixels as subsequent processing units can significantly accelerate data processing efficiency. Numerous studies have demonstrated that combining superpixel segmentation with SAR classification can effectively suppress the influence of speckle noise. The superpixel segmentation method used in this paper is Felz. Figure 3 The effects of different superpixel segmentation methods are demonstrated. In the comparison of segmentation results for SAR images, Quick Shift and Watershed perform relatively poorly, struggling to capture complete feature edge contours. SLIC performs well overall, but its performance is inferior to Felz at fine boundaries. The Felz segmentation method can extract fine feature edges relatively completely while requiring fewer superpixels, which is crucial for the clustering accuracy and efficiency of deep networks.

[0064] After superpixel segmentation, segmentation labels with the same dimensions as the input image can be obtained. These segmentation labels are then fused with the convolutional feature map to generate a new feature map. See [link to specific processing steps] for details. Figure 4The generated result image is generated by performing "mean" or "highest frequency" processing on the pixel block corresponding to each superpixel position of the superpixel segmentation label in the feature map. "Mean" processing replaces the pixel block with the average value of all pixels within it, while "highest frequency" processing replaces the pixel block with the highest frequency pixel value from the statistical results. The "mean" method is mainly used in the early "fusion" step to preserve feature information in small pixel blocks and prevent it from disappearing with increasing training iterations. The "highest frequency" method is used for further clustering operations after the "fusion" process. Extensive experiments show that the "highest frequency" method is faster than the "fusion" method for clustering.

[0065] ③ Hyperparameter Constraints. In water body extraction studies based on SAR imagery, the main method for distinguishing between water and non-water bodies is by reflectivity intensity. Water surfaces primarily undergo specular reflection, appearing as pixels with extremely low reflectivity in SAR images. Thresholding segmentation methods extract water bodies by binarizing the image using one or more threshold values. In SAR images, pixels with higher reflectivity are more likely to be non-water bodies, while pixels with lower reflectivity are more likely to be water bodies. The range between high-confidence water and non-water bodies is termed the "intermediate region" in this paper. The key to thresholding segmentation lies in understanding this "intermediate region."

[0066] The pixel blocks within the image are segmented. As mentioned above, the deep learning clustering method used in this invention, without applying additional constraints, can misclassify images with excessively high or low water content (significant omission error). To overcome this problem, the network introduces two hyperparameters to limit the disappearance of excessively low or high pixel value regions during the clustering process. See [link to details]. Figure 5 In Uint8 data, the pixel range is [0, 255], and the constraints are shown in Formula 3.

[0067] v n+1 =r×v n +(1-r)×v n+1 ,n<90 or n>160,r=0.05×|v n -v n+1 |#(3)

[0068] Among them, v n and v n+1 ... Figure 5 In the diagram, the red box indicates that water blocks with lower pixel values ​​are transformed into non-water blocks with higher pixel values ​​during clustering. The role of hyperparameters is to suppress this process. The green box represents the normal clustering process, which is not affected by hyperparameters.

[0069] Finally, the cross-entropy loss function is obtained by calculating the cross-entropy between cluster feature maps and feature maps.

[0070] (3) Binarization after clustering.

[0071] The clustering results output by convolutional neural networks often don't only contain two classes. To obtain a result image that only includes water and non-water body categories, binarization processing is required. First, the pixel intensity information from the original image is combined with the clustering results output by the network to obtain a clustering result image that best reflects the features of the original image. This process is similar to the highest frequency method mentioned above, but instead of using the highest frequency of the corresponding class in the original image to represent the corresponding class in the clustering image, a weighted average is used. In the clustering image, the backscattering intensity range covered by all pixels representing each class is very wide, and pixels with higher backscattering intensity frequencies best represent that class. If the highest frequency pixel has an absolute advantage in quantity, then using it directly to replace the pixels of that class is very appropriate. However, in some cases, the highest frequency cannot truly reflect the overall pixel value distribution, such as... Figure 6 The highest frequency pixel value is approximately 125, but there is another local highest frequency pixel at approximately 170, and the two are not significantly different in number. To obtain a value that reflects the overall backscattering intensity distribution of this type of pixel to represent it, this paper uses a weighted average to obtain this value, as shown in Formula 4. First, the weighting range is determined; here, only the weighted average within the first 60% of pixel values ​​is calculated. Figure 6 The orange-yellow range in the middle.

