A remote sensing multispectral image water body recognition method based on a dual-channel segmentation network

Abstract

The present application relates to a kind of water body identification method of remote sensing multispectral image based on dual-channel segmentation network. RGB remote sensing image and water body index are combined with dual-channel semantic segmentation network to solve the problem of water body identification of remote sensing multispectral image, the characteristics of RGB remote sensing image and water body index are used, by constructing dual-channel semantic segmentation neural network, multi-layer feature fusion module and attention module are introduced, the features of original remote sensing image and water body index are fused from different scales and different feature levels, the water body information is highlighted and other information such as vegetation, soil and building is suppressed to the greatest extent, and the identification accuracy of water body area is improved. The water body identification method of transmission line corridor remote sensing multispectral image based on the dual-channel semantic segmentation network described in the present application has good water body identification performance.

Description

Technical Field

[0001] This invention relates to the fields of deep learning, computer vision, and water body recognition in remote sensing images, and particularly to a method for water body recognition in remote sensing multispectral images based on a dual-channel segmentation network. Background Technology

[0002] Transmission lines are a crucial component of the power system, used to transport electrical energy from power plants to various users. The presence of water bodies in transmission line corridors can cause various problems during maintenance and management, such as equipment damage, decreased insulation performance, and increased accident risks. With the continuous development of the power system, the safety of transmission line corridors is receiving increasing attention. Therefore, early detection and accurate identification of water bodies in transmission line corridors are essential for ensuring the reliability and stability of the power system.

[0003] Traditional remote sensing methods typically rely on manual analysis or rule-based image processing techniques, which can be inefficient and susceptible to subjective interference. With the rapid rise of deep learning technology, automated water body identification methods based on semantic segmentation networks offer new opportunities to address this problem. This application introduces a dual-channel semantic segmentation network to better capture water features in remote sensing images of power transmission line corridors. This application employs two parallel channels, using the original remote sensing image and water body indices, to maximize the highlighting of water body information while suppressing information such as vegetation, soil, and buildings, thereby improving the accuracy of water body area identification. Summary of the Invention

[0004] The purpose of this invention is to provide a method for water body identification based on remote sensing multispectral images using a dual-channel segmentation network, in order to solve the problems of poor versatility and adaptability of existing technologies.

[0005] To achieve the above objectives, the technical solution of the present invention is: a method for water body identification based on remote sensing multispectral images using a dual-channel segmentation network, comprising the following steps:

[0006] Step S1: Data preprocessing. The collected remote sensing images are preprocessed, and the water bodies in the remote sensing images are labeled at the pixel level to obtain labels and establish a multispectral image water body dataset.

[0007] Step S2: Construct RGB image data by selecting the infrared (Red), green, and blue bands of the multispectral water body data and superimposing them to form a three-channel RGB image data.

[0008] Step S3: Construct water body index data. Water body index is an indicator used for water body detection and monitoring in remote sensing image analysis. By calculating the ratio or index between different bands, the spectral characteristics of water bodies are highlighted, thereby distinguishing them from other land features. Select near-infrared (NIR), infrared (Red), green, and shortwave infrared (SWIR) bands of multispectral image water body data, calculate the normalized difference vegetation index (NDVI), normalized difference water enhancement index (AWElnsh), and normalized difference water index (NDWI) respectively, and combine the three indices into a three-channel water body index data.

[0009] Step S4: Construct a dual-channel semantic segmentation neural network. The dual-channel semantic segmentation neural network includes a feature extraction module, a hierarchical feature fusion module, and a feature output module. The RGB image data generated in step S2 and the water index data generated in step S3 are used as the data inputs for the two channels, respectively. They are processed by the feature extraction module and the features of the two input channels are fused by the hierarchical feature fusion module to obtain deep features. The feature output module then outputs the water segmentation results of the multispectral image.

[0010] Step S5: Construct the loss function of the dual-channel semantic segmentation neural network, that is, use the improved Focal loss function to construct the loss function of the dual-channel semantic segmentation network;

[0011] Step S6: Train the dual-channel semantic segmentation neural network. Use the RGB image data generated in step S2 and the water index data generated in step S3 as input to the dual-channel semantic segmentation neural network, and use the labels marked in step S1 as output to train the dual-channel semantic segmentation neural network. Use PyTorch as the deep neural network learning framework and NVIDIA GeForce RTX4060 as the GPU for model training. Use the Adam optimizer, set the initial learning rate to 0.001, and use the hierarchical descent learning rate method. The training times are 200 epochs.

[0012] Step S7: Identify water bodies in multispectral images. After processing in steps S2 and S3, the multispectral images to be identified are input into the dual-channel semantic segmentation neural network trained in step S6 to obtain the results of multispectral image identification of water bodies.

[0013] In one embodiment of the present invention, the data preprocessing in step S1 specifically involves preprocessing the collected remote sensing images, including radiometric correction, atmospheric correction, and geographic correction, to eliminate noise and errors in the images.

[0014] In one embodiment of the present invention, in step S3, the Normalized Difference Vegetation Index (NDVI) is calculated as follows:

[0015] NDVI = (NIR - Red) / (NIR + Red)

[0016] In the formula: NDVI is the normalized vegetation index, NIR is the reflectance in the near-infrared band, and Red is the reflectance in the infrared band.

[0017] The normalized water enhancement index (AWElnsh) is calculated as follows:

[0018] AWElnsh=4*(Green-SWIR)-(0.25*NIR+2.75*SWIR)

[0019] In the formula: AWEElnsh is the normalized water enhancement index, Green is the reflectance of the green light band, SWIR represents the reflectance of the short-wave infrared band, and NIR is the reflectance of the near-infrared band.

[0020] The Normalized Difference Water Index (NDWI) is calculated as follows:

[0021] NDWI = (NIR - SWIR) / (NIR + SWIR)

[0022] In the formula: NDWI is the normalized differential water index, NIR is the reflectance in the near-infrared band, and SWIR represents the reflectance in the short-wave infrared band.

[0023] In one embodiment of the present invention, in step S4, the feature extraction module, in order to fully capture the spatial information of the input data and obtain data features of different scales from different levels, sequentially uses three networks consisting of two convolutional layers with a kernel size of 3×3 and a max pooling layer with a kernel size of 3×3, outputting three feature layers S1, S2, and S3 with feature sizes of 1 / 2, 1 / 4, and 1 / 8 of the input image. The feature layer obtained from the RGB image channel is S1. 11 S 12 S 13 The water index image channel yields a feature layer of S. 21 S 22 S 23 .

[0024] In one embodiment of the present invention, in step S4, the hierarchical feature extraction module extracts the feature layer S extracted by the feature extraction module. 11 S 12 S 13 S 21 S 22 S 23 All of them pass through an attention module, which learns weight allocation to enable the network to pay more attention to the features of the water body area, thereby improving the accuracy of water body area identification and improving the identification accuracy from low-resolution feature maps S. 11 and S 21The fusion process begins, and after upsampling, it merges with S. 12 and S 22 After fusion, and then upsampling, it is combined with S 13 and S 23 Fusion. The attention module performs max pooling and average pooling on the feature layer, then concatenates the results of these two pooling operations, and then performs global average pooling, convolutional layers, and activation functions. The resulting feature is multiplied by the input feature, and the weights are applied to each of the original feature spaces to obtain the new features of the attention mechanism.

[0025] In one embodiment of the present invention, in step S4, the feature output module consists of a global average pooling layer, two convolutional layers with a kernel size of 3×3 and a deconvolutional layer with a kernel size of 2×2. The deconvolutional layer is used to upsample the feature map of the last convolutional layer to obtain the same size as the input image. Finally, pixel-by-pixel classification is performed on the upsampled feature map.

[0026] In one embodiment of the present invention, the specific method for constructing the loss function of the dual-channel semantic segmentation neural network in step S5 is as follows: In order to effectively solve the problems of imbalance between positive and negative samples and imbalance between easy and difficult classification samples, an improved Focal loss function is used, the mathematical expression of which is as follows:

[0027]

[0028] FL(P t )=(1-P t ) γ *lg(P t )*P t 2

[0029] In the formula: FL is the loss value, P is the model's estimated probability of the class; y = 1 is the true class; γ is the balance factor that adjusts the difficulty of classifying samples, γ > 0.

[0030] Compared to existing technologies, this invention offers the following advantages: The method of this invention employs a dual-channel semantic segmentation network combining RGB remote sensing imagery and water body indices to address the problem of water body identification in remote sensing multispectral images. Leveraging the characteristics of RGB remote sensing imagery and water body indices, a dual-channel semantic segmentation neural network is constructed, incorