A SAR Remote Sensing Image Monitoring Method and System for Ulva prolifera Based on Maximum Likelihood Classification

Abstract

This invention relates to the field of remote sensing image recognition technology for Ulva prolifera, specifically a SAR Ulva prolifera remote sensing image monitoring method and system based on maximum likelihood classification. The method includes acquiring the original SAR image of the monitoring area, denoted as the first image; obtaining a second image through radiometric correction, noise reduction, and geometric correction; selecting sample regions from the second image; calculating the image features of each sample region to establish a training sample library; training a maximum likelihood classifier; using the trained maximum likelihood classifier for classification to obtain a preliminary Ulva prolifera classification result image, denoted as the third image; and performing image post-processing on the third image to obtain an optimized Ulva prolifera distribution map. This invention utilizes maximum likelihood classification to classify SAR images, improving the accuracy of Ulva prolifera region identification.

Description

Technical Field

[0001] This invention relates to the field of Ulva prolifera remote sensing image recognition technology, specifically to a SAR Ulva prolifera remote sensing image monitoring method and system based on maximum likelihood classification. Background Technology

[0002] Spaceborne synthetic aperture radar (SAR), as an active microwave remote sensing sensor, possesses all-weather, all-day imaging capabilities and is unaffected by weather conditions such as clouds and rain. Therefore, it plays an irreplaceable role in monitoring marine algae blooms that require high timeliness and continuity. Algae typically exhibits a higher backscattering coefficient in SAR images than in the surrounding seawater. Based on this characteristic, SAR images can be used to detect the distribution range of algae.

[0003] Currently, methods for extracting *Ulva prolifera* information from SAR imagery mainly rely on grayscale threshold-based segmentation algorithms. This method analyzes the differences in grayscale or backscattering intensity between *Ulva prolifera* and seawater in the image, sets a fixed threshold, and identifies pixel regions exceeding the threshold as *Ulva prolifera* regions. However, this method relies solely on the pixel's own grayscale or backscattering intensity, making it sensitive to the inherent speckle noise of SAR images. It easily misclassifies interfering targets with similar high backscattering characteristics (such as ships or wave-broken areas in seawater) as *Ulva prolifera*, and also easily misses weakly scattering *Ulva prolifera* information, resulting in high misclassification and omission rates. This makes traditional methods insufficiently robust under complex sea conditions and difficult to meet monitoring accuracy requirements.

[0004] Therefore, it is necessary to propose a classification method that can comprehensively utilize multiple feature information in SAR images and has stronger anti-interference capabilities in order to achieve accurate identification of Ulva prolifera areas. Summary of the Invention

[0005] (1) Technical problems to be solved

[0006] The purpose of this invention is to provide a method and system for monitoring SAR seaweed remote sensing images based on maximum likelihood classification, so as to identify seaweed areas in SAR images.

[0007] (2) Technical solution

[0008] To achieve the above objectives, this invention provides a SAR remote sensing image monitoring method for *Ulva prolifera* based on maximum likelihood classification, the method comprising the following steps:

[0009] S1. Obtain the original SAR image of the monitoring area, denoted as the first image. The original SAR image of the monitoring area is a fully polarimetric SAR image. Perform radiometric correction, noise reduction and geometric correction on the first image to obtain a preprocessed SAR image, denoted as the second image.

[0010] S2, from the second image, select the pre-identified seaweed sample area, seawater sample area, ship sample area, and marine structure sample area; calculate the image features of each sample area, including backscattering intensity features, polarization decomposition features, and texture features; and establish a training sample library based on the image features of all sample areas.

[0011] S3. Based on the training sample library, a maximum likelihood classifier is trained, and the trained maximum likelihood classifier is used to classify each pixel in the second image to obtain a preliminary classification result image of Ulva prolifera, which is denoted as the third image; the pixel regions that are classified as Ulva prolifera in the third image are denoted as Ulva prolifera pixel regions.

[0012] S4, perform image post-processing on the third image to obtain an optimized distribution map of Ulva prolifera; the image post-processing includes removing noise points and merging adjacent Ulva prolifera pixel regions.

[0013] Further, the method of performing radiometric correction, noise reduction, and geometric correction on the first image to obtain a preprocessed SAR image, denoted as the second image, includes:

[0014] The first image is radiometrically calibrated. The radiometric calibration steps are as follows: calculate the gray value of each pixel in the first image, convert the gray value of each pixel in the first image into a backscattering coefficient value representing the scattering ability of ground objects according to the calibration parameters attached to the satellite sensor, and reconstruct the radiometrically corrected image using the backscattering coefficient value.

[0015] The radiometrically corrected image is subjected to noise suppression; the noise suppression step is to use an adaptive speckle noise filtering algorithm based on local statistical characteristics to filter out speckle noise from the radiometrically corrected image, thereby obtaining a noise-reduced image.

[0016] The denoised image is then geometrically corrected. The geometric correction steps involve using satellite orbital parameters and digital elevation model data to perform terrain correction and geocoding on the denoised image, thereby correcting the denoised image to a pre-specified geographic coordinate system to obtain a second image.

[0017] Furthermore, the method for calculating image features for each sample region, including backscattering intensity features, polarization decomposition features, and texture features, and establishing a training sample library based on the image features of all sample regions includes:

[0018] The sample areas are numbered and designated as the first sample area to the Nth sample area, where N is the number of sample areas. A mapping relationship is established between the sample areas and the sample area types, including seaweed sample areas, seawater sample areas, ship sample areas, and marine structure sample areas.

[0019] Extract the backscattering coefficient values ​​of all pixels from the first sample region to the Nth sample region from the second image; calculate the arithmetic mean of the backscattering coefficient values ​​of all pixels from the first sample region to the Nth sample region, and denot it as the first backscattering intensity feature to the Nth backscattering intensity feature.

