Coastal wetland intelligent interpretation and classification method based on spectral remote sensing data

Abstract

The invention discloses a coastal wetland intelligent interpretation and classification method based on spectral remote sensing data, and belongs to the technical field of image processing, and the method comprises the following steps: S1, collecting a spectral remote sensing image, and extracting global spatial spectrum information of a global image block; s2, acquiring local spatial spectrum information of the local image blocks; s3, determining prediction probabilities of ground feature categories corresponding to the global spatial spectrum information and the local spatial spectrum information; and S4, optimizing the coding network and the image block segmentation network to complete wetland classification. According to the method, data characteristics of hyperspectral and multispectral spectral remote sensing images are fully considered, a local image segmentation network is specifically constructed to randomly generate local image blocks, redundant neighborhood information is effectively eliminated, a coding network is constructed to respectively extract spatial spectrum characteristics of global image blocks and the local image blocks, a classification probability is obtained through a classifier, and the classification efficiency is improved. And stopping training the model until the classification probability is greater than a threshold value.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to a method for intelligent interpretation and classification of coastal wetlands from spectral remote sensing data. Background Technology

[0002] Remote sensing technology possesses large-scale and high-frequency observation capabilities, offering significant advantages in monitoring coastal wetland land cover types. Currently, coastal wetland monitoring methods based on remote sensing data primarily rely on manual interpretation, a process dependent on extensive experience and knowledge, and is time-consuming and labor-intensive. There is an urgent need to improve the automation and intelligence of remote sensing monitoring. In coastal wetland areas, ecological vegetation is the main land cover type. Compared to visible light and synthetic aperture radar, spectral remote sensing methods, primarily hyperspectral and multispectral, can provide richer information. Spectral remote sensing, based on the "red edge effect" of vegetation in the near-infrared band, can more accurately identify vegetation. Therefore, accurate identification of coastal wetland land cover types based on spectral remote sensing data is an effective way to promote the informatization and intelligent protection and management of wetlands.

[0003] Taking pixel-by-pixel spectral remote sensing image classification as an example, target spectral classification models, which use a single spectrum as input, are highly susceptible to noise interference. Meanwhile, the performance of spatial spectral classification models based on convolutional neural networks is limited by factors such as fixed convolution size and interference from neighboring unrelated pixels, which can easily lead to serious problems such as unclear classification boundaries and fragmentation of ground features. Therefore, there is an urgent need for intelligent interpretation models of ground feature types with high accuracy and strong generalization. Summary of the Invention

[0004] To address the problems of low accuracy in wetland feature classification and difficulty in boundary delineation in existing methods, this invention proposes an intelligent interpretation and classification method for coastal wetlands based on spectral remote sensing data.

[0005] The technical solution of this invention is: a method for intelligent interpretation and classification of coastal wetlands based on spectral remote sensing data, comprising the following steps:

[0006] S1. Acquire spectral remote sensing images and extract global spatial spectral information from global image patches;

[0007] S2. Obtain the local spatial spectrum information of the local image patch based on the global spatial spectrum information of the global image patch;

[0008] S3. Determine the predicted probability of the land cover category corresponding to the global spatial spectrum information and the local spatial spectrum information;

[0009] S4. Based on the predicted probabilities of land cover categories corresponding to global and local spatial spectral information, optimize the coding network and image patch segmentation network to complete wetland classification.

[0010] Furthermore, S1 includes the following sub-steps:

[0011] S11. Acquire spectral remote sensing images, perform 2D convolution operation on the image patches of the spectral remote sensing images, and extract global spatial information of the global image patches;

[0012] S12. Perform a 1D convolution operation on the spectral information of the center pixel of the global image patch to extract the global spectral information of the global image patch.

[0013] S13. Perform a weighted summation of the global spatial information and global spectral information of the global image patch to obtain the global spatial-spectral information of the global image patch.

[0014] Furthermore, in S11, the global spatial information of the global image patch. The expression is:

[0015] ;

[0016] in, This is the weight matrix of the first convolutional layer. The bias of the first convolutional layer. A global image patch of hyperspectral data;

[0017] In S12, global spectral information of the global image patch. The expression is:

[0018] ;

[0019] in, This is the weight matrix of the second convolutional layer. This is the bias of the second convolutional layer;

[0020] In S13, global spatial spectrum information of the global image patch. The expression is:

[0021] ;

[0022] in, This represents a set of learnable parameters.

