Mangrove forest extraction method based on orthogonal matched filtering-weighted least square

Abstract

The invention discloses a mangrove forest extraction method based on orthogonal matched filtering-weighted least square, and relates to the technical field of remote sensing and image processing. According to the method, through preprocessing and band expansion, the whitened orthogonal subspace projection and the whitened matched filter method are combined, and mangrove forest feature information is effectively enhanced; then, a weighted least square filtering technology is introduced, and the edge information of the mangrove forest image is further enhanced; and finally, carrying out adaptive threshold segmentation on the remote sensing image by using an Otsu method so as to realize target extraction of the mangrove forest region. The method has the advantages of low cost, high precision, short time and the like. The problems that in the prior art, the labor cost is high, the efficiency is low, a neural network method lacks enough representativeness and coverage, the mangrove forest cannot be effectively recognized and extracted under various complex backgrounds, and the adaptability and the accuracy are not high are solved.

Description

Technical Field

[0001] The invention belongs to the technical field of remote sensing and image processing, and in particular relates to a mangrove extraction method based on orthogonal matching filtering-weighted least squares. Background Art

[0002] As one of the important marine-terrestrial interactive ecosystems, the health and stability of the mangrove ecosystem is crucial to maintaining marine biodiversity and marine ecological balance. Therefore, in order to achieve the sustainable development of mangroves, it is necessary to strengthen management and supervision, take scientific environmental protection measures, ensure the benign interaction of mangroves, protect the marine ecological environment, and achieve sustainable use of resources.

[0003] At present, the efficiency of traditional manual inspection and supervision is too slow, and multiple field measurements are required, which consumes a lot of manpower and material resources and gradually cannot meet the growing monitoring needs. In recent years, with the development of remote sensing technology, the accuracy of optical sensors has become higher and higher. Remote sensing has gradually become one of the most important monitoring methods with its advantages of large scale, short cycle, real-time monitoring and low cost. For ecosystems such as mangroves in nature reserves, remote sensing monitoring has unique advantages. The continuity of its sensor system and image acquisition capabilities can quickly obtain the spatial distribution information of mangroves. The Gaofen series of satellites can play an active and important role in improving my country's natural resource asset property rights system and use management system, demarcating ecological protection red lines, implementing resource paid use system and ecological compensation system, and reforming the ecological environment protection management system. The Gaofen-6 (GF-6) remote sensing satellite is a typical representative of them. It has the characteristics of high resolution, wide coverage, high-quality imaging, high-efficiency imaging, and high localization rate. As an important branch in the field of remote sensing data processing, remote sensing image target detection has developed many target detection methods, such as spectral angle mapping, orthogonal subspace projection, matched filter and constrained energy minimization.

[0004] Although manual visual interpretation can obtain accurate results, it is labor-intensive and inefficient, and cannot effectively cope with large-scale data. Although neural networks have developed rapidly in recent years, they often require large-scale data sets for sufficient training before they can play a role in practical applications. Although there are some training data sets, these existing data sets are still difficult to meet the growing demand for remote sensing applications. Existing data sets are often limited to specific regions or specific crop types and lack sufficient representativeness and coverage. In addition, the quality and annotation accuracy of the data set also affect the performance and generalization ability of the neural network model. When facing large-area data, the running time also seriously constrains the need for rapid observation. At present, research on remote sensing image data processing focuses on data enhancement and semi-supervised learning. Data enhancement technology improves the utilization rate and generalization ability of the model by transforming and expanding existing data, while semi-supervised learning uses a small amount of labeled data and a large amount of unlabeled data to improve the generalization ability and adaptability of the model through self-supervised or semi-supervised learning methods. However, this method still has problems such as inability to effectively identify and extract mangroves under various complex backgrounds, and low adaptability and accuracy.

[0005] Based on this, the present invention proposes a mangrove extraction method based on orthogonal matching filtering-weighted least squares, aiming to solve the problems existing in the above-mentioned prior art and provide more reliable technical support for mangrove resource monitoring and protection. Summary of the invention

[0006] In view of this, the main purpose of the present invention is to provide a mangrove extraction method based on orthogonal matching filtering-weighted least squares to solve the problems of high labor cost and low efficiency in the above-mentioned prior art, lack of sufficient representativeness and coverage of the neural network method, inability to effectively identify and extract mangroves under various complex backgrounds, and low adaptability and accuracy; it effectively improves the extraction performance under complex backgrounds of mangroves, and has the advantages of high efficiency, simplicity, reliability, and high accuracy.