[0020] From the second image, the coherence matrix of each pixel belonging to the first sample region to the Nth sample region is extracted; the coherence matrix of each pixel is decomposed using a polarimetric target decomposition algorithm to obtain the entropy value, anisotropy value, and average scattering angle value corresponding to each pixel; the arithmetic mean of the entropy values ​​of all pixels in the first sample region to the Nth sample region is calculated and denoted as the first entropy value feature to the Nth entropy value feature; the arithmetic mean of the anisotropy values ​​of all pixels in the first sample region to the Nth sample region is calculated and denoted as the first anisotropy value feature to the Nth anisotropy value feature; the arithmetic mean of the average scattering angle values ​​of all pixels in the first sample region to the Nth sample region is calculated and denoted as the first average scattering angle value feature to the Nth average scattering angle value feature; the first entropy value feature to the Nth entropy value feature, the first anisotropy value feature to the Nth anisotropy value feature, and the first average scattering angle value feature to the Nth average scattering angle feature are combined to obtain the first polarization decomposition feature to the Nth polarization decomposition feature.

[0021] Calculate the texture features of the first sample region to the Nth sample region respectively, and denote them as the first texture feature to the Nth texture feature.

[0022] Based on the mapping relationship between sample regions and sample region types, establish mapping relationships between the first backscattering intensity feature to the Nth backscattering intensity feature, the first polarization decomposition feature to the Nth polarization decomposition feature, the first texture feature to the Nth texture feature, and the sample region type. Combine the first backscattering intensity feature to the Nth backscattering intensity feature, the first polarization decomposition feature to the Nth polarization decomposition feature, the first texture feature to the Nth texture feature, and the corresponding sample region type into N sample records. Each sample record corresponds to a sample region, and each sample record contains the sample region type and the backscattering intensity feature, polarization decomposition feature, and texture feature corresponding to the sample region.

[0023] Further, the method for calculating the texture features of the first sample region to the Nth sample region respectively, denoted as the first texture feature to the Nth texture feature, includes:

[0024] A square window with a pre-defined size of M rows by M columns is used as the neighborhood window, where M is a pre-defined odd number. Texture feature calculation operations are performed on the first sample region to the Nth sample region respectively. The steps of the texture feature calculation operation are as follows: According to the neighborhood window, for each pixel in the kth sample region, an index calculation step is performed to obtain the contrast index value, homogeneity index value, and information entropy index value of each pixel in the kth sample region; the arithmetic mean of the contrast index value, homogeneity index value, and information entropy index value of all pixels in the kth sample region is calculated to obtain the kth contrast index value, kth homogeneity index value, and kth information entropy index value; k is iterated through to obtain the first contrast index value to the Nth contrast index value, the first homogeneity index value to the Nth homogeneity index value, and the first information entropy index value to the Nth information entropy index value; the first contrast index value to the Nth contrast index value, the first homogeneity index value to the Nth homogeneity index value, and the first information entropy index value to the Nth information entropy index value are combined to obtain the first texture feature to the Nth texture feature.

[0025] Furthermore, the method for calculating the performance index includes:

[0026] Centered on the pixel to be calculated, the backscattering coefficient values ​​of all pixels within the coverage area of ​​the neighborhood window of the pixel to be calculated are extracted to obtain the gray-level co-occurrence matrix; based on the gray-level co-occurrence matrix, the contrast index value, the homogeneity index value, and the information entropy index value of the pixel to be calculated are obtained.

[0027] Furthermore, the method of extracting the backscattering coefficient values ​​of all pixels within the coverage area of ​​the neighborhood window of the pixel to be calculated, centered on the pixel to be calculated, and calculating the gray-level co-occurrence matrix includes:

[0028] Centered on the pixel to be calculated, extract the backscattering coefficient values ​​of all pixels within the coverage area of ​​the neighborhood window of the pixel to be calculated; search for the minimum and maximum values ​​of the backscattering coefficient values ​​of all pixels within the coverage area of ​​the neighborhood window of the pixel to be calculated, and denote them as follows: , ;according to , The intensity levels of all pixels within the neighborhood window coverage area of ​​the pixel to be calculated are obtained by using the intensity calculation formula, along with the pre-set total number of intensity quantization levels L. The intensity levels of all pixels within the neighborhood window coverage area of ​​the pixel to be calculated are statistically analyzed to obtain the gray-level co-occurrence matrix.

[0029] The strength calculation formula is as follows:

[0030] ;

[0031] in, Indicates the intensity level of a pixel; This represents the backscattering coefficient value of a pixel; Indicates to Round down to the nearest integer.

[0032] Furthermore, the method for statistically analyzing the intensity levels of all pixels within the neighborhood window coverage area of ​​the pixel to be calculated to obtain the gray-level co-occurrence matrix includes:

[0033] Within the neighborhood window, traverse i (taking integers from 1 to L) and j (taking integers from 1 to L) to search for the number of pixel pairs that satisfy the first matching statistical condition, denoted as . The number of pixel pairs that satisfy the second matching statistical condition is denoted as . ,Will Divide by get ; Let be the element in the i-th row and j-th column of the gray-level co-occurrence matrix.

[0034] The first matching statistical condition is that within the neighborhood window, in a pixel pair consisting of two pixels, one pixel has an intensity level of i and the other pixel has an intensity level of j, and the spatial distance between the two pixels is equal to a preset standard pixel distance d, and the angle between the line connecting the two pixels and a preset horizontal line is equal to a preset standard angle. .

[0035] The second matching statistical condition is that within the neighborhood window, in a pixel pair consisting of two pixels, the spatial distance between the two pixels is equal to a preset standard pixel distance d, and the angle between the line connecting the two pixels and a preset horizontal line is equal to a preset standard angle. .

[0036] Further, the method for calculating the contrast index value, homogeneity index value, and information entropy index value of the pixel to be calculated based on the gray-level co-occurrence matrix includes:

[0037] The contrast ratio, homogeneity, and information entropy of the pixel to be calculated are obtained based on the gray-level co-occurrence matrix.

[0038] The formula for calculating the contrast index value of the pixel to be calculated is as follows:

[0039] ;

[0040] in, This represents the contrast ratio value of the pixel to be calculated.