[0023] Furthermore, S2 includes the following sub-steps:

[0024] S21. Perform channel dimension feature reduction on the global spatial spectral information of the global image patch to obtain the dimension-reduced features;

[0025] S22. Based on the global spatial spectrum information of the global image patch and the reduced features, obtain the position of the center pixel of the global image patch through a gated loop unit;

[0026] S23. Based on the position of the center pixel of the global image block, obtain the local image block, and obtain the local spatial spectrum information through the local coding network.

[0027] Furthermore, in S21, the features after dimensionality reduction The expression is:

[0028] ;

[0029] in, The weights of a 1×1 convolutional layer For the bias of a 1×1 convolutional layer, This refers to the global spatial spectrum information of the global image patch;

[0030] In S22, the expression for obtaining the position of the center pixel of the image patch is:

[0031] ;

[0032] in, This is the output of the GRU hidden layer at the previous time step. The normalized coordinates of the center of the local image patch at the next time step. The equation for calculating GRU at the current time is given.

[0033] Furthermore, S3 includes the following sub-steps:

[0034] S31. Vectorize the global spatial spectrum information of the global image patch through global average pooling operation to obtain the predicted probability of land cover category;

[0035] S32. Vectorize the local spatial spectral information of local image patches through global average pooling operation to obtain the predicted probability of land cover categories.

[0036] Furthermore, in S31, the predicted probability of the land cover category corresponding to the global image patch. The expression is:

[0037] ;

[0038] in, This is a global average pooling operation. This is the output of the GRU hidden layer at the previous time step. For ground feature classifier, This refers to the global spatial spectral information of the image patch;

[0039] In S32, the predicted probability of the land cover category corresponding to the local image patch. The expression is:

[0040] ;

[0041] in, This represents the local spatial spectral information of the image patch.

[0042] Furthermore, S4 includes the following sub-steps:

[0043] S41. Randomly generate local image patches and optimize the parameters of the coding network and the ground cover classifier;

[0044] S42. Optimize the image patch segmentation network by fixing the weights of the coding network and the land cover classifier;

[0045] S43. Fix the weights of the fixed image patch segmentation network, adjust the encoding network and classifier, and adjust the prediction accuracy of the prediction probability of land cover categories.

[0046] Furthermore, the classification loss function of the land cover classifier The expression is:

[0047] ;

[0048] in, One-hot encoding for modal tags, To use a land cover classifier The obtained predicted probability, It is a logarithmic function. This refers to the global spatial spectral information or the local spatial spectral information of an image patch.

[0049] Discriminative loss constraints are designed for wetland feature characteristics to improve the discriminative power of features between different classes.

[0050] Discrimination loss function of land cover classifier The expression is:

[0051] ;

[0052] in, The cosine similarity between two input vectors is represented by the following expression: Represents a logical function. Indicates an index. Indicates the number of input vectors;

[0053] ,if and If they belong to the same category, then The value is 1 if they do not belong to the same category. It is 0.

[0054] Optimization of the overall loss function of the land cover classifier The expression is:

[0055] ;

[0056] The image patch segmentation network maximizes the total reward through a policy gradient algorithm to determine the optimal local image patch location; the total reward at time t+1... The expression is:

[0057] ;

[0058] in, Let be the probability of correctly predicting the class at time t+1. Let be the probability of correctly predicting the class at time t;

[0059] The image patch segmentation network optimizes by maximizing the discounted reward, and its expression is:

[0060] ;

[0061] in, As a discount reward, Let be the total reward at time t. For a predefined discount factor, This represents the total number of moments.

[0062] The beneficial effects of this invention are:

[0063] (1) This invention fully considers the data characteristics of hyperspectral and multispectral remote sensing images, constructs a local image segmentation network to randomly generate local image blocks, effectively eliminates redundant neighborhood information, constructs an encoding network to extract the spatial spectral features of global image blocks and local image blocks respectively, obtains the classification probability through a classifier, and stops training the model until the classification probability is greater than the threshold.

