Water body detection method and system for synthetic aperture radar image based on lognormal distribution and multi-scale feature fusion

Abstract

The application belongs to the technical field of image processing, and relates to a water body detection method and system for synthetic aperture radar images based on a logarithmic normal distribution and multi-scale feature fusion. The method comprises: performing a logarithm operation on an input synthetic aperture radar image; performing local window parameter estimation on the image after the logarithm operation by using windows of different scales to obtain the mean and standard deviation of pixels in the windows of different scales; calculating a Gaussian distance according to the mean and standard deviation of the pixels in the windows of different scales to obtain a distance map corresponding to the window of each scale; performing a first filtering according to the distance map corresponding to the window of each scale to eliminate high-intensity outliers and ground object pixels; fusing the distance maps corresponding to the windows of different scales; performing a second filtering according to the fused distance map to eliminate shadow pixels to obtain water body pixels. The application can expand the difference between feature approximate SAR water bodies and shadows, so as to accurately distinguish the two.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology and relates to a method for water body detection and segmentation of synthetic aperture radar images. Specifically, it relates to a method and system for water body detection of synthetic aperture radar images based on log-normal distribution and multi-scale feature fusion. Background Technology

[0002] In SAR (Synthetic Aperture Radar) images, water bodies appear dark due to the weak echo signals received by the sensor, caused by specular reflection of the radar signal on the smooth water surface. Water bodies exhibit a distinct bimodal distribution in the intensity histogram of SAR images; therefore, threshold-based methods were widely used for water body extraction in the early development of this field. However, the accuracy of threshold-based SAR water body segmentation is generally lower than that of optical remote sensing images. Currently, there are many methods for extracting water bodies from SAR images, which can be summarized into three main categories: (1) threshold-based methods, (2) image feature-based methods, and (3) deep learning-based methods, as well as combinations and variations of related methods. The classification of SAR water body detection techniques is shown in Table 1.

[0003] Table 1. Classification of SAR water body detection technologies In recent years, deep learning algorithms have been widely used and have achieved remarkable results. However, such methods require a large amount of data for training to support the construction of reasonable parameters for the model, and the number of parameters is huge. Its application prospects are broad, but it is heavily dependent on hardware. The method of water body detection in SAR images based on statistical characteristics, which belongs to the image feature method, is mainly constrained by two types of factors that affect the detection accuracy: (1) high intensity outliers in water body pixels lead to overestimation of statistical parameters; (2) the similarity between shadows and water bodies in statistical features makes it difficult to distinguish between the two.

[0004] The grayscale / threshold method is susceptible to interference from terrain and wind / wave (roughness), easily leading to false detections. Furthermore, low-scattering features, such as smooth asphalt roads, airport runways, exposed saline-alkali land, and frozen calm water surfaces, have grayscale values ​​highly similar to water bodies, making them difficult to distinguish. Image feature methods involve complex feature processes and rely heavily on expert experience; inappropriate feature combinations not only fail to improve performance but may even introduce noise and reduce accuracy. Machine learning methods heavily depend on high-quality labeled data, while low-quality or insufficient data easily leads to overfitting or poor generalization ability.

[0005] This invention can be classified as an image feature method. Image feature methods are easily affected by the environment and are sensitive to noise, which leads to inaccurate feature analysis. Moreover, it is difficult to distinguish similar features of different land features, thus affecting the accuracy of SAR water body detection. In practical applications, this problem is specifically manifested as follows: (1) SAR image water body detection is affected by high-intensity outlier interference, which causes statistical parameters to be overestimated, resulting in an increased false detection rate; (2) Water bodies are similar to shadows and are not easy to distinguish, resulting in reduced detection accuracy. Summary of the Invention

[0006] To address the aforementioned issues, this patent proposes a method and system for water body detection based on log-normal distribution and multi-scale feature fusion of synthetic aperture radar images.