[0041] The formula for calculating the homogeneity index value of the pixels to be calculated is as follows:

[0042] ;

[0043] in, This represents the homogeneity index value of the pixels to be calculated.

[0044] The formula for calculating the information entropy index value of the pixel to be calculated is as follows:

[0045] ;

[0046] in, This represents the information entropy index value of the pixel to be calculated.

[0047] Further, the method of training a maximum likelihood classifier based on the training sample library, and using the trained maximum likelihood classifier to classify each pixel in the second image to obtain a preliminary classification result image of *Ulva prolifera*, denoted as the third image, includes:

[0048] Based on the sample region type, backscattering intensity features, polarization decomposition features, and texture features of all sample records in the training sample library, the maximum likelihood estimation algorithm is used to calculate the feature mean vector and covariance matrix of each category in the categories of seaweed, seawater, ships, and marine structures. The feature mean vector and covariance matrix of each category are set as the built-in parameters of the pre-set maximum likelihood classifier to obtain the trained maximum likelihood classifier.

[0049] Calculate the backscattering intensity feature, polarization decomposition feature, and texture feature of each pixel in the second image; input the backscattering intensity feature, polarization decomposition feature, and texture feature of each pixel in the second image into the trained maximum likelihood classifier, and output the likelihood probabilities of the seaweed category, seawater category, ship category, and marine structure category, respectively denoted as the likelihood probability of the seaweed category, the likelihood probability of the seawater category, the likelihood probability of the ship category, and the likelihood probability of the marine structure category; the category corresponding to the maximum value among the likelihood probabilities of the seaweed category, the likelihood probability of the seawater category, the likelihood probability of the ship category, and the likelihood probability of the marine structure category is taken as the category to which the pixel belongs.

[0050] Traverse all pixels in the second image and calculate the category of each pixel; record and spatially arrange the categories of all pixels to generate the third image.

[0051] Based on the same inventive concept, this invention also provides a SAR Ulva prolifera remote sensing image monitoring system based on maximum likelihood classification, the system comprising, in sequence: a data reading module, a training sample library construction module, a maximum likelihood classification module, and an image post-processing module.

[0052] The data reading module is used to acquire the original SAR image of the monitoring area, denoted as the first image. The original SAR image of the monitoring area is a fully polarimetric SAR image. The first image is subjected to radiometric correction, noise reduction and geometric correction to obtain a preprocessed SAR image, denoted as the second image.

[0053] The training sample library construction module is used to select pre-identified seaweed sample areas, seawater sample areas, ship sample areas, and marine structure sample areas from the second image; calculate the image features of each sample area, including backscattering intensity features, polarization decomposition features, and texture features; and establish a training sample library based on the image features of all sample areas.

[0054] The maximum likelihood classification module is used to train a maximum likelihood classifier based on the training sample library, and use the trained maximum likelihood classifier to classify each pixel in the second image to obtain a preliminary classification result image of Ulva prolifera, which is denoted as the third image; the pixel regions that are classified as Ulva prolifera in the third image are identified and denoted as Ulva prolifera pixel regions.

[0055] The image post-processing module is used to perform image post-processing on the third image to obtain an optimized distribution map of Ulva prolifera; the image post-processing includes removing noise points and merging adjacent Ulva prolifera pixel regions.

[0056] (3) Beneficial effects

[0057] Compared with the prior art, the beneficial effects of the present invention are:

[0058] By comprehensively utilizing backscattering intensity features, polarization decomposition features, and texture features in fully polarimetric SAR images, and training a maximum likelihood classifier based on the fused multidimensional features for pixel-level classification, the ability of *Ulva prolifera* to be better separated from interfering targets such as seawater, ships, and marine structures in the high-dimensional feature space is improved, thereby enhancing the accuracy of *Ulva prolifera* identification. Attached Figure Description

[0059] Figure 1 This is a flowchart of the SAR Ulva prolifera remote sensing image monitoring method based on maximum likelihood classification according to Embodiment 1 of the present invention.

[0060] Figure 2 This is a schematic diagram of the module composition of the SAR Ulva prolifera remote sensing image monitoring system based on maximum likelihood classification in Embodiment 2 of the present invention. Detailed Implementation

[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] Before providing examples, it is necessary to describe the application scenario of the present invention, which is applied to the identification of seaweed areas in SAR images.

[0063] Example 1: As Figure 1 As shown, this embodiment provides a SAR remote sensing image monitoring method for *Ulva prolifera* based on maximum likelihood classification. The method includes the following steps:

[0064] S1. Obtain the original SAR image of the monitoring area, denoted as the first image. The original SAR image of the monitoring area is a fully polarimetric SAR image. Perform radiometric correction, noise reduction and geometric correction on the first image to obtain a preprocessed SAR image, denoted as the second image.

[0065] S2, from the second image, select the pre-identified seaweed sample area, seawater sample area, ship sample area, and marine structure sample area; calculate the image features of each sample area, including backscattering intensity features, polarization decomposition features, and texture features; and establish a training sample library based on the image features of all sample areas.

[0066] S3. Based on the training sample library, a maximum likelihood classifier is trained, and the trained maximum likelihood classifier is used to classify each pixel in the second image to obtain a preliminary classification result image of Ulva prolifera, which is denoted as the third image; the pixel regions that are classified as Ulva prolifera in the third image are denoted as Ulva prolifera pixel regions.

[0067] S4, perform image post-processing on the third image to obtain an optimized distribution map of Ulva prolifera; the image post-processing includes removing noise points and merging adjacent Ulva prolifera pixel regions.