[0007] To achieve the above object, the technical solution of the present invention is achieved as follows:

[0008] A mangrove extraction method based on orthogonal matched filtering-weighted least squares, comprising:

[0009] Step 1: Obtain the original multispectral remote sensing image, and preprocess and fuse the remote sensing image;

[0010] Step 2: Separate land and sea from multispectral fusion remote sensing images;

[0011] Step 3: After separating the land and sea of ​​the multispectral fusion remote sensing image, the average spectrum of the prior information is obtained in the spectral library based on the spectral information of the mangroves collected multiple times;

[0012] Step 4: After obtaining the prior information, construct the characteristic index through band expansion, and further distinguish the mangroves through the characteristic index;

[0013] Step 5: Perform two-dimensional and normalized processing on the remote sensing image data obtained by step 4;

[0014] Step 6: Based on the orthogonal subspace projection and matched filter method, whitening and regularization terms are added to the remote sensing image data processed in step 5;

[0015] Step 7: By combining the orthogonal subspace projection optimized in step 6 with the matched filter method, a new orthogonal matched filtering method is obtained;

[0016] Step 8: Introduce the weighted least squares method into the new orthogonal matched filter method obtained in step 7 to obtain a mangrove extraction method based on orthogonal matched filter-weighted least squares. Finally, use the Otsu method for threshold segmentation to achieve the recognition and extraction of mangroves.

[0017] As a preferred embodiment of the present invention, the process of preprocessing and fusing the remote sensing image in step 1 includes:

[0018] Step 1.1: Select the area to be acquired, obtain the original multispectral remote sensing image of the area, and preprocess the image;

[0019] Among them, the process of preprocessing the original multispectral remote sensing images includes radiometric calibration, atmospheric correction and orthorectification processes;

[0020] Step 1.2: After preprocessing the original multispectral remote sensing image, image fusion is performed based on the multispectral image and the panchromatic image to obtain a high-resolution multispectral fused remote sensing image;

[0021] Among them, the fusion method of image fusion is Gram-Schmidt method.

[0022] As a preferred embodiment of the present invention, when performing land-sea separation processing on the multispectral fusion remote sensing image in step 2, the vector elements of the multispectral fusion remote sensing image are first edited using an embedded GIS, and finally mask processing is performed.

[0023] As a preferred embodiment of the present invention, the characteristic index in step 4 includes normalized vegetation index, enhanced vegetation index, normalized difference water index and total suspended matter concentration index.

[0024] As a preferred embodiment of the present invention, the process of two-dimensionalizing and normalizing the remote sensing image data in step 5 includes:

[0025] Step 5.1: Convert the three-dimensional multispectral image into a two-dimensional matrix and map the data between 0 and 1, removing the dimensions and dimensional units of data of different dimensions;

[0026] Step 5.2: Then, perform average normalization on the data to achieve the effect of centering the data at 0, and then standardize the data; the specific process is:

[0027] The known covariance matrix is:

[0028]

[0029] Since the data has been normalized, the mean Therefore, formula (1) is transformed into:

[0030]

[0031] Then perform eigenvalue decomposition on the covariance matrix C and calculate the eigenvalues ​​and eigenvectors:

[0032] C=UAU T (3)

[0033] In formula (3), U is the eigenvector, Λ is a diagonal matrix, and the elements on the diagonal are the eigenvalues.

[0034] As a preferred embodiment of the present invention, the solution process of the orthogonal subspace projection method in step 5 includes:

[0035] Step 6.1.1: First, let X be the L×1-dimensional pixel spectral vector, K represent the dimension of the endmember spectral matrix, α i represents the abundance of the corresponding end member, and the expression is:

[0036]

[0037] The OSP model consists of the target signal d and the background end member U, so we get:

[0038]

[0039] In formula (5), is the abundance vector of the target signal, is the abundance vector of background endmembers, n is noise or model error;

[0040] Step 6.1.2: Introduce an orthogonal subspace projector, defined as:

[0041]

[0042] In formula (6), ⊥ is the orthogonal complement space, I is the K×K dimensional unit matrix, (U T U) -1 U T is the pseudo-inverse matrix of the U matrix, U T Represents the transposition of the U matrix;

[0043] The constructed filter detector is:

[0044]

[0045] As a preferred embodiment of the present invention, the solution process of the matched filter method in step 5 includes:

[0046] Step 6.2.1: Assume that the image to be identified has B bands, and the vector form of each pixel is:

[0047] x s =[x1, x2, x3, …, x B ] T (s=1,2,3,…,S) (8)

[0048] In formula (8), S = LEN × WID, LEN is the length of the image size, WID is the width of the image size, and the sensor of each band will obtain a grayscale image G of size S b :

[0049] G b (len,wid)(b=1,2,3,…,B;len=1,2,3,…,LEN;wid=1,2,3,…,WID)(9)

[0050] Step 6.2.2: Rearrange the 3D image data into:

[0051] X = [x1, x2, x3, ..., x S ] (10)

[0052] It is expressed as:

[0053] X(b,s)(b=1,2,3,…,B; s=1,2,3,…,S) (11)

[0054] The relationship between formula (9) and formula (11) is:

[0055] G b (len,wid)=X(b,(len-1)WID+wid) (12)

[0056] Step 6.2.3: Based on the prior information, the matched filter (MF) algorithm for single target detection is formulated as a quadratic programming problem:

[0057]

[0058] In formula (13), μ1 is the mean vector of the target pixel, μ0 is the mean vector of the background pixel, and W is the linear filter;

[0059] The Lagrange multiplier method is introduced to solve the unknown vector W:

[0060]

[0061] The covariance matrix of the background and the background mean vector According to the available data, it is estimated that:

[0062]

[0063] The linear detector has the form:

[0064]

[0065] The target mean vector is obtained by averaging the existing spectral library or label information:

[0066]