[0007] The technical solution adopted in this invention is as follows: A water body detection method based on log-normal distribution and multi-scale feature fusion of synthetic aperture radar images includes the following steps: Perform a logarithmic operation on the input synthetic aperture radar image; For the image after logarithmic operation, local window parameters are estimated using windows of different scales to obtain the mean and standard deviation of pixels within windows of different scales. Gaussian distance is calculated based on the mean and standard deviation of pixels within windows of different scales, resulting in a distance map for each scale window. The first filtering is performed based on the distance map of the window corresponding to each scale to eliminate high-intensity outliers and ground feature pixels; The distance maps corresponding to windows of different scales are fused together, and a second filtering is performed based on the fused distance map to eliminate shadow pixels and obtain water body pixels.

[0008] Furthermore, the estimation of local window parameters using windows of different scales includes: Acquire data within a local window of the image and estimate the cutoff factor based on the acquired data; Calculate the cutoff threshold based on the cutoff factor; The data is truncated according to the truncation threshold to obtain the truncated data; Estimate the mean and standard deviation of the Gaussian distribution based on the truncated data.

[0009] Furthermore, the truncation factor, truncation threshold, and truncated data are calculated using the following formula: in, λ As the cutoff factor, μ The mean of the data. σThe standard deviation of the data. T This is the truncation threshold; X The input data after taking the logarithm. This is the truncated data.

[0010] Furthermore, the truncated data is calculated using the following formula. true mean and true standard deviation : in, express A pixel value within, , , , .

[0011] Furthermore, the Gaussian distance is calculated using the following formula: in, This represents the Gaussian distance.

[0012] Furthermore, the first filtering uses the maximum inter-class variance method to perform binary classification on the data in the distance map to obtain the classification threshold. Pixels in the distance map that are greater than the threshold are assigned a value of 0, while the values ​​of pixels at other positions remain unchanged.

[0013] Furthermore, the process of fusing distance maps corresponding to windows of different scales and performing a second filtering based on the fused distance map includes: multiplying the distance maps corresponding to windows of different scales to amplify the difference between water and shadows to obtain a fused distance map; classifying the fused distance map using the maximum inter-class variance method to obtain an inter-class threshold; assigning a value of 0 to pixels greater than the threshold, and the remaining pixels are water pixels.

[0014] A water body detection system based on synthetic aperture radar images using log-normal distribution and multi-scale feature fusion, comprising: The logarithm operation module is used to perform a logarithm operation on the input synthetic aperture radar image; The local window parameter estimation module is used to estimate the local window parameters of the image after logarithmic operation using windows of different scales, and obtain the mean and standard deviation of pixels within the window of different scales. The Gaussian distance calculation module is used to calculate the Gaussian distance based on the mean and standard deviation of pixels within windows of different scales, and obtain the distance map corresponding to each scale window. The first filtering module is used to perform the first filtering based on the distance map of the window corresponding to each scale, eliminating high-intensity outliers and ground feature pixels; The second filtering module is used to fuse the distance maps of windows at different scales, and then perform a second filtering based on the fused distance map to eliminate shadow pixels and obtain water pixels.

[0015] The beneficial effects of this invention are as follows: 1) This invention improves the parameter estimation method, significantly reducing processing time while maintaining statistical accuracy when estimating parameters for truncated clutter. Furthermore, traditional truncation thresholds often require manually setting the truncation factor; this invention proposes an adaptive method for calculating the truncation factor.

[0016] 2) By fusing different “distance maps” (multiplying corresponding pixels), this invention can amplify the differences between SAR water bodies and shadows with similar features, thereby accurately distinguishing the two. Attached Figure Description

[0017] Figure 1 This is a flowchart of the steps for water body detection in synthetic aperture radar images based on log-normal distribution and multi-scale feature fusion.

[0018] Figure 2 This is a schematic diagram of the water body detection method for synthetic aperture radar images based on log-normal distribution and multi-scale feature fusion.

[0019] Figure 3 It is a local window parameter estimation technique process.

[0020] Figure 4 This is a comparison of the effects of specifying a cutoff factor and adaptively calculating a cutoff factor, where: (a) SAR image; (b) specified cutoff factor 2.0; (c) adaptively calculated cutoff factor 3.0.