[0068] For example, the area with the most severe seaweed infestation in a certain sea area was selected as the research object. Although seaweed is non-toxic, its massive proliferation can block sunlight, consume oxygen in seawater, and affect the normal growth of other marine organisms, thus leading to a series of ecological and environmental problems. In this embodiment, a SAR satellite was used to acquire the original SAR image of the area on a certain day, which is denoted as the first image. The SAR satellite has 12 imaging modes, including spotting, striping, full polarization, and wave modes, providing rich scattering information for ground object identification. According to the calibration parameters attached to the satellite sensor, a radiometric calibration method was used to convert the gray value of each pixel in the acquired first image into a backscattering coefficient value and reconstruct a radiometrically corrected image. The radiometrically corrected image was processed using the Lee-Sigma adaptive filtering algorithm based on local statistical characteristics to filter out speckle noise and obtain a denoised image. Using the satellite's orbital parameters and digital elevation model data, terrain correction and geocoding were performed on the denoised image to correct it to the WGS84 geographic coordinate system, resulting in the second image.

[0069] From the second image, pre-identified sample regions for seaweed, seawater, ships, and marine structures are selected. Image features for each sample region are calculated, including backscattering intensity, polarization decomposition, and texture features. Backscattering intensity is obtained by calculating the arithmetic mean of the backscattering coefficients of all pixels within the sample region. Polarization decomposition is obtained by extracting the coherence matrix of each pixel within the sample region, decomposing it using the Cloude-Pottier decomposition algorithm, and then combining the arithmetic mean of the resulting entropy, anisotropy, and average scattering angle values. Texture features are calculated by analyzing the spatial distribution of pixel intensity within the sample region. All the above features from all sample regions are collected to establish a training sample library containing sample region types and their corresponding feature vectors.

[0070] Based on the established training sample library, a classifier is trained using the maximum likelihood estimation algorithm. Using the trained classifier, each pixel in the second image is classified: backscattering intensity features, polarization decomposition features, and texture features are extracted from the pixel, combined into a feature vector, and input into the classifier. The likelihood probability of the pixel belonging to each preset category is calculated, and the pixel is assigned to the category with the highest likelihood probability. After traversing all pixels in the second image, a preliminary classification result image is generated, identifying the pixel regions of the *Ulva prolifera* category, denoted as the third image.

[0071] Finally, post-processing was performed on the third image. Morphological filtering was used to remove isolated noise points from the image. Based on spatial adjacency, adjacent *Ulva prolifera* pixel regions were merged to form a complete and continuous *Ulva prolifera* region. After the above processing, an optimized *Ulva prolifera* distribution map was obtained. Experiments showed that the overall accuracy of this method in identifying *Ulva prolifera* in this region was 92.3%, while the overall accuracy of the traditional thresholding method was only 76.5%.

[0072] Further, the method of performing radiometric correction, noise reduction, and geometric correction on the first image to obtain a preprocessed SAR image, denoted as the second image, includes:

[0073] The first image is radiometrically calibrated. The radiometric calibration steps are as follows: calculate the gray value of each pixel in the first image, convert the gray value of each pixel in the first image into a backscattering coefficient value representing the scattering ability of ground objects according to the calibration parameters attached to the satellite sensor, and reconstruct the radiometrically corrected image using the backscattering coefficient value.

[0074] The radiometrically corrected image is subjected to noise suppression; the noise suppression step is to use an adaptive speckle noise filtering algorithm based on local statistical characteristics to filter out speckle noise from the radiometrically corrected image, thereby obtaining a noise-reduced image.

[0075] The denoised image is then geometrically corrected. The geometric correction steps involve using satellite orbital parameters and digital elevation model data to perform terrain correction and geocoding on the denoised image, thereby correcting the denoised image to a pre-specified geographic coordinate system to obtain a second image.

[0076] For example, the radiometric calibration process involves first calculating the grayscale value of each pixel in the first image. Based on the calibration parameters provided with the satellite sensor, the grayscale value of each pixel is converted into a backscattering coefficient value representing the scattering capability of ground objects. After the conversion of all pixels is completed, a new image is reconstructed using these backscattering coefficient values; this image is the radiometrically corrected image. This step converts the original sensor-recorded values ​​into scattering intensity values ​​with clear physical meaning, which is a prerequisite for subsequent quantitative analysis and feature extraction.

[0077] The noise suppression step involves processing the radiometrically corrected image using a Lee-Sigma adaptive filtering algorithm based on local statistical properties. This algorithm estimates the statistical characteristics of noise within a local window of the image and dynamically adjusts the filtering intensity accordingly, thereby smoothing speckle noise while preserving as much edge and detail information as possible. After this step, the denoised image is obtained.

[0078] The geometric correction process involves processing the denoised image using the satellite's orbital parameters and the digital elevation model data of the covered area at the time of image acquisition. This processing includes two parts: terrain correction and geocoding. Terrain correction aims to eliminate pixel position distortion caused by terrain undulations; geocoding maps the image pixel coordinates to a unified geodetic coordinate system. Through this step, the image is ultimately corrected to the WGS84 geographic coordinate system, resulting in a second image with precise geolocation.

[0079] Furthermore, the method for calculating image features for each sample region, including backscattering intensity features, polarization decomposition features, and texture features, and establishing a training sample library based on the image features of all sample regions includes:

[0080] The sample areas are numbered and designated as the first sample area to the Nth sample area, where N is the number of sample areas. A mapping relationship is established between the sample areas and the sample area types, including seaweed sample areas, seawater sample areas, ship sample areas, and marine structure sample areas.

[0081] Extract the backscattering coefficient values ​​of all pixels from the first sample region to the Nth sample region from the second image; calculate the arithmetic mean of the backscattering coefficient values ​​of all pixels from the first sample region to the Nth sample region, and denot it as the first backscattering intensity feature to the Nth backscattering intensity feature.

[0082] From the second image, the coherence matrix of each pixel belonging to the first sample region to the Nth sample region is extracted; the coherence matrix of each pixel is decomposed using a polarimetric target decomposition algorithm to obtain the entropy value, anisotropy value, and average scattering angle value corresponding to each pixel; the arithmetic mean of the entropy values ​​of all pixels in the first sample region to the Nth sample region is calculated and denoted as the first entropy value feature to the Nth entropy value feature; the arithmetic mean of the anisotropy values ​​of all pixels in the first sample region to the Nth sample region is calculated and denoted as the first anisotropy value feature to the Nth anisotropy value feature; the arithmetic mean of the average scattering angle values ​​of all pixels in the first sample region to the Nth sample region is calculated and denoted as the first average scattering angle value feature to the Nth average scattering angle value feature; the first entropy value feature to the Nth entropy value feature, the first anisotropy value feature to the Nth anisotropy value feature, and the first average scattering angle value feature to the Nth average scattering angle feature are combined to obtain the first polarization decomposition feature to the Nth polarization decomposition feature.