[0072]

[0073] The threshold segmentation method of this invention differs significantly from traditional threshold segmentation methods. The pixel value distribution range of the original image is [0, 255] (Uint8), and the threshold needs to segment approximately 256 pixel classes. More importantly, speckle noise in the image is difficult to remove. This invention clusters regions between high and low scattering intensities, resulting in a final cluster map with approximately four categories, effectively distinguishing water bodies from non-water bodies.

[0074] As a preferred embodiment, this invention combines a specific case of the Yellow River Delta for water extraction and accuracy assessment.

[0075] (1) Data processing.

[0076] Preprocessing of Sentinel-1 data was performed using the Python module Snappy from SNAP. This paper employs Range Doppler Terrain Correction to geocode radar geometric images, with the reference DEM being SRTM 3Sec (Shuttle Radar Topographic Mission 3second). After geocoding, the images are projected into the WGS 1984UTM Zone 50N coordinate system.

[0077] The preprocessed image needs to be cropped into a 2000×2000 pixel training image, which is then used to generate a clustering result image through a deep network training framework. The binarization process of this invention is relatively simplified. Extensive experiments have shown that for high-quality input images, selecting any threshold between 105 and 125 (Uint8 data type) can yield a relatively ideal classification result, offering a wide range of choices.

[0078] (2) Accuracy assessment.

[0079] ① Accuracy Evaluation Indicators. Accuracy evaluation for classification problems often employs confusion matrix calculations (Table 1). The confusion matrix is ​​a relatively basic, intuitive, and simple method for evaluating the accuracy of non-classification problems. This paper evaluates the accuracy of water extraction using the following indicators: Omission Error (OE), Commission Error (CE), Dice coefficient (DC), user accuracy, producer accuracy, and accuracy rate. Specific calculation formulas are shown in 5-10.

[0080] The reference results are from the high-resolution optical image Gaofen-1, which replaces the role of the ground truth labels: Missing error, representing the proportion of incorrectly classified non-water body sampling points to the proportion of sampling points whose ground truth label is water; Misclassification error, representing the proportion of incorrectly classified water body sampling points to the proportion of sampling points whose ground truth label is water; Dice coefficient, used to describe the similarity between the classification result and the ground truth label; Accuracy, describing the proportion of correctly classified sampling points to the total number of sampling points; Precision, the proportion of correctly classified water body sampling points to the total number of water body sampling points; Sensitivity, describing the proportion of all correctly classified water body sampling points to the total number of sampling points whose ground truth label is water; Specificity, describing the proportion of all non-water body sampling points to the total number of sampling points whose ground truth label is non-water body.

[0081] In the confusion matrix, TP (True Positive) represents the number of water body sample points in the reference results that are classified as water bodies, FN represents the number of water body sample points in the reference results that are classified as non-water bodies, FP represents the number of non-water body sample points in the reference results that are classified as water bodies, and TN represents the number of non-water body sample points in the reference results that are classified as non-water bodies.

[0082] Table 1 Confusion Matrix

[0083]

[0084] Table 2 shows a comparison of optical and SAR image parameters. As mentioned earlier, the Yellow River Delta coastal zone is greatly affected by tides. Under the influence of strong anthropogenic factors, the water body in this area is undergoing drastic dynamic changes. Therefore, the acquisition times of Sentinel-1 and Gaofen-1 images should be as close as possible to reduce errors caused by other factors. Due to the limited data sources of Gaofen-1 images, the validation scope of this paper mainly focuses on the coastal area of ​​the Yellow River Delta, such as... Figure 7 As shown. The Sentinel-1 image verification time was selected on March 18, 2019, and the Gaofen-1 images were selected on March 15 and March 19, 2019, with a time difference of less than one week.