[0021] Figure 5 The comparison shows the effects of multi-scale feature fusion, including: (a) SAR water body map (river); (b) water body label map; (c) filtered distance map obtained from an 11×11 pixel window; (d) heat map display effect of (c); (e) filtered distance map obtained from a 21×21 pixel window; (f) heat map display effect of (e); (g) distance map fused from (c) and (e); (h) heat map display effect of (g); and (i) filtered water body segmentation map of (g).

[0022] Figure 6The results are compared using different methods, including: (a) SAR water body map (paddy field); (b) water body label map; (c) U-Net detection results; (d) DeepLabv3+ detection results; (e) OTSU detection results; (f) MISP-GGD detection results; and (g) detection results of this method.

[0023] Figure 7 This is a block diagram of the water body detection system based on log-normal distribution and multi-scale feature fusion of synthetic aperture radar images according to the present invention. Detailed Implementation

[0024] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0025] The core idea of ​​this invention is "removing ground features first, then segmenting water bodies," meaning that ground features and high-intensity outliers are first removed based on the statistical characteristics of SAR water bodies, and then water bodies and shadows are segmented while retaining the water bodies. The water body and shadow differentiation involved in this scheme is completely different from existing methods.

[0026] The key points of this invention are mainly reflected in the following two aspects: 1) Truncation Gaussian Parameter Estimation: An improved statistical parameter estimation method is proposed to significantly reduce processing time while maintaining statistical accuracy when estimating parameters for truncated clutter. For cases where the truncation threshold needs to be manually set, this solution proposes an adaptive method for calculating the truncation threshold.

[0027] 2) Multi-scale feature fusion: Gaussian distance calculation relies on the statistical parameters of data within a local window, and different window sizes cause variations in these parameters. Based on this characteristic, this paper proposes a multi-scale fusion scheme that can amplify the differences between SAR water bodies and shadows that are difficult to distinguish, thereby differentiating between the two.

[0028] In one embodiment, the present invention provides a water body detection method based on log-normal distribution and multi-scale feature fusion of synthetic aperture radar images, such as... Figure 1 As shown, it includes the following steps: Perform a logarithmic operation on the input synthetic aperture radar image; For the image after logarithmic operation, local window parameters are estimated using windows of different scales to obtain the mean and standard deviation of pixels within windows of different scales. Gaussian distance is calculated based on the mean and standard deviation of pixels within windows of different scales to obtain a distance map for each scale window. The first filtering is performed based on the distance map of the window corresponding to each scale to eliminate high-intensity outliers and ground feature pixels; The distance maps corresponding to windows of different scales are fused together, and a second filtering is performed based on the fused distance map to eliminate shadow pixels and obtain water body pixels.

[0029] In one embodiment, the estimation of local window parameters using windows of different scales includes: Acquire data within a local window of the image and estimate the cutoff factor based on the acquired data; Calculate the cutoff threshold based on the cutoff factor; The data is truncated according to the truncation threshold to obtain the truncated data; Estimate the mean and standard deviation of the Gaussian distribution based on the truncated data.

[0030] In one embodiment, the first filtering uses the maximum inter-class variance method to perform binary classification on the data in the distance map to obtain a classification threshold, assigns a value of 0 to pixels in the distance map that are greater than the threshold, and keeps the values ​​of pixels at other locations unchanged.

[0031] In one embodiment, fusing the distance maps corresponding to windows of different scales and performing a second filtering based on the fused distance map includes: multiplying the distance maps corresponding to windows of different scales to amplify the difference between water and shadows to obtain the fused distance map; classifying the fused distance map using the maximum inter-class variance method to obtain an inter-class threshold; assigning a value of 0 to pixels greater than the threshold, and the remaining pixels are water pixels.

[0032] In one embodiment, such as Figure 2 As shown, the water body detection method of synthetic aperture radar image based on log-normal distribution and multi-scale feature fusion of the present invention mainly includes four parts: logarithmic operation, local window parameter statistics, Gaussian distance calculation and filtering, and multi-scale fusion and filtering.

[0033] (1) Logarithmic operation. Perform a logarithmic operation on the input SAR image. SAR water bodies approximately follow a log-normal distribution. After taking the logarithm, they approximately follow a Gaussian distribution (or normal distribution), that is, the log-normal distribution is transformed into a Gaussian distribution.