[0083] Calculate the texture features of the first sample region to the Nth sample region respectively, and denote them as the first texture feature to the Nth texture feature.

[0084] Based on the mapping relationship between sample regions and sample region types, establish mapping relationships between the first backscattering intensity feature to the Nth backscattering intensity feature, the first polarization decomposition feature to the Nth polarization decomposition feature, the first texture feature to the Nth texture feature, and the sample region type. Combine the first backscattering intensity feature to the Nth backscattering intensity feature, the first polarization decomposition feature to the Nth polarization decomposition feature, the first texture feature to the Nth texture feature, and the corresponding sample region type into N sample records. Each sample record corresponds to a sample region, and each sample record contains the sample region type and the backscattering intensity feature, polarization decomposition feature, and texture feature corresponding to the sample region.

[0085] For example, 132 sample regions were selected, including 38 regions of seaweed, 45 regions of seawater, 28 regions of ships, and 21 regions of marine structures. These 132 sample regions are designated as the first to the 132nd sample regions. The type of each sample region was pre-collected manually. Backscattering coefficient values ​​of all pixels within each sample region were extracted from the second image, and the arithmetic mean of these values ​​was calculated to obtain the backscattering intensity feature of the corresponding sample region. This feature reflects the overall average scattering capability of the region as a single value. The coherence matrix of all pixels within each sample region was extracted from the second image, and the Cloude-Pottier decomposition algorithm was used to process each coherence matrix to obtain the entropy, anisotropy, and average scattering angle value of each pixel. Subsequently, the arithmetic mean of the entropy, anisotropy, and average scattering angle values ​​of all pixels within each sample region was calculated. The three average values ​​were combined to form a polarization decomposition feature representing the scattering mechanism characteristics of the sample region. The texture features of each sample region were calculated and recorded. Finally, based on the established mapping relationship between sample region types, the backscattering intensity features, polarization decomposition features, and texture features of each sample region are combined with their corresponding sample region types to form sample records. A total of 132 sample records are collected to obtain the training sample library. Taking the first sample region as an example, the sample record corresponding to the first sample region includes: {first backscattering intensity feature, first polarization decomposition feature, first texture feature, sample region type: *Ulva prolifera* sample region}.

[0086] Further, the method for calculating the texture features of the first sample region to the Nth sample region respectively, denoted as the first texture feature to the Nth texture feature, includes:

[0087] A square window with a pre-defined size of M rows by M columns is used as the neighborhood window, where M is a pre-defined odd number. Texture feature calculation operations are performed on the first sample region to the Nth sample region respectively. The steps of the texture feature calculation operation are as follows: According to the neighborhood window, for each pixel in the kth sample region, an index calculation step is performed to obtain the contrast index value, homogeneity index value, and information entropy index value of each pixel in the kth sample region; the arithmetic mean of the contrast index value, homogeneity index value, and information entropy index value of all pixels in the kth sample region is calculated to obtain the kth contrast index value, kth homogeneity index value, and kth information entropy index value; k is iterated through to obtain the first contrast index value to the Nth contrast index value, the first homogeneity index value to the Nth homogeneity index value, and the first information entropy index value to the Nth information entropy index value; the first contrast index value to the Nth contrast index value, the first homogeneity index value to the Nth homogeneity index value, and the first information entropy index value to the Nth information entropy index value are combined to obtain the first texture feature to the Nth texture feature.

[0088] For example, the preset value of M is 5, thereby achieving a balance between representing the local texture structure and computational complexity.

[0089] Furthermore, the method for calculating the performance index includes:

[0090] Centered on the pixel to be calculated, the backscattering coefficient values ​​of all pixels within the coverage area of ​​the neighborhood window of the pixel to be calculated are extracted to obtain the gray-level co-occurrence matrix; based on the gray-level co-occurrence matrix, the contrast index value, the homogeneity index value, and the information entropy index value of the pixel to be calculated are obtained.

[0091] Furthermore, the method of extracting the backscattering coefficient values ​​of all pixels within the coverage area of ​​the neighborhood window of the pixel to be calculated, centered on the pixel to be calculated, and calculating the gray-level co-occurrence matrix includes:

[0092] Centered on the pixel to be calculated, extract the backscattering coefficient values ​​of all pixels within the coverage area of ​​the neighborhood window of the pixel to be calculated; search for the minimum and maximum values ​​of the backscattering coefficient values ​​of all pixels within the coverage area of ​​the neighborhood window of the pixel to be calculated, and denote them as follows: , ;according to , The intensity levels of all pixels within the neighborhood window coverage area of ​​the pixel to be calculated are obtained by using the intensity calculation formula, along with the pre-set total number of intensity quantization levels L. The intensity levels of all pixels within the neighborhood window coverage area of ​​the pixel to be calculated are statistically analyzed to obtain the gray-level co-occurrence matrix.

[0093] The strength calculation formula is as follows:

[0094] ;

[0095] in, Indicates the intensity level of a pixel; This represents the backscattering coefficient value of a pixel; Indicates to Round down to the nearest integer.

[0096] For example, when the pixel to be calculated is located at the edge of the second image, and the neighborhood window of the pixel to be calculated extends beyond the range of the second image, a mirror filling method is used to obtain the backscattering coefficient value of the virtual pixel. That is, the pixel values ​​within the image boundary are reflected by mirroring to fill the portion of the neighborhood window that extends beyond the boundary, thereby ensuring that each pixel to be calculated can obtain complete neighborhood window data for subsequent calculations. The total number of intensity quantization levels is set to 16.