[0064] (2) The present invention designs a two-stage training strategy to optimize the encoding network and the image block segmentation network. First, local image blocks are generated through random local image block segmentation operations to train the image encoding network and classifier. Second, the weights of the image encoding network and classifier are fixed to train the image block segmentation network to obtain the optimal local image blocks. Attached Figure Description

[0065] Figure 1 A flowchart of a method for intelligent interpretation and classification of coastal wetlands based on spectral remote sensing data;

[0066] Figure 2 This is a schematic diagram of the image classification method of the present invention;

[0067] Figure 3 This is a schematic diagram of the spatial-spectral feature extraction network structure of the present invention;

[0068] Figure 4 This is a schematic diagram of the image block segmentation network structure of the present invention. Detailed Implementation

[0069] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0070] like Figure 1 As shown, this invention provides a method for intelligent interpretation and classification of coastal wetlands based on spectral remote sensing data, comprising the following steps:

[0071] S1. Acquire spectral remote sensing images and extract global spatial spectral information from global image patches;

[0072] S2. Obtain the local spatial spectrum information of the local image patch based on the global spatial spectrum information of the global image patch;

[0073] S3. Determine the predicted probability of the land cover category corresponding to the global spatial spectrum information and the local spatial spectrum information;

[0074] S4. Based on the predicted probabilities of land cover categories corresponding to global and local spatial spectral information, optimize the coding network and image patch segmentation network to complete wetland classification.

[0075] In this embodiment of the invention, S1 includes the following sub-steps:

[0076] S11. Acquire spectral remote sensing images, perform 2D convolution operation on the image patches of the spectral remote sensing images, and extract global spatial information of the global image patches;

[0077] S12. Perform a 1D convolution operation on the spectral information of the center pixel of the global image patch to extract the global spectral information of the global image patch.

[0078] S13. Perform a weighted summation of the global spatial information and global spectral information of the global image patch to obtain the global spatial-spectral information of the global image patch.

[0079] Take the area around the corresponding pixel in the input spectral remote sensing image. The neighboring pixels represent the spatial spectral information of the center pixel. It takes the value of an odd number that is not less than 15;

[0080] Perform N-layer 2D convolution operations on the image patch. The value is an odd number not less than 5, used to extract the global spatial information of the input image patch; N-layer 1D convolution operation is performed on the spectral information of the center pixel of the image patch. The value is an odd number not less than 5, which is used to extract the global spectral information of the input image patch.

[0081] In this embodiment of the invention, in S11, the global spatial information of the global image patch... The expression is:

[0082] ;

[0083] in, This is the weight matrix of the first convolutional layer. The bias of the first convolutional layer. A global image patch of hyperspectral data;

[0084] In S12, global spectral information of the global image patch. The expression is:

[0085] ;

[0086] in, This is the weight matrix of the second convolutional layer. This is the bias of the second convolutional layer;

[0087] In S13, global spatial spectrum information of the global image patch. The expression is:

[0088] ;

[0089] in, This represents a set of learnable parameters.

[0090] In this embodiment of the invention, S2 includes the following sub-steps:

[0091] S21. Perform channel dimension feature reduction on the global spatial spectral information of the global image patch to obtain the dimension-reduced features;

[0092] S22. Based on the global spatial spectrum information of the global image patch and the reduced features, obtain the position of the center pixel of the global image patch through a gated loop unit;

[0093] S23. Based on the position of the center pixel of the global image block, obtain the local image block, and obtain the local spatial spectrum information through the local coding network.

[0094] Use one layer Convolution performs dimensionality reduction on channel-dimensional features, preserving spatial information while reducing unnecessary channel information interference. The size of local image patches is less than... For odd numbers, the center pixel position for the next step is randomly selected from the parameterized distribution by inputting the feature map from the previous step. Local image patches are then obtained based on these center pixel positions. Local spatial spectral features are obtained through a local coding network. The local coding network has the same structure as the global coding network, and the parameters are independent.

[0095] In this embodiment of the invention, in S21, the dimensionality-reduced features The expression is:

[0096] ;

[0097] in, The weights of a 1×1 convolutional layer For the bias of a 1×1 convolutional layer, This refers to the global spatial spectrum information of the global image patch;

[0098] In S22, the expression for obtaining the position of the center pixel of the image patch is:

[0099] ;

[0100] in, This is the output of the GRU hidden layer at the previous time step. The normalized coordinates of the center of the local image patch at the next time step. The equation for calculating GRU at the current time is given.

[0101] In this embodiment of the invention, S3 includes the following sub-steps:

[0102] S31. Vectorize the global spatial spectrum information of the global image patch through global average pooling operation to obtain the predicted probability of land cover category;

[0103] S32. Vectorize the local spatial spectral information of local image patches through global average pooling operation to obtain the predicted probability of land cover categories.