[0034] (1) in, I For the input SAR image, I ln For the image after taking the logarithm, ln (·) represents the logarithmic operation.

[0035] (2) Local window parameter estimation. Parameter estimation is performed based on the logarithmic image, as follows: Figure 3 As shown.

[0036] a) Cutoff factor estimation. Obtain the data within a V×V window and estimate the cutoff factor based on this data. The mathematical expression is as follows: (2) in, λ As the cutoff factor, μ The mean of the data. σ denoted as the standard deviation of the data.

[0037] Where V×V represents the size of a local window (or patch) in the image, and the data refers to the pixel values ​​of the synthetic aperture radar image; for a local window, it refers to the pixel values ​​of the window image. V can be flexibly set according to the image size, for example, one-twentieth of the number of rows or columns, with a minimum of 3 pixels × 3 pixels. In this embodiment, 11 pixels × 11 pixels are used.

[0038] b) Calculation of the cutoff threshold. The cutoff threshold is calculated based on the cutoff factor: (3) in, T The truncation threshold is used. Data is truncated according to the following formula: (4) in, X The input is the logarithm of the data. This is the truncated data.

[0039] c) Gaussian distribution statistical parameter estimation. The mean and standard deviation of the Gaussian distribution are estimated based on the truncated data. Since the data is truncated, direct statistical analysis of the truncated data will result in some differences in statistical parameters compared to the data before truncation; therefore, the truncation factor is included in the statistics.

[0040] make The data is truncated, and is the set of pixels within the local window before truncation. X It follows a Gaussian distribution and can be denoted as: . X The cumulative distribution function is denoted as ,in x express X A specific pixel value within the range. For The probability density function can be expressed as: (5) (6) Equation (5) can be further derived as follows: (7) in, Θ Let be the cumulative distribution function, then Θ ( T , μ, σ The expression for ) is: (8) Based on formulas (5) to (8), the samples are truncated. true mean and true standard deviation It can be calculated using maximum likelihood estimation, and its mathematical expression is: (9) (10) in, express A pixel value within, , , , .

[0041] The reason why this invention performs truncated statistics is because X The image contains noise (land or shore pixels, ship pixels, speckle noise, etc.), and this noise does not conform to statistical laws, directly affecting... X Parameter estimation can lead to significant deviations in the mean and standard deviation of water bodies that conform to statistical laws. In this invention... It is based on the theoretical mean and standard deviation obtained from the statistics of water pixels that satisfy a Gaussian distribution after truncation, that is, using a finite sample after truncation to estimate the true mean of the population sample. and true standard deviation .

[0042] (3) Gaussian distance calculation and filtering.

[0043] a) Gaussian distance calculation. Based on the local window parameter estimation described above, each pixel in the image can correspond to a set of... The Gaussian distance is calculated as follows: (11) A "distance map" can be obtained using equation (11).

[0044] b) Filtering. The Otsu's method (Otsu's method) is used to perform binary classification on the data within the "distance map" to obtain the classification threshold. For high-intensity outliers and ground features, their distance... D The distance to the image is relatively large, while the distance to water and shadows is relatively small. Using this characteristic, pixels in the "distance map" that are greater than a threshold can be assigned a value of 0, while leaving the values ​​of pixels at other locations unchanged.

[0045] (4) Multi-scale fusion and filtering.

[0046] a) Multi-scale fusion. Through the above operations, high-intensity outliers and ground object pixels have been filtered out, leaving only water bodies and shadows. SAR water bodies and shadows are similar in grayscale and statistical characteristics, but differences still exist. The statistical characteristics of water bodies are more pronounced, meaning they appear farther away on the distance map. Different "distance maps" can be obtained by estimating parameters and calculating Gaussian distances for windows of different sizes. Multiplying these different distance maps effectively amplifies the differences between water bodies and shadows, resulting in a "distance map with greater differences."

[0047] For windows of different sizes, you can choose two windows of different sizes, such as 11×11 and 21×21. Alternatively, you can choose three, four or more windows. From a practical application perspective, it is preferable to use two windows.