[0097] Furthermore, the method for statistically analyzing the intensity levels of all pixels within the neighborhood window coverage area of ​​the pixel to be calculated to obtain the gray-level co-occurrence matrix includes:

[0098] Within the neighborhood window, traverse i (taking integers from 1 to L) and j (taking integers from 1 to L) to search for the number of pixel pairs that satisfy the first matching statistical condition, denoted as . The number of pixel pairs that satisfy the second matching statistical condition is denoted as . ,Will Divide by get ; Let be the element in the i-th row and j-th column of the gray-level co-occurrence matrix.

[0099] The first matching statistical condition is that within the neighborhood window, in a pixel pair consisting of two pixels, one pixel has an intensity level of i and the other pixel has an intensity level of j, and the spatial distance between the two pixels is equal to a preset standard pixel distance d, and the angle between the line connecting the two pixels and a preset horizontal line is equal to a preset standard angle. .

[0100] The second matching statistical condition is that within the neighborhood window, in a pixel pair consisting of two pixels, the spatial distance between the two pixels is equal to a preset standard pixel distance d, and the angle between the line connecting the two pixels and a preset horizontal line is equal to a preset standard angle. .

[0101] Further, the method for calculating the contrast index value, homogeneity index value, and information entropy index value of the pixel to be calculated based on the gray-level co-occurrence matrix includes:

[0102] The contrast ratio, homogeneity, and information entropy of the pixel to be calculated are obtained based on the gray-level co-occurrence matrix.

[0103] The formula for calculating the contrast index value of the pixel to be calculated is as follows:

[0104] ;

[0105] in, This represents the contrast ratio value of the pixel to be calculated.

[0106] The formula for calculating the homogeneity index value of the pixels to be calculated is as follows:

[0107] ;

[0108] in, This represents the homogeneity index value of the pixels to be calculated.

[0109] The formula for calculating the information entropy index value of the pixel to be calculated is as follows:

[0110] ;

[0111] in, This represents the information entropy index value of the pixel to be calculated.

[0112] Further, the method of training a maximum likelihood classifier based on the training sample library, and using the trained maximum likelihood classifier to classify each pixel in the second image to obtain a preliminary classification result image of *Ulva prolifera*, denoted as the third image, includes:

[0113] Based on the sample region type, backscattering intensity features, polarization decomposition features, and texture features of all sample records in the training sample library, the maximum likelihood estimation algorithm is used to calculate the feature mean vector and covariance matrix of each category in the categories of seaweed, seawater, ships, and marine structures. The feature mean vector and covariance matrix of each category are set as the built-in parameters of the pre-set maximum likelihood classifier to obtain the trained maximum likelihood classifier.

[0114] Calculate the backscattering intensity feature, polarization decomposition feature, and texture feature of each pixel in the second image; input the backscattering intensity feature, polarization decomposition feature, and texture feature of each pixel in the second image into the trained maximum likelihood classifier, and output the likelihood probabilities of the seaweed category, seawater category, ship category, and marine structure category, respectively denoted as the likelihood probability of the seaweed category, the likelihood probability of the seawater category, the likelihood probability of the ship category, and the likelihood probability of the marine structure category; the category corresponding to the maximum value among the likelihood probabilities of the seaweed category, the likelihood probability of the seawater category, the likelihood probability of the ship category, and the likelihood probability of the marine structure category is taken as the category to which the pixel belongs.

[0115] Traverse all pixels in the second image and calculate the category of each pixel; record and spatially arrange the categories of all pixels to generate the third image.

[0116] Example 2: Based on the same inventive concept, such as Figure 2 As shown, this embodiment also provides a SAR Ulva prolifera remote sensing image monitoring system based on the maximum likelihood classification method. The system includes, in sequence, a data reading module, a training sample library construction module, a maximum likelihood classification module, and an image post-processing module.

[0117] The data reading module is used to acquire the original SAR image of the monitoring area, denoted as the first image. The original SAR image of the monitoring area is a fully polarimetric SAR image. The first image is subjected to radiometric correction, noise reduction and geometric correction to obtain a preprocessed SAR image, denoted as the second image.

[0118] The training sample library construction module is used to select pre-identified seaweed sample areas, seawater sample areas, ship sample areas, and marine structure sample areas from the second image; calculate the image features of each sample area, including backscattering intensity features, polarization decomposition features, and texture features; and establish a training sample library based on the image features of all sample areas.

[0119] The maximum likelihood classification module is used to train a maximum likelihood classifier based on the training sample library, and use the trained maximum likelihood classifier to classify each pixel in the second image to obtain a preliminary classification result image of Ulva prolifera, which is denoted as the third image; the pixel regions that are classified as Ulva prolifera in the third image are identified and denoted as Ulva prolifera pixel regions.

[0120] The image post-processing module is used to perform image post-processing on the third image to obtain an optimized distribution map of Ulva prolifera; the image post-processing includes removing noise points and merging adjacent Ulva prolifera pixel regions.

[0121] It should be noted that the specific methods by which each module performs operations in the system described in the above embodiments have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0122] Finally, it should be noted that although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A SAR remote sensing image monitoring method for *Ulva prolifera* based on maximum likelihood classification, characterized in that, The method includes the following steps: S1, acquire the original SAR image of the monitoring area, denoted as the first image, the original SAR image of the monitoring area is a fully polarimetric SAR image; perform radiometric correction, noise reduction and geometric correction processing on the first image to obtain the preprocessed SAR image, denoted as the second image; S2, from the second image, select the pre-identified seaweed sample area, seawater sample area, ship sample area, and marine structure sample area; calculate the image features of each sample area, including backscattering intensity features, polarization decomposition features, and texture features; and establish a training sample library based on the image features of all sample areas. S3. Based on the training sample library, train a maximum likelihood classifier, and use the trained maximum likelihood classifier to classify each pixel in the second image to obtain a preliminary classification result image of Ulva prolifera, which is denoted as the third image; the pixel regions that are classified as Ulva prolifera in the third image are identified and denoted as Ulva prolifera pixel regions. S4, perform image post-processing on the third image to obtain an optimized distribution map of Ulva prolifera; the image post-processing includes removing noise points and merging adjacent Ulva prolifera pixel regions.