[0085] Table 2. Parameters of Gaofen-1 and Sentinel-1

[0086]

[0087] ②Accuracy verification results.

[0088] Classification error points were identified in the sample points through manual visual interpretation. The distribution of these error points is shown in the figure. Figure 8 To clarify the difference in error accuracy between coastal and non-coastal areas, we statistically analyzed the error situation in different regions, as shown in Table 3. The overall error parameters are shown in Table 4.

[0089] Table 3. Statistics of Error Points in Coastal and Non-Coastal Areas

[0090]

[0091] Table 4 Confusion Matrix and Accuracy Calculation

[0092]

[0093]

[0094] Statistical data shows that the extraction accuracy for water bodies is higher in coastal areas (referring to...). Figure 8The accuracy of the extraction of non-water areas (in the yellow area, the same below) is 2.93 percentage points lower than that of non-coastal areas. The extraction accuracy of coastal areas is 4.34 percentage points lower than that of non-coastal areas. The overall extraction accuracy of water areas is 90.7%, lower than the overall extraction accuracy of land (96.4%). The extraction accuracy of coastal areas (93.29%) is lower than that of non-coastal areas (97.63%). In summary, the classification accuracy of the method presented in this paper can perform well in complex classification areas, but there are certain differences between water and non-water areas, and between coastal and non-coastal areas. The classification accuracy is relatively low in areas with dense artificial structures.

[0095] The method of this invention has a missing classification error of 9.3%, a misclassification error of 3.82%, a Dice coefficient of 93.36%, an accuracy rate (i.e., overall classification accuracy) of 93.55%, a user accuracy of 96.18% for water body extraction, and a producer accuracy of 90.7%.

[0096] (3) Accuracy comparison with other methods

[0097] To verify the accuracy advantage of our method, we utilized Support Vector Machine (SVM), Otsu's maximum inter-class difference method, Minimum Error Thresholding (MET), and Neural Net Classification (NN) implemented on the ENVI platform. Neural Net Classification was based on supervised classification's standard backpropagation, with a hidden layer of 1 and a learning rate of 0.2. The error point distributions of each classification method are shown below. Figures 9 to 12 The comparison information on water extraction accuracy is shown in Table 5 below. The method presented in this paper has significant advantages in both user accuracy and producer accuracy.

[0098] Table 5 Comparison of accuracy of different classification methods

[0099]

[0100] The following is a detailed signal and data processing procedure for a large-scale, rapid remote sensing water body extraction method based on deep learning:

[0101] 1) Preprocessing: Preprocess the original remote sensing images, including radiometric correction, atmospheric correction and other operations, to eliminate noise and non-water body information in the images.

[0102] 2) Image segmentation: The preprocessed image is segmented into training images that can meet the computing power requirements of the computer.

[0103] 3) Minimum error segmentation: Based on the minimum error segmentation method, the clustering range of the deep learning framework is defined near the segmentation threshold.

[0104] 4) Deep learning clustering: Using deep learning frameworks, such as convolutional neural networks (CNN) or autoencoders (AE), to cluster pixels within a clustering range.

[0105] 5) Hyperparameter adjustment: During training, for training images with extremely uneven proportions of water and non-water bodies, two hyperparameters are adjusted to prevent the loss of water or non-water body pixel blocks as the number of training iterations increases.

[0106] 6) Binarization: The clustered images are binarized based on the pixel features of the original training images. During binarization, water pixels are set to 1, and non-water pixels are set to 0.

[0107] 7) Generate a classification map: Combine the binarized images to generate a water-non-water body classification map.

[0108] The following is a detailed working principle of a large-scale, rapid remote sensing water body extraction system based on deep learning:

[0109] 1) Data Preprocessing Module: This module is responsible for processing the raw remote sensing data. First, orbit correction and thermal noise removal are performed. Next, radiometric calibration, filtering, and terrain correction are performed. Finally, the corrected data is converted to dB units and image cropping is performed to obtain the remote sensing image to be processed.