[0067] Step 6.2.4: Combining the data from equations (14) to (18), perform Mahalanobis distance calculation between the target and the background to obtain:

[0068]

[0069] As a preferred embodiment of the present invention, the process of adding whitening and regularization terms to the remote sensing image data in step 5 includes:

[0070] Step 6.3.1: By combining the features x of each dimension rot,i Divide by the square root of its corresponding eigenvalue To scale the data, we get the data matrix after PCA whitening:

[0071]

[0072] Adding the regularization term constant ε to the data matrix, the final result of the PCA whitening matrix is:

[0073]

[0074] Step 6.3.2: Perform whitening weighted operation on equations (5)-(7) and (21) to obtain:

[0075]

[0076] So the detector is redefined as:

[0077]

[0078] Step 6.3.3: Orthogonal subspace projector in (23) Construct a sparse matrix and obtain the sample space based on whitening:

[0079]

[0080] So MF is rewritten as:

[0081]

[0082] In formula (25), is a constant.

[0083] As a preferred embodiment of the present invention, the process of combining the orthogonal subspace projection and matched filter method optimized in step 6 to obtain the orthogonal matched filtering method in step 7 includes:

[0084] Step 7.1: First, use the new measurement method to find the projection vector W. In the whitened space, W is found by minimizing and obtaining the new output:

[0085]

[0086] Step 7.2: Next, replace d in equation (22) white As the target vector in the matched filter A new orthogonal matched filtering optimization model is formed, and the constrained minimum problem becomes:

[0087]

[0088] Step 7.3: Finally, the new orthogonal matched filter optimization model is expanded and simplified as follows:

[0089]

[0090] In formula (28), Σ OMF Represents the autocovariance matrix of the weighted whitened samples. In the whitened space, the covariance matrix is ​​the unit matrix. Solving the new orthogonal matched filter optimization model yields:

[0091]

[0092] Therefore, the final output of OMF is:

[0093]

[0094] As a preferred embodiment of the present invention, the process of introducing the weighted least squares method into the new orthogonal matched filtering method obtained in step 7 in step 8 includes:

[0095] Step 8.1.1: Perform weighted least squares filtering on the image after whitening weighted filtering:

[0096]

[0097] In formula (31), u is the filtered image to be solved, λ is the regularization parameter, and D x , D y is the discrete difference operator, A x , A y is a diagonal matrix;

[0098] Step 8.1.2: Derivative the above equation (31) to obtain the minimum value of u:

[0099] u=(I+λL yOMF ) -1 y OMF (32)

[0100] In formula (32), is a Laplacian matrix;

[0101] At this point, the final result of u is the output of the orthogonal matched filter-weighted least squares method.

[0102] Compared with the prior art, the present invention provides a mangrove extraction method based on orthogonal matching filtering-weighted least squares, which has the following beneficial effects:

[0103] 1. This method is based on the orthogonal subspace projection (OSP), matched filter (MF) and weighted least squares filtering (WLS) methods, and adds optimization means such as whitening and regularization terms to construct an orthogonal matched filter-weighted least squares method. This method uses the principle of enhancing the target information of mangroves and weakening the background information. The Otsu method (OTSU) is used for threshold segmentation, and the extraction results are displayed by binarization. Finally, various indicators are used for calibration and verification. It has the advantages of low cost, high accuracy, and short time. It solves the problems of high labor cost and low efficiency in the existing technology, lack of sufficient representativeness and coverage of neural network methods, inability to effectively identify and extract mangroves under various complex backgrounds, and low adaptability and accuracy.

[0104] 2. This method uses the orthogonal matching filter-weighted least squares method to identify and extract mangrove features, which can effectively identify and extract mangroves under various complex backgrounds, and has the advantages of good adaptability and high accuracy;

[0105] 3. Unlike machine learning, this method only requires traditional mathematical calculations of the target spectral information, does not require a lot of learning and training, and has fast calculation speed and high efficiency;

[0106] 4. This method can better handle complex situations under different backgrounds and reduce misidentification and false positives by designing a reasonable weighting mechanism. At the same time, the edges of mangroves usually have a clear contrast with the surrounding environment. The method of the present invention can effectively retain these edge features while smoothing the internal information of large tracts of mangroves in the image, which makes the extraction of mangrove boundaries clearer and more accurate, and can provide more reliable technical support for the monitoring and protection of mangrove resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0107] Figure 1 This is a flow chart of the mangrove extraction method based on orthogonal matching filtering-weighted least squares of the present invention;

[0108] Figure 2 Three mangrove areas selected in the map for the present invention;

[0109] Figure 3 For the present invention Figure 2 The original image of the pre-processed large-scale area A selected in the middle;

[0110] Figure 4 For the present invention Figure 2 The original image of the pre-processed mesoscale area B selected in the middle box;

[0111] Figure 5 For the present invention Figure 2 The original image of the preprocessed small-scale area C selected in the middle;

[0112] Figure 6 This is a diagram showing the enhancement effect of the method of the present invention on mangroves before filtering using the weighted least squares method;

[0113] Figure 7 This is a diagram showing the enhancement effect of the method of the present invention on mangroves after filtering using the weighted least squares method;

[0114] Figure 8 The method of the present invention and other methods Figure 7 Binarized comparison image after Otsu threshold segmentation based on .