[0104] The ground cover classifier performs gradient backpropagation using the cross-entropy loss function, while the image patch segmentation network maximizes the total reward through a policy gradient algorithm, thereby determining the optimal local image patch location. A probability threshold is set; when the predicted probability exceeds the threshold, the model stops training.

[0105] In this embodiment of the invention, in S31, the predicted probability of the land cover category corresponding to the global image patch. The expression is:

[0106] ;

[0107] in, This is a global average pooling operation. This is the output of the GRU hidden layer at the previous time step. For ground feature classifier, This refers to the global spatial spectral information of the image patch;

[0108] In S32, the predicted probability of the land cover category corresponding to the local image patch. The expression is:

[0109] ;

[0110] in, This represents the local spatial spectral information of the image patch.

[0111] In this embodiment of the invention, S4 includes the following sub-steps:

[0112] S41. Randomly generate local image patches and optimize the parameters of the coding network and the ground cover classifier;

[0113] S42. Optimize the image patch segmentation network by fixing the weights of the coding network and the land cover classifier;

[0114] S43. Fix the weights of the fixed image patch segmentation network, adjust the encoding network and classifier, and adjust the prediction accuracy of the prediction probability of land cover categories.

[0115] A two-stage training strategy is designed to optimize the encoding network and the image patch segmentation network: without using the image patch segmentation network, the parameters of the encoding network and the classifier are optimized; with the weights of the encoding network and the classifier fixed, the image patch segmentation network is optimized, and then the weights of the encoding network and the classifier are fine-tuned.

[0116] In this embodiment of the invention, the classification loss function of the land cover classifier is... The expression is:

[0117] ;

[0118] in, One-hot encoding for modal tags, To use a land cover classifier The obtained predicted probability, It is a logarithmic function. This refers to the global spatial spectral information or the local spatial spectral information of an image patch.

[0119] Discriminative loss constraints are designed for wetland feature characteristics to improve the discriminative power of features between different classes.

[0120] Discrimination loss function of land cover classifier The expression is:

[0121] ;

[0122] in, The cosine similarity between two input vectors is represented by the following expression: Represents a logical function. Indicates an index. Indicates the number of input vectors;

[0123] ,if and If they belong to the same category, then The value is 1 if they do not belong to the same category. It is 0.

[0124] Optimization of the overall loss function of the land cover classifier The expression is:

[0125] ;

[0126] The image patch segmentation network maximizes the total reward through a policy gradient algorithm to determine the optimal local image patch location; the total reward at time t+1... The expression is:

[0127] ;

[0128] in, Let be the probability of correctly predicting the class at time t+1. Let be the probability of correctly predicting the class at time t;

[0129] The image patch segmentation network optimizes by maximizing the discounted reward, and its expression is:

[0130] ;

[0131] in, As a discount reward, Let be the total reward at time t. For a predefined discount factor, This represents the total number of moments.

[0132] like Figure 2 As shown, the overall technical framework of this invention consists of an image patch segmentation network, a global coding network, a local coding network, and a classifier, enabling adaptive selection of the neighborhood size and range of image patches and accurate classification of land cover types.

[0133] like Figure 3 As shown, the global coding network and the local coding network have the same structural framework, including 2D convolution to extract spatial information, 1D convolution to extract spectral information, and then obtaining spatial-spectral information through adaptive fusion.

[0134] like Figure 4 As shown, the image patch segmentation network consists of convolutional layers and a GRU. The convolutional layers are used for feature dimensionality reduction, while the GRU combines the hidden layer outputs at different time steps to obtain the position of the center pixel of the image patch.

[0135] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A method for intelligent interpretation and classification of coastal wetlands from spectral remote sensing data, characterized in that, Includes the following steps: S1. Acquire spectral remote sensing images and extract global spatial spectral information from global image patches; S2. Obtain the local spatial spectrum information of the local image patch based on the global spatial spectrum information of the global image patch; S3. Determine the predicted probability of the land cover category corresponding to the global spatial spectrum information and the local spatial spectrum information; S4. Based on the predicted probabilities of land cover categories corresponding to global and local spatial spectral information, optimize the coding network and image patch segmentation network to complete wetland classification.