[0110] 2) Convolutional Network Clustering Module: This module utilizes deep learning (such as convolutional neural networks) for pixel clustering. First, the original image is segmented into superpixels, generating segmentation labels. Then, the original image undergoes three convolutional layers to output feature maps. Next, the feature maps are further processed through original image fusion, segmentation label clustering, and two hyperparameter constraint steps to generate clustered feature maps. Finally, a loss function is generated using the clustered feature maps and the original feature maps. The parameters of each node in the network are continuously optimized through backpropagation to minimize the loss function.

[0111] 3) Post-clustering binarization module: This module combines the pixel intensity information from the original image with the clustering results output by the convolutional network to obtain a clustering result image that best reflects the features of the original image. In this process, pixel intensity information helps to refine the clustering results and improve the accuracy of distinguishing between water bodies and non-water bodies.

[0112] The deep learning-based large-scale rapid remote sensing water body extraction system achieves efficient and accurate water body extraction through the collaborative work of the three modules mentioned above. The data preprocessing module ensures image quality, while the convolutional network clustering module utilizes deep learning technology for pixel clustering, improving the accuracy of water body detection. Finally, the post-clustering binarization module combines the clustering results with the original pixel intensity information, further improving the quality of the classification image. The entire system comprehensively utilizes remote sensing data processing technology and deep learning methods to improve the efficiency and accuracy of remote sensing water body extraction.

[0113] Through the above signal and data processing steps, large-scale and rapid remote sensing water body extraction based on deep learning can be achieved. This method utilizes the powerful feature extraction capabilities of deep learning to complete large-scale remote sensing water body extraction tasks in a short time, improving the accuracy and efficiency of water body detection.

[0114] Example 1:

[0115] In a large-scale remote sensing monitoring project for a reservoir, a deep learning-based method for large-scale, rapid remote sensing water body extraction was employed. First, a data preprocessing module performed trajectory correction, thermal noise removal, radiometric calibration, filtering, and terrain correction on the raw remote sensing data. The corrected data was then converted to dB units and cropped. Next, a convolutional network clustering module performed superpixel segmentation on the original image, generating segmentation labels. These labels were then processed through three convolutional layers to output feature maps. These feature maps were further processed through original image fusion, segmentation label clustering, and two hyperparameter constraints to generate clustered feature maps. Finally, a post-clustering binarization module combined the pixel intensity information from the original image with the clustering results output by the network to obtain a clustering result map that best reflects the characteristics of the original image. This method successfully extracted the reservoir's water body extent, providing accurate data support for reservoir management.

[0116] Example 2:

[0117] In a remote sensing monitoring project of an urban river, a large-scale, rapid remote sensing water body extraction method based on deep learning was adopted to extract river water bodies. First, the raw remote sensing data was preprocessed using a data preprocessing module, including orbit correction, thermal noise removal, radiometric calibration, filtering, and terrain correction. Then, a convolutional network clustering module was used to perform superpixel segmentation and clustering on the raw image, generating feature maps. These feature maps were further processed through raw image fusion, segmentation label clustering, and two hyperparameter constraints to generate clustered feature maps. Finally, a post-clustering binarization module combined the pixel intensity information from the original image with the clustering results output by the network to obtain a clustering result map that best reflects the characteristics of the original image. This method successfully extracted the river's water body extent, providing strong data support for urban river management.