[0115] Among them, Figure 6 In the figure, Figure (a) shows the enhancement effect of mangroves in region A before filtering by weighted least squares method, Figure (b) shows the enhancement effect of mangroves in region B before filtering by weighted least squares method, and Figure (c) shows the enhancement effect of mangroves in region C before filtering by weighted least squares method.

[0116] exist Figure 7 In the figure, Figure (a) shows the enhancement effect of mangroves in region A after filtering by weighted least squares method, Figure (b) shows the enhancement effect of mangroves in region B after filtering by weighted least squares method, and Figure (c) shows the enhancement effect of mangroves in region C after filtering by weighted least squares method.

[0117] exist Figure 8 In FIG. (a), the present invention is Figure 2 The original image of area A in FIG. (b) is the original image of area A in FIG. Figure 2 The true value diagram of region A in FIG. 1 is shown in FIG. 1 , and FIG. (c) is shown in FIG. 1 . Figure 2 The orthogonal matched filter-weighted least squares method (OMF-WLS) extraction diagram of region A in FIG. 1 is shown in FIG. 2 . Figure 2 The matched filter (MF) extraction diagram of region A in FIG. 1 is shown in FIG. 1 , and FIG. (e) is shown in FIG. 1 of the present invention. Figure 2 The constrained energy minimization (CEM) extraction diagram of region A in FIG. 1 is a diagram of the present invention in FIG. Figure 2 The adaptive coherence estimator (ACE) extraction diagram of region A in FIG. 1 is shown in FIG. 1 , and FIG. (g) is the adaptive coherence estimator (ACE) extraction diagram of region A in FIG. Figure 2 The maximum likelihood estimation (MLE) extraction diagram of region A in FIG. 1 is shown in FIG. 1 . FIG. (h) is the maximum likelihood estimation (MLE) extraction diagram of region A in FIG. 1 . Figure 2 The original image of area B in FIG. (i) is the original image of area B in FIG. Figure 2 The true value diagram of region B in FIG. (j) is the true value diagram of the present invention in Figure 2 The orthogonal matched filter-weighted least squares method (OMF-WLS) extraction diagram of region B in FIG. (k) is the present invention in Figure 2 The matched filter (MF) extraction diagram of region B in FIG. (l) is the present invention in Figure 2 The constrained energy minimization (CEM) extraction diagram of region B in FIG. (m) is the present invention in Figure 2 The adaptive coherence estimator (ACE) extraction diagram of region B in FIG. (n) is the present invention in Figure 2 The maximum likelihood estimation (MLE) extraction diagram of region B in FIG. (o) is the present invention in Figure 2 The original image of area C in FIG. (p) is the original image of area C in FIG. Figure 2 The truth graph of region C in FIG. (q) is the truth graph of the present invention in Figure 2 The orthogonal matched filter-weighted least squares method (OMF-WLS) extraction diagram of region C in FIG. (r) is the present invention in Figure 2 The matched filter (MF) extraction diagram of region C in FIG. (s) is the present invention in Figure 2 The constrained energy minimization (CEM) extraction diagram of region C in FIG. (t) is the present invention in Figure 2 The adaptive coherence estimator (ACE) extraction diagram of region C in FIG. 1 is the adaptive coherence estimator (ACE) extraction diagram of region C in FIG. 1 , and FIG. 1 is the adaptive coherence estimator (ACE) extraction diagram of region C in FIG. 1 . Figure 2 Maximum likelihood estimation (MLE) extraction plot for region C in the middle. DETAILED DESCRIPTION

[0118] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0119] Embodiment 1:

[0120] Please refer to the instruction manual Figure 1 -Attached Figure 8 , the present invention provides a technical solution:

[0121] A mangrove extraction method based on orthogonal matched filtering-weighted least squares, comprising:

[0122] Step 1: Obtain the original multispectral remote sensing image, and preprocess and fuse the remote sensing image. The specific process includes:

[0123] Step 1.1: Please refer to the attached Figure 2 , select three areas (see Appendix Figure 3 , Attachment Figure 4 , Attachment Figure 5 ), obtain its original multispectral remote sensing image and preprocess the image;

[0124] Among them, the process of preprocessing the original multispectral remote sensing images includes radiometric calibration, atmospheric correction and orthorectification processes;

[0125] Step 1.2: After preprocessing the original multispectral remote sensing image, image fusion is performed based on the multispectral image and the panchromatic image to obtain a high-resolution multispectral fused remote sensing image;

[0126] Among them, the fusion method used for image fusion is the Gram-Schmidt (GS) method;

[0127] Step 2: In order to eliminate the interference factors of the ocean, the multispectral fusion remote sensing image is separated into land and sea;

[0128] The process of land-sea separation of multispectral fusion remote sensing images includes: firstly, using embedded GIS (ArcGIS) to edit the vector elements of multispectral fusion remote sensing images, and finally performing mask processing;

[0129] Step 3: After separating the land and sea of ​​the multispectral fusion remote sensing image, the average spectrum of the prior information is obtained in the spectral library based on the spectral information of the mangroves collected multiple times;