2. The SAR remote sensing image monitoring method for *Ulva prolifera* based on maximum likelihood classification as described in claim 1, characterized in that, The method of performing radiometric correction, noise reduction, and geometric correction on the first image to obtain a preprocessed SAR image, denoted as the second image, includes: The first image is radiometrically calibrated. The radiometric calibration steps are as follows: calculate the gray value of each pixel in the first image, convert the gray value of each pixel in the first image into a backscattering coefficient value representing the scattering ability of ground objects according to the calibration parameters attached to the satellite sensor, and reconstruct the radiometrically corrected image using the backscattering coefficient value. The radiometrically corrected image is subjected to noise suppression; the noise suppression step is to use an adaptive speckle noise filtering algorithm based on local statistical characteristics to filter out speckle noise from the radiometrically corrected image, thereby obtaining a noise-reduced image. The denoised image is then geometrically corrected. The geometric correction steps involve using satellite orbital parameters and digital elevation model data to perform terrain correction and geocoding on the denoised image, thereby correcting the denoised image to a pre-specified geographic coordinate system to obtain a second image.

3. The SAR remote sensing image monitoring method for *Ulva prolifera* based on maximum likelihood classification as described in claim 2, characterized in that... The method of calculating image features for each sample region, including backscattering intensity features, polarization decomposition features, and texture features, and establishing a training sample library based on the image features of all sample regions includes: The sample areas are numbered and designated as the first sample area to the Nth sample area, where N is the number of sample areas. A mapping relationship between sample areas and sample area types is established, where the sample area types include seaweed sample areas, seawater sample areas, ship sample areas, and marine structure sample areas. Extract the backscattering coefficient values ​​of all pixels from the first sample region to the Nth sample region from the second image; calculate the arithmetic mean of the backscattering coefficient values ​​of all pixels from the first sample region to the Nth sample region, and denot it as the first backscattering intensity feature to the Nth backscattering intensity feature; From the second image, the coherence matrix of each pixel belonging to the first sample region to the Nth sample region is extracted; the coherence matrix of each pixel is decomposed using a polarimetric target decomposition algorithm to obtain the entropy value, anisotropy value, and average scattering angle value corresponding to each pixel; the arithmetic mean of the entropy values ​​of all pixels in the first sample region to the Nth sample region is calculated and denoted as the first entropy value feature to the Nth entropy value feature; the arithmetic mean of the anisotropy values ​​of all pixels in the first sample region to the Nth sample region is calculated and denoted as the first anisotropy value feature to the Nth anisotropy value feature; the arithmetic mean of the average scattering angle values ​​of all pixels in the first sample region to the Nth sample region is calculated and denoted as the first average scattering angle value feature to the Nth average scattering angle value feature; the first entropy value feature to the Nth entropy value feature, the first anisotropy value feature to the Nth anisotropy value feature, and the first average scattering angle value feature to the Nth average scattering angle value feature are combined to obtain the first polarization decomposition feature to the Nth polarization decomposition feature; Calculate the texture features of the first sample region to the Nth sample region respectively, and denote them as the first texture feature to the Nth texture feature; Based on the mapping relationship between sample regions and sample region types, establish mapping relationships between the first backscattering intensity feature to the Nth backscattering intensity feature, the first polarization decomposition feature to the Nth polarization decomposition feature, the first texture feature to the Nth texture feature, and the sample region type. Combine the first backscattering intensity feature to the Nth backscattering intensity feature, the first polarization decomposition feature to the Nth polarization decomposition feature, the first texture feature to the Nth texture feature, and the corresponding sample region type into N sample records. Each sample record corresponds to a sample region, and each sample record contains the sample region type and the backscattering intensity feature, polarization decomposition feature, and texture feature corresponding to the sample region.

4. The SAR Ulva prolifera remote sensing image monitoring method based on maximum likelihood classification as described in claim 3, characterized in that, The method for calculating the texture features of the first sample region to the Nth sample region respectively, denoted as the first texture feature to the Nth texture feature, includes: A square window with a pre-defined size of M rows by M columns is used as the neighborhood window, where M is a pre-defined odd number. Texture feature calculation operations are performed on the first sample region to the Nth sample region respectively. The steps of the texture feature calculation operation are as follows: According to the neighborhood window, for each pixel in the kth sample region, an index calculation step is performed to obtain the contrast index value, homogeneity index value, and information entropy index value of each pixel in the kth sample region; the arithmetic mean of the contrast index value, homogeneity index value, and information entropy index value of all pixels in the kth sample region is calculated to obtain the kth contrast index value, kth homogeneity index value, and kth information entropy index value; k is iterated through to obtain the first contrast index value to the Nth contrast index value, the first homogeneity index value to the Nth homogeneity index value, and the first information entropy index value to the Nth information entropy index value; the first contrast index value to the Nth contrast index value, the first homogeneity index value to the Nth homogeneity index value, and the first information entropy index value to the Nth information entropy index value are combined to obtain the first texture feature to the Nth texture feature.

5. The SAR Ulva prolifera remote sensing image monitoring method based on maximum likelihood classification as described in claim 4, characterized in that, The method for calculating the performance indicators includes: Centered on the pixel to be calculated, the backscattering coefficient values ​​of all pixels within the coverage area of ​​the neighborhood window of the pixel to be calculated are extracted to obtain the gray-level co-occurrence matrix; based on the gray-level co-occurrence matrix, the contrast index value, the homogeneity index value, and the information entropy index value of the pixel to be calculated are obtained.