[0048] b) Filtering. Filtering is the final segmentation of water bodies and shadows, and removes shadows. For distance maps with greater differences, the Otsu's method is used to classify the two to obtain a threshold between classes. Pixel values ​​greater than this threshold are assigned a value of 0, and the remaining pixels are water body pixels. Setting the water body pixel values ​​to 1 (binary image) creates a water body mask, which can effectively extract water bodies.

[0049] The beneficial effects of this invention are: Key point 1: Truncation Gaussian parameter estimation.

[0050] Beneficial technical effects: This scheme improves the parameter estimation method, significantly reducing processing time while maintaining statistical accuracy when estimating parameters for truncated clutter. Furthermore, traditional truncation thresholds often require manual setting of the truncation factor; this scheme proposes an adaptive method for calculating the truncation factor.

[0051] (1) Parameter estimation comparison. Two single-channel SAR images of different sizes were selected, and the parameter estimation was performed three times for each image. The time taken for the parameter estimation was recorded. For accuracy, the mean of the estimated parameters was recorded, including the mean of the mean and the mean of the standard deviation. The comparison results are shown in Table 2.

[0052] Table 2 Comparison of Processing Times Operating device specifications: CPU: Core (TM) i9-10900K, Memory: 32 GB RAM, Operating system: 64-bit Win10, Matlab version: R2014b.

[0053] (2) Adaptive Truncation Factor. A 256×256 image is used for illustration. The effects of specifying a truncation factor and adaptively calculating a truncation factor are compared. From... Figure 4As can be seen, the adaptive truncation factor can calculate a suitable truncation factor based on SAR image data. However, a specified truncation factor often fails to provide a complete fit, for example: Figure 4 In (b), only some pixels of the ship at sea are removed.

[0054] Key Point 2: Multi-scale feature fusion.

[0055] Beneficial technical effect: By fusing different "distance maps" (multiplying corresponding pixels), the differences between SAR water bodies and shadows with similar features can be amplified, thereby accurately distinguishing between the two. An example and effect demonstration are provided here using two "distance maps" (distance maps before and after fusion), such as... Figure 5 As shown. From Figure 5 From (c) to (h), it can be seen that as fusion proceeds, the difference between the water body and the shadow increases. Figure 5 (as shown in h). Water bodies and shadows can be distinguished using a simple Otsu's method, such as... Figure 5 As shown in (i). Figure 6 List the results of this scheme and compare them with those of other existing algorithms.

[0056] Other embodiments of the present invention: 1) This method can be further extended in some technical details. For example, after obtaining the binary image of the water body, morphological filtering can be performed to eliminate the existing discrete points or noise points.

[0057] 2) The method described here uses two different window sizes for parameter estimation. In practical applications, multiple windows of different sizes can be selected for statistics, and then the different distance maps can be multiplied together.

[0058] Another embodiment of the present invention provides a water body detection system based on log-normal distribution and multi-scale feature fusion of synthetic aperture radar images, such as... Figure 7 As shown, it includes: The logarithm operation module is used to perform a logarithm operation on the input synthetic aperture radar image; The local window parameter estimation module is used to estimate the local window parameters of the image after logarithmic operation using windows of different scales, and obtain the mean and standard deviation of pixels within the window of different scales. The Gaussian distance calculation module is used to calculate the Gaussian distance based on the mean and standard deviation of pixels within windows of different scales, and obtain the distance map for each scale window. The first filtering module is used to perform the first filtering based on the distance map of the window corresponding to each scale, eliminating high-intensity outliers and ground feature pixels; The second filtering module is used to fuse the distance maps of windows at different scales, and then perform a second filtering based on the fused distance map to eliminate shadow pixels and obtain water body pixels.

[0059] The above division of modules is merely illustrative. In practical applications, the functions described above can be assigned to different functional modules as needed to complete all or part of the functions described in the aforementioned method. The specific working process of each module can be found in the corresponding process in the aforementioned method embodiments. Each module can be implemented entirely or partially through software, hardware, or a combination thereof.

[0060] Another embodiment of the present invention provides a computer device (computer, server, etc.) including a memory and a processor, the memory storing a computer program configured to be executed by the processor, the computer program including instructions for performing steps of the method of the present invention.