[0118] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0119] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for large-scale, rapid remote sensing water body extraction based on deep learning, characterized in that, include: The preprocessed image is segmented into training images that meet the computing power requirements of the computer. According to the minimum error segmentation method, the clustering range of the deep learning framework is defined near the segmentation threshold. The deep learning framework is used to cluster the pixels in this range. The intermediate region corresponding to the segmentation threshold is the range between high-confidence water bodies and non-water bodies. High-confidence water bodies are pixel regions with extremely low reflectivity in SAR images, and high-confidence non-water bodies are pixel regions with extremely high reflectivity in SAR images. When training on training images with extremely uneven proportions of water and non-water bodies, two hyperparameters are used to prevent the loss of water or non-water body pixel blocks as the number of training iterations increases. The two hyperparameters are set based on the pixel range [0, 255] of Uint8 data, and the specific constraints are defined by the formula: ； in, and The first and The pixel values ​​of the superpixel blocks output by the second training iteration, with hyperparameters of 90 and 160 respectively; The clustered images are binarized using the pixel features of the original training images, ultimately generating a water-non-water body classification map.

2. A large-scale, rapid remote sensing water body extraction system based on deep learning, characterized in that, include: The data preprocessing module performs track correction and thermal noise removal after the raw data is input. After radiometric calibration, the processed data is then filtered and topographically corrected. The corrected data is converted to dB units for image cropping; The convolutional network clustering module generates segmentation labels after superpixel segmentation of the original image. The original image then undergoes three convolutional layers to output feature maps. These feature maps are further processed through original image fusion, segmentation label clustering, and two hyperparameter constraint steps to generate clustering feature maps. Finally, a loss function is generated using the clustering feature maps and the feature maps, and the parameters of each node in the network are continuously optimized through backpropagation. The two hyperparameter constraints refer to limiting the disappearance of excessively low or high pixel value regions during the clustering process. Specific constraints are expressed by the following formula: ， in, and The first and The pixel values ​​of the superpixel blocks output by the second training iteration, with hyperparameters of 90 and 160 respectively; The clustering binarization module combines the pixel intensity information in the original image with the clustering results output by the network to obtain a clustering result image that best reflects the features of the original image.

3. The large-scale rapid remote sensing water body extraction system based on deep learning as described in claim 2, characterized in that, After preprocessing, the intensity data is converted into dB data using the following formula: ； in, I The pixel intensity is the preprocessed value, and dB is the decibel unit data generated by the conversion.

4. The large-scale rapid remote sensing water body extraction system based on deep learning as described in claim 2, characterized in that, The image unit after preprocessing by the preprocessing module is dB, and the data type is 32-bit floating point. Before the convolutional neural network clusters the images, the data type is converted to an unsigned 8-bit integer. Before the data type conversion, all training images are linearly stretched by 1%, and the stretching range of all images is the same.

5. The large-scale rapid remote sensing water body extraction system based on deep learning as described in claim 2, characterized in that, In the early stages of training, low-dimensional features need to be preserved by "fusion of input images." As the number of training iterations increases, small areas of water in the images will gradually disappear as low-dimensional features, thus reducing the final clustering accuracy. The original image and the feature map output from the convolutional layer are fused to obtain a new feature map containing both feature information and information from the original image. The fusion method involves weighting the original image and the feature map according to a certain ratio. As the number of training iterations increases, the weight of the original image gradually decreases, while the weight of the feature map gradually increases. The image fusion formula is: ， in To obtain the pixel values ​​of the "new feature map" after fusing the original image, Indicates the number of training sessions. This indicates the pixel value corresponding to the input image. This represents the pixel value corresponding to the output feature map of the convolutional layer.

6. The large-scale rapid remote sensing water body extraction system based on deep learning as described in claim 2, characterized in that, The new feature map generated after fusion still needs to be processed by segmentation labels.

7. A computer device, characterized in that, The computer device includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the steps of the deep learning-based large-scale rapid remote sensing water extraction system as described in claim 2.

8. A computer-readable storage medium, characterized in that, The storage medium is used to store a computer program, which, when executed by a processor, causes the processor to perform the steps of the deep learning-based large-scale rapid remote sensing water extraction system as described in claim 2.

9. An information data processing terminal, characterized in that, The information data processing terminal is used to implement the large-scale rapid remote sensing water body extraction system based on deep learning as described in claim 2.

Images