2. The method for intelligent interpretation and classification of coastal wetlands based on spectral remote sensing data according to claim 1, characterized in that, S1 includes the following sub-steps: S11. Acquire spectral remote sensing images, perform 2D convolution operation on the image patches of the spectral remote sensing images, and extract global spatial information of the global image patches; S12. Perform a 1D convolution operation on the spectral information of the center pixel of the global image patch to extract the global spectral information of the global image patch. S13. Perform a weighted summation of the global spatial information and global spectral information of the global image patch to obtain the global spatial-spectral information of the global image patch.

3. The method for intelligent interpretation and classification of coastal wetlands based on spectral remote sensing data according to claim 2, characterized in that, In S11, the global spatial information of the global image patch The expression is: ； in, This is the weight matrix of the first convolutional layer. The bias of the first convolutional layer. A global image patch of hyperspectral data; In S12, the global spectral information of the global image patch The expression is: ； in, This is the weight matrix of the second convolutional layer. This is the bias of the second convolutional layer; In S13, the global spatial spectrum information of the global image patch The expression is: ； in, This represents a set of learnable parameters.

4. The method for intelligent interpretation and classification of coastal wetlands based on spectral remote sensing data according to claim 1, characterized in that, S2 includes the following sub-steps: S21. Perform channel dimension feature reduction on the global spatial spectral information of the global image patch to obtain the dimension-reduced features; S22. Based on the global spatial spectrum information of the global image patch and the reduced features, obtain the position of the center pixel of the global image patch through a gated loop unit; S23. Based on the position of the center pixel of the global image block, obtain the local image block, and obtain the local spatial spectrum information through the local coding network.

5. The method for intelligent interpretation and classification of coastal wetlands based on spectral remote sensing data according to claim 4, characterized in that, In S21, the features after dimensionality reduction The expression is: ； in, The weights of a 1×1 convolutional layer For the bias of a 1×1 convolutional layer, This refers to the global spatial spectrum information of the global image patch; In step S22, the expression for obtaining the position of the center pixel of the image block is: ； in, This is the output of the GRU hidden layer at the previous time step. The normalized coordinates of the center of the local image patch at the next time step. The equation for calculating GRU at the current time is given.

6. The method for intelligent interpretation and classification of coastal wetlands based on spectral remote sensing data according to claim 1, characterized in that, S3 includes the following sub-steps: S31. Vectorize the global spatial spectral information of the global image patch through global average pooling operation to obtain the predicted probability of land cover category; S32. Vectorize the local spatial spectral information of local image patches through global average pooling operation to obtain the predicted probability of land cover categories.

7. The method for intelligent interpretation and classification of coastal wetlands based on spectral remote sensing data according to claim 6, characterized in that, In step S31, the predicted probability of the land cover category corresponding to the global image patch. The expression is: ； in, This is a global average pooling operation. This is the output of the GRU hidden layer at the previous time step. For ground feature classifier, This refers to the global spatial spectral information of the image patch; In step S32, the predicted probability of the land cover category corresponding to the local image patch. The expression is: ； in, This represents the local spatial spectral information of the image patch.

8. The method for intelligent interpretation and classification of coastal wetlands based on spectral remote sensing data according to claim 1, characterized in that, S4 includes the following sub-steps: S41. Randomly generate local image patches and optimize the parameters of the coding network and the ground cover classifier; S42. Optimize the image patch segmentation network by fixing the weights of the coding network and the land cover classifier; S43. Fix the weights of the fixed image patch segmentation network, adjust the encoding network and classifier, and adjust the prediction accuracy of the prediction probability of land cover categories.

9. The method for intelligent interpretation and classification of coastal wetlands based on spectral remote sensing data according to claim 8, characterized in that, The classification loss function of the land cover classifier The expression is: ； in, One-hot encoding for modal tags, To use a land cover classifier The obtained predicted probability, It is a logarithmic function. This refers to the global spatial spectral information or the local spatial spectral information of an image patch. The discriminant loss function of the land cover classifier The expression is: ； in, The cosine similarity between two input vectors is represented by the following expression: Represents a logical function. Indicates an index. Indicates the number of input vectors; The optimized overall loss function of the land cover classifier The expression is: ； The image patch segmentation network maximizes the total reward using a policy gradient algorithm to determine the optimal local image patch location; the total reward at time t+1... The expression is: ； in, Let be the probability of correctly predicting the class at time t+1. Let be the probability of correctly predicting the class at time t; The image patch segmentation network is optimized by maximizing the discount reward, and its expression is: ； in, As a discount reward, Let be the total reward at time t. For a predefined discount factor, This represents the total number of moments.

Images