[0061] Another embodiment of the present invention provides a computer-readable storage medium (such as ROM / RAM, disk, optical disk) that stores a computer program, which, when executed by a computer, implements the steps of the method of the present invention.

[0062] Another embodiment of the present invention provides a computer program product, the computer program product including a computer program, which, when executed by a computer, implements the steps of the method of the present invention.

[0063] The specific embodiments of the present invention disclosed above are intended to help understand the content of the present invention and to implement it accordingly. Those skilled in the art will understand that various substitutions, changes, and modifications are possible without departing from the spirit and scope of the present invention. The present invention should not be limited to the content disclosed in the embodiments of this specification; the scope of protection of the present invention is defined by the claims.

Claims

1. A method for water body detection based on synthetic aperture radar images using log-normal distribution and multi-scale feature fusion, characterized in that, Includes the following steps: Perform a logarithmic operation on the input synthetic aperture radar image; For the image after logarithmic operation, local window parameters are estimated using windows of different scales to obtain the mean and standard deviation of pixels within windows of different scales. Gaussian distance is calculated based on the mean and standard deviation of pixels within windows of different scales, resulting in a distance map for each scale window. The first filtering is performed based on the distance map of the window corresponding to each scale to eliminate high-intensity outliers and ground feature pixels; The distance maps corresponding to windows of different scales are fused together, and a second filtering is performed based on the fused distance map to eliminate shadow pixels and obtain water body pixels.

2. The method according to claim 1, characterized in that, The method of estimating local window parameters using windows of different scales includes: Acquire data within a local window of the image and estimate the cutoff factor based on the acquired data; Calculate the cutoff threshold based on the cutoff factor; The data is truncated according to the truncation threshold to obtain the truncated data; Estimate the mean and standard deviation of the Gaussian distribution based on the truncated data.

3. The method according to claim 2, characterized in that, The truncation factor, truncation threshold, and truncated data are calculated using the following formula: in, λ As the cutoff factor, μ The mean of the data. σ The standard deviation of the data. T This is the truncation threshold; X The input data after taking the logarithm. This is the truncated data.

4. The method according to claim 3, characterized in that, The truncated data is calculated using the following formula. true mean and true standard deviation : in, express A pixel value within, , , , .

5. The method according to claim 4, characterized in that, The Gaussian distance is calculated using the following formula: in, This represents the Gaussian distance.

6. The method according to claim 1, characterized in that, The first filtering uses the maximum inter-class variance method to perform binary classification on the data in the distance map to obtain the classification threshold. Pixels in the distance map that are greater than the threshold are assigned a value of 0, while the values ​​of pixels at other positions remain unchanged.

7. The method according to claim 1, characterized in that, The process of fusing distance maps corresponding to windows of different scales and performing a second filtering based on the fused distance map includes: multiplying the distance maps corresponding to windows of different scales to amplify the difference between water and shadows to obtain a fused distance map; classifying the fused distance map using the maximum inter-class variance method to obtain an inter-class threshold; assigning a value of 0 to pixels greater than the threshold, and the remaining pixels are water pixels.

8. A water body detection system based on synthetic aperture radar images using log-normal distribution and multi-scale feature fusion, characterized in that, include: The logarithm operation module is used to perform a logarithm operation on the input synthetic aperture radar image; The local window parameter estimation module is used to estimate the local window parameters of the image after logarithmic operation using windows of different scales, and obtain the mean and standard deviation of pixels within the window of different scales. The Gaussian distance calculation module is used to calculate the Gaussian distance based on the mean and standard deviation of pixels within windows of different scales, and obtain the distance map corresponding to each scale window. The first filtering module is used to perform the first filtering based on the distance map of the window corresponding to each scale, eliminating high-intensity outliers and ground feature pixels; The second filtering module is used to fuse the distance maps of windows at different scales, and then perform a second filtering based on the fused distance map to eliminate shadow pixels and obtain water pixels.

9. A computer device, characterized in that, It includes a memory and a processor, the memory storing a computer program configured to be executed by the processor, the computer program including instructions for performing the method of any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a computer, implements the method according to any one of claims 1 to 7.

Images