6. The SAR remote sensing image monitoring method for *Ulva prolifera* based on maximum likelihood classification as described in claim 5, characterized in that... The method for calculating the gray-level co-occurrence matrix by extracting the backscattering coefficient values ​​of all pixels within the coverage area of ​​a neighborhood window of the pixel to be calculated, centered on the pixel to be calculated, includes: Centered on the pixel to be calculated, extract the backscattering coefficient values ​​of all pixels within the coverage area of ​​the neighborhood window of the pixel to be calculated; search for the minimum and maximum values ​​of the backscattering coefficient values ​​of all pixels within the coverage area of ​​the neighborhood window of the pixel to be calculated, and denote them as follows: , ;according to , The intensity levels of all pixels within the neighborhood window coverage area of ​​the pixel to be calculated are obtained by using the intensity calculation formula, along with the pre-set total number of intensity quantization levels L; the intensity levels of all pixels within the neighborhood window coverage area of ​​the pixel to be calculated are statistically analyzed to obtain the gray-level co-occurrence matrix. The strength calculation formula is as follows: ； in, Indicates the intensity level of a pixel; This represents the backscattering coefficient value of a pixel; Indicates to Round down to the nearest integer.

7. The SAR remote sensing image monitoring method for *Ulva prolifera* based on maximum likelihood classification as described in claim 6, characterized in that... The method for obtaining the gray-level co-occurrence matrix by statistically analyzing the intensity levels of all pixels within the neighborhood window coverage area of ​​the pixel to be calculated includes: Within the neighborhood window, traverse i (taking integers from 1 to L) and j (taking integers from 1 to L) to search for the number of pixel pairs that satisfy the first matching statistical condition, denoted as . The number of pixel pairs that satisfy the second matching statistical condition is denoted as . ,Will Divide by get ; The element in the i-th row and j-th column of the gray-level co-occurrence matrix; The first matching statistical condition is that within the neighborhood window, in a pixel pair consisting of two pixels, one pixel has an intensity level of i and the other pixel has an intensity level of j, and the spatial distance between the two pixels is equal to a preset standard pixel distance d, and the angle between the line connecting the two pixels and a preset horizontal line is equal to a preset standard angle. ; The second matching statistical condition is that within the neighborhood window, in a pixel pair consisting of two pixels, the spatial distance between the two pixels is equal to a preset standard pixel distance d, and the angle between the line connecting the two pixels and a preset horizontal line is equal to a preset standard angle. .

8. The SAR remote sensing image monitoring method for *Ulva prolifera* based on maximum likelihood classification as described in claim 7, characterized in that, The method for calculating the contrast index value, homogeneity index value, and information entropy index value of the pixel to be calculated based on the gray-level co-occurrence matrix includes: The contrast index, homogeneity index, and information entropy index of the pixel to be calculated are obtained based on the gray-level co-occurrence matrix. The formula for calculating the contrast index value of the pixel to be calculated is as follows: ； in, This represents the contrast ratio value of the pixel to be calculated. The formula for calculating the homogeneity index value of the pixels to be calculated is as follows: ； in, This represents the homogeneity index value of the pixels to be calculated; The formula for calculating the information entropy index value of the pixel to be calculated is as follows: ； in, This represents the information entropy index value of the pixel to be calculated.

9. The SAR remote sensing image monitoring method for *Ulva prolifera* based on maximum likelihood classification as described in claim 8, characterized in that, The method of training a maximum likelihood classifier based on the training sample library, and using the trained maximum likelihood classifier to classify each pixel in the second image to obtain a preliminary classification result image of *Ulva prolifera*, denoted as the third image, includes: Based on the sample region type, backscattering intensity features, polarization decomposition features, and texture features of all sample records in the training sample library, the maximum likelihood estimation algorithm is used to calculate the feature mean vector and covariance matrix of each category in the categories of seaweed, seawater, ships, and marine structures. The feature mean vector and covariance matrix of each category are set as the built-in parameters of the pre-set maximum likelihood classifier to obtain the trained maximum likelihood classifier. Calculate the backscattering intensity feature, polarization decomposition feature, and texture feature of each pixel in the second image; input the backscattering intensity feature, polarization decomposition feature, and texture feature of each pixel in the second image into the trained maximum likelihood classifier, and output the likelihood probabilities of the seaweed category, seawater category, ship category, and marine structure category, respectively denoted as the likelihood probability of the seaweed category, the likelihood probability of the seawater category, the likelihood probability of the ship category, and the likelihood probability of the marine structure category; the category corresponding to the maximum value among the likelihood probabilities of the seaweed category, the likelihood probability of the seawater category, the likelihood probability of the ship category, and the likelihood probability of the marine structure category is taken as the category to which the pixel belongs. Traverse all pixels in the second image and calculate the category of each pixel; record and spatially arrange the categories of all pixels to generate the third image.

10. A SAR remote sensing image monitoring system for *Ulva prolifera* based on maximum likelihood classification, characterized in that: The system is used to perform the method according to any one of claims 1-9, characterized in that the system comprises, in sequence, a data reading module, a training sample library construction module, a maximum likelihood classification module, and an image post-processing module; The data reading module is used to acquire the original SAR image of the monitoring area, denoted as the first image. The original SAR image of the monitoring area is a fully polarimetric SAR image. The first image is subjected to radiometric correction, noise reduction and geometric correction to obtain a preprocessed SAR image, denoted as the second image. The training sample library construction module is used to select pre-identified seaweed sample areas, seawater sample areas, ship sample areas, and marine structure sample areas from the second image; calculate the image features of each sample area, including backscattering intensity features, polarization decomposition features, and texture features; and establish a training sample library based on the image features of all sample areas. The maximum likelihood classification module is used to train a maximum likelihood classifier based on the training sample library, and use the trained maximum likelihood classifier to classify each pixel in the second image to obtain a preliminary classification result image of Ulva prolifera, which is denoted as the third image; the pixel regions that are classified as Ulva prolifera in the third image are identified and denoted as Ulva prolifera pixel regions. The image post-processing module is used to perform image post-processing on the third image to obtain an optimized distribution map of Ulva prolifera; the image post-processing includes removing noise points and merging adjacent Ulva prolifera pixel regions.

Images