[0130] Among them, the spectral library is a .sli file obtained using ENVI, imported into MATLAB, and converted into a matrix form;

[0131] Step 4: Through band expansion, construct characteristic index to strengthen the target spectral characteristics, and further distinguish mangroves through characteristic index;

[0132] Among them, the constructed characteristic indices include the normalized difference vegetation index (NDVI), enhanced vegetation index (EVI), normalized difference water index (NDWI) and total suspended matter concentration index (TSM);

[0133] Step 5: Perform two-dimensional and normalized processing on the remote sensing image data to complete data preprocessing. The specific process includes:

[0134] Step 5.1: Convert the three-dimensional multispectral image into a two-dimensional matrix and map the data between 0 and 1 to remove the dimensions and dimensional units of data of different dimensions;

[0135] Step 5.2: Then, the data is averaged and normalized to achieve the effect of centering the data at 0, and then the data is standardized to put the features on the same scale; the covariance matrix is ​​known to be:

[0136]

[0137] Since the data has been normalized, the mean Therefore, formula (1) can be transformed into:

[0138]

[0139] Then perform eigenvalue decomposition on the covariance matrix C and calculate the eigenvalues ​​and eigenvectors:

[0140] C=UAU T (3)

[0141] In formula (3), U is the eigenvector, Λ is a diagonal matrix, and the elements on the diagonal are the eigenvalues;

[0142] Step 6: Based on the orthogonal subspace projection and matched filter method, optimization operations such as whitening and regularization are added to the remote sensing image data processed in step 5;

[0143] Among them, the solution process of orthogonal subspace projection and matched filter method includes:

[0144] Step 6.1: The solution process of the Orthogonal Subspace Projection (OSP) method is:

[0145] Step 6.1.1: First, let X be the L×1-dimensional pixel spectral vector, K represents the dimension of the endmember spectral matrix, where α i represents the abundance of the corresponding end member, and the expression is:

[0146]

[0147] The OSP model is composed of the target signal d and the background end element U, so we can get:

[0148]

[0149] In formula (5), is the abundance vector of the target signal, is the abundance vector of background endmembers, n is noise or model error;

[0150] Step 6.1.2: In order to extract the target signal of interest and suppress all other background features, an orthogonal subspace projector is introduced, where ⊥ represents the orthogonal complementary space and is defined as follows:

[0151]

[0152] The constructed filter detector can be expressed as:

[0153]

[0154] In formula (6), I is the identity matrix of K×K dimensions, (U T U) -1 U T is the pseudo-inverse matrix of the U matrix, where U T represents the transposition of the U matrix; from formula (7), we can see that when extracting endmembers, prior knowledge of image endmembers is required;

[0155] Step 6.2: The solution process of the matched filter (MF) method is

[0156] Step 6.2.1: Assuming that the image to be identified has B bands, the vector form of each pixel can be expressed as:

[0157] x s =[x1, x2, x3, …, x B ] T (s=1,2,3,…,S) (8)

[0158] In formula (8), S = LEN × WID, where LEN is the length of the image size, WID is the width of the image size, and the sensor of each band will obtain a grayscale image G of size S b :

[0159] G b (len,wid)(b=1,2,3,…,B; len=1,2,3,…,LEN; wid=1,2,3,…,WID) (9)

[0160] Step 6.2.2: The 3D image data can be rearranged as follows:

[0161] X = [x1, x2, x3, ..., x S ] (10)

[0162] So it can be expressed as:

[0163] X(b,s)(b=1,2,3,…,B; s=1,2,3,…,S) (11)

[0164] The relationship between formula (9) and formula (11) is:

[0165] G b (len,wid)=X(b,(len-1)WID+wid) (12)

[0166] Step 6.2.3: Based on the known prior information, the matched filter (MF) algorithm for single target detection can be reduced to a simple quadratic programming problem:

[0167]

[0168] In formula (13), μ1 is the mean vector of the target pixel, μ0 is the mean vector of the background pixel, and W is a linear filter. By introducing the Lagrange multiplier method, the unknown vector W is solved:

[0169]

[0170] The covariance matrix of the background and the background mean vector Based on the available data, it can be estimated that:

[0171]

[0172] The linear detector has the form:

[0173]

[0174] The target mean vector is obtained by averaging the existing spectral library or label information:

[0175]

[0176] Step 6.2.4: Combine the data from equations (14) to (18) and perform Mahalanobis distance calculation between the target and the background, and finally obtain:

[0177]

[0178] Step 6.3: The process of adding optimization operations such as whitening and regularization to the remote sensing image data processed in step 5 includes:

[0179] Step 6.3.1: By transforming each dimension of feature x rot，i Divide by the square root of its corresponding eigenvalue To scale our decorrelated data, we get the PCA whitened data matrix:

[0180]

[0181] Sometimes we encounter the eigenvalue λ i Very close to 0. This may cause numerical instability or overflow when dividing by a number close to zero. To solve this problem, the eigenvalues ​​are scaled and a small constant ε is added to ensure that division by values ​​close to zero is avoided. The value is generally ε≈10 -5 This constant is often called a regularization term. By adding this constant, we can ensure that we do not encounter numerical instability or overflow issues when performing division operations. So the final result of the PCA whitening matrix is:

[0182]

[0183] Step 6.3.2: Combine equations (5)-(7) and equation (21) to perform whitening weighted operation to obtain:

[0184]

[0185] So the detector can be redefined as:

[0186]

[0187] Step 6.3.3: Finally, the orthogonal subspace projector in (23) Construct a sparse matrix to effectively improve the target information and suppress the background information, and obtain the sample space based on whitening as follows:

[0188]

[0189] Therefore, MF can be rewritten as:

[0190]

[0191] In formula (25), is a constant;

[0192] Step 7: By combining the orthogonal subspace projection optimized in step 6 with the matched filter method, a new method, the orthogonal matched filter method, is obtained. The specific process includes:

[0193] Step 7.1: First, use a new measurement method to find the projection vector W. In the whitened space, W can be found by minimizing and get the new output:

[0194]

[0195] Step 7.2: Next, replace d in equation (22) white As the target vector in the matched filter A new orthogonal matching filter (OMF) optimization model is formed, and the constrained minimum problem becomes:

[0196]

[0197] Step 7.3: Finally, expand and simplify the model into:

[0198]

[0199] In formula (28), ∑ OMF Represents the autocovariance matrix of the weighted whitened samples. In the whitened space, the covariance matrix is ​​the unit matrix. Solving the model yields:

[0200]

[0201] Therefore, the final OMF output is:

[0202]

[0203] Step 8: In the new orthogonal matched filtering method obtained in step 7, the weighted least square method is introduced to obtain the final method of this embodiment - the mangrove extraction method based on orthogonal matched filtering-weighted least squares. Finally, the Otsu method is used for threshold segmentation to realize the recognition and extraction process of mangroves;

[0204] Step 8.1: In the new orthogonal matched filtering method, the weighted least squares method (Maximum Likelihood Estimation, MLE) is introduced to obtain the mangrove extraction method based on orthogonal matched filtering-weighted least squares. The process includes:

[0205] Step 8.1.1: Perform weighted least squares filtering on the image after whitening weighted filtering:

[0206]

[0207] In formula (31), u is the filtered image to be solved, and λ is the regularization parameter used to balance u and y OMF When the edge gradient of the input image changes greatly, we hope that the constraint is smaller to preserve the structural information of the image; when the edge gradient of the image changes slightly, we think that these details are not important, so the constraint can be larger. x , D y is the discrete difference operator, A x , A y is a diagonal matrix;

[0208] Step 8.1.2: By taking the derivative of equation (31), the minimum value of u is obtained as:

[0209] u=(I+λL yOMF ) -1 y OMF (32)

[0210] In formula (32), is a Laplacian matrix; so far, the final result of u is the output of the Orthogonal Matching Filter-Weighted Least Squares (OMF-WLS) method verified in this embodiment;

[0211] Step 8.2: Please refer to the attached Figure 8 , the filtered u image will be binarized and segmented based on the Otsu method to achieve the result of accurate extraction of mangrove targets.

[0212] Embodiment 2:

[0213] As the instruction manual Figure 1 -Attached Figure 8 As shown, in order to verify the feasibility of the mangrove extraction method based on orthogonal matching filtering-weighted least squares as described in Example 1, this example is designed for verification:

[0214] 1. Experimental Procedure

[0215] Using the method described in step 4 of Example 1, the method of constructing the characteristic index of mangroves is used to construct and strengthen the target spectral characteristics, so as to reflect and extract the ground feature characteristics of mangroves. These characteristic indexes are usually new image data obtained by mathematical operation and combination of different bands in remote sensing images, which are used to reflect certain attributes or conditions of surface features. The characteristic indexes constructed by the present invention are mainly divided into two categories, one is the vegetation index class, which mainly reflects the growth state and coverage of vegetation by combining infrared bands and visible light bands. Common vegetation indices include normalized difference vegetation index (NDVI) and the like. These indices can effectively distinguish mangroves from background information because mangroves usually have a high vegetation coverage. The second category is the water body index class, which mainly reflects the distribution and type of water bodies by combining blue light bands and near infrared bands. Common water body indices include normalized difference water body index (NDWI) and the like. Through these indices, waters and land can be effectively distinguished because water body indexes usually have higher values ​​in waters, while land areas are lower.

[0216] Since the PCA whitening in the method described in Example 1 of the present invention is implemented based on singular value decomposition (SVD), in order to avoid highly correlated features in the input data, singular values ​​close to zero are present in the singular value decomposition results, and even overfitting occurs. This embodiment will only select four indexes, namely NDVI (normalized vegetation index), EVI (enhanced vegetation index), NDWI (normalized difference water index) and TSM (total suspended matter concentration index), to further distinguish mangroves through these indexes. Finally, there will be 8 bands after the band expansion is completed. Table 1 shows the characteristic indexes after band expansion, where B1 represents the blue light band, B2 represents the green light band, B3 represents the red light band, and B4 represents the near infrared band.

[0217] Table 1: Characteristic index after band expansion

[0218]

[0219] 2. Results Analysis

[0220] In order to better quantitatively evaluate the recognition and extraction capabilities of mangroves, this example randomly selects samples from the research objects in the Yingluo Port area, averages them to obtain target prior information, and then uses ENVI to export the spectrum library. Combined with the selected sample data, the confusion matrix is ​​used to calculate the four indicators of mangrove overall accuracy (Overall Accuracy, OA), average accuracy (Average Accuracy, AA), producer accuracy (Producer's Accuracy, PA) and Kappa coefficient. The formula is as follows:

[0221]

[0222] Among them, TP (True Positive) represents true positive examples, that is, the number of examples predicted by the model as positive and the actual label is also positive; TN (True Negative) represents true negative examples, that is, the number of examples predicted by the model as negative and the actual label is also negative; FP (False Positive) represents false positive examples, that is, the number of examples predicted by the model as positive but the actual label is negative; FN (False Negative) represents false negative examples, that is, the number of examples predicted by the model as negative but the actual label is positive.

[0223] Therefore, through the above experiments Figure 8 Comparing the mangrove identification and extraction results before and after calibration between the method described in Example 1 of the present invention and other methods, it can be seen that the method described in Example 1 of the present invention can effectively improve the mangrove identification and extraction effect. Figure 8 (a) to (g) represent the original image of region A, the truth image, the orthogonal matched filter-weighted least squares method (OMF-WLS) extraction image of embodiment 1 of the present invention, the matched filter (MF) extraction image, the constrained energy minimization (CEM) extraction image, the adaptive coherence estimator (ACE) extraction image and the maximum likelihood estimation (MLE) extraction image. Similarly, Figure 8 (h)-(n) and Figure 8 (o)-(u) represent the extraction results of different methods for regions B and C with the same arrangement order as region A.

[0224] Finally, these indices (Table 2, Table 3, Table 4) were used to quantitatively evaluate the extraction effect of mangroves. The closer the value is to 1, the better the effect.

[0225] Table 2: Extraction results and indicators of Yingluo Port Area A

[0226]

[0227] Table 3: Extraction results and indicators of Yingluo Port area B

[0228]

[0229]

[0230] Table 4: Extraction results and indicators of Yingluo Port area C

[0231]

[0232] Therefore, it is proved that the method described in Example 1 of the present invention has various indicators above 94% in both large-scale areas and medium- and small-scale areas, and no matter what complex background is used, the extraction and identification effect of mangroves by the method described in Example 1 of the present invention tends to be stable and better than other methods.

[0233] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0234] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A mangrove extraction method based on orthogonal matched filtering-weighted least squares, characterized by: include: Step 1: Obtain the original multispectral remote sensing image, and preprocess and fuse the remote sensing image; Step 2: Separate land and sea from multispectral fusion remote sensing images; Step 3: After separating the land and sea of ​​the multispectral fusion remote sensing image, the average spectrum of the prior information is obtained in the spectral library based on the spectral information of the mangroves collected multiple times; Step 4: After obtaining the prior information, construct the characteristic index through band expansion, and further distinguish the mangroves through the characteristic index; Step 5: Perform two-dimensional and normalized processing on the remote sensing image data obtained by step 4; Step 6: Based on the orthogonal subspace projection and matched filter method, whitening and regularization terms are added to the remote sensing image data processed in step 5; Step 7: By combining the orthogonal subspace projection optimized in step 6 with the matched filter method, a new orthogonal matched filtering method is obtained; Step 8: Introduce the weighted least squares method into the new orthogonal matched filter method obtained in step 7 to obtain a mangrove extraction method based on orthogonal matched filter-weighted least squares. Finally, use the Otsu method for threshold segmentation to achieve the recognition and extraction of mangroves.

2. The method for extracting mangroves based on orthogonal matched filtering-weighted least squares as claimed in claim 1, characterized in that: The process of preprocessing and fusing the remote sensing image described in step 1 includes: Step 1.1: Select the area to be acquired, obtain the original multispectral remote sensing image of the area, and preprocess the image; Among them, the process of preprocessing the original multispectral remote sensing images includes radiometric calibration, atmospheric correction and orthorectification processes; Step 1.2: After preprocessing the original multispectral remote sensing image, image fusion is performed based on the multispectral image and the panchromatic image to obtain a high-resolution multispectral fused remote sensing image; Among them, the fusion method of image fusion is Gram-Schmidt method.

3. The method for extracting mangroves based on orthogonal matched filtering-weighted least squares as claimed in claim 1, characterized in that: When performing land-sea separation processing on the multispectral fusion remote sensing image as described in step 2, the vector elements of the multispectral fusion remote sensing image are first edited using the embedded GIS, and finally mask processing is performed.

4. The method for extracting mangroves based on orthogonal matched filtering-weighted least squares as claimed in claim 1, characterized in that: The characteristic indexes in step 4 include normalized vegetation index, enhanced vegetation index, normalized difference water index and total suspended matter concentration index.

5. The method for extracting mangroves based on orthogonal matched filtering-weighted least squares as claimed in claim 1, characterized in that: The process of two-dimensionalizing and normalizing the remote sensing image data described in step 5 includes: Step 5.1: Convert the three-dimensional multispectral image into a two-dimensional matrix and map the data between 0 and 1, removing the dimensions and dimensional units of data of different dimensions; Step 5.2: Then, perform average normalization on the data to achieve the effect of centering the data at 0, and then standardize the data; the specific process is: The known covariance matrix is: Since the data has been normalized, the mean Therefore, formula (1) is transformed into: Then perform eigenvalue decomposition on the covariance matrix C and calculate the eigenvalues ​​and eigenvectors: C=UΛU T (3) In formula (3), U is the eigenvector, Λ is a diagonal matrix, and the elements on the diagonal are the eigenvalues.

6. The method for extracting mangroves based on orthogonal matching filtering-weighted least squares as claimed in claim 1, characterized in that: The solution process of the orthogonal subspace projection method in step 5 includes: Step 6.1.1: First, let X be the L×1-dimensional pixel spectral vector, K represent the dimension of the endmember spectral matrix, α i Represents the abundance of the corresponding end member, and the expression is: The OSP model consists of the target signal d and the background end member U, so we get: In formula (5), is the abundance vector of the target signal, is the abundance vector of background endmembers, n is noise or model error; Step 6.1.2: Introduce an orthogonal subspace projector, defined as: In formula (6), ⊥ is the orthogonal complement space, I is the K×K dimensional unit matrix, (U T U) -1 U T is the pseudo-inverse matrix of the U matrix, U T Represents the transposition of the U matrix; The constructed filter detector is:

7. The method for extracting mangroves based on orthogonal matched filtering-weighted least squares as claimed in claim 1, characterized in that: The solution process of the matched filter method described in step 5 includes: Step 6.2.1: Assume that the image to be identified has B bands, and the vector form of each pixel is: x s =[x1,x2,x3,…,x B ] T (s=1,2,3,…,S) (8) In formula (8), S = LEN × WID, LEN is the length of the image size, WID is the width of the image size, and the sensor of each band will obtain a grayscale image G of size S b : G b (len,wid)(b=1,2,3,…,B;len=1,2,3,…,LEN;wid=1,2,3,…,WID) (9) Step 6.2.2: Rearrange the 3D image data into: X=[x1,x2,x3,…,x S ] (10) It is expressed as: X(b,s)(b=1,2,3,…,B; s=1,2,3,…,S) (11) The relationship between formula (9) and formula (11) is: G b (len,wid)=X(b,(len-1)WID+wid) (12) Step 6.2.3: Based on the prior information, the matched filter algorithm for single target detection is classified as a quadratic programming problem: In formula (13), μ1 is the mean vector of the target pixel, μ0 is the mean vector of the background pixel, and W is the linear filter; The Lagrange multiplier method is introduced to solve the unknown vector W: The covariance matrix of the background and the background mean vector According to the available data, it is estimated that: The linear detector has the form: The target mean vector is obtained by averaging the existing spectral library or label information: Step 6.2.4: Combining the data from equations (14) to (18), perform Mahalanobis distance calculation between the target and the background to obtain:

8. The method for extracting mangroves based on orthogonal matching filtering-weighted least squares as claimed in claim 6, characterized in that: The process of adding whitening and regularization terms to the remote sensing image data in step 5 includes: Step 6.3.1: By combining the features x of each dimension rot,i Divide by the square root of its corresponding eigenvalue To scale the data, we get the data matrix after PCA whitening: Adding the regularization term constant ε to the data matrix, the final result of the PCA whitening matrix is: Step 6.3.2: Perform whitening weighted operation on equations (5)-(7) and (21) to obtain: So the detector is redefined as: Step 6.3.3: For the orthogonal subspace projector P in equation (23) ⊥ U Construct a sparse matrix and obtain the sample space based on whitening: So MF is rewritten as: In formula (25), is a constant.

9. The method for extracting mangroves based on orthogonal matched filtering-weighted least squares as claimed in claim 1, characterized in that: The process of combining the orthogonal subspace projection and matched filter method optimized in step 6 to obtain the orthogonal matched filtering method in step 7 includes: Step 7.1: First, use the new measurement method to find the projection vector W. In the whitened space, W is found by minimizing and obtaining the new output: Step 7.2: Next, replace d in equation (22) white As the target vector in the matched filter A new orthogonal matched filtering optimization model is formed, and the constrained minimum problem becomes: Step 7.3: Finally, the new orthogonal matched filter optimization model is expanded and simplified as follows: In formula (28), Σ OMF Represents the autocovariance matrix of the weighted whitened samples. In the whitened space, the covariance matrix is ​​the unit matrix. Solving the new orthogonal matched filter optimization model yields: Therefore, the final output of OMF is:

10. The method for extracting mangroves based on orthogonal matched filtering-weighted least squares according to claim 1, characterized in that: The process of introducing the weighted least squares method into the new orthogonal matched filtering method obtained in step 7 in step 8 includes: Step 8.1.1: Perform weighted least squares filtering on the image after whitening weighted filtering: In formula (31), u is the filtered image to be solved, λ is the regularization parameter, and D x , D y is the discrete difference operator, A x , A y is a diagonal matrix; Step 8.1.2: Derivative the above equation (31) to obtain the minimum value of u: In formula (32), is a Laplacian matrix; At this point, the final result of u is the output of the orthogonal matched filter-weighted least squares method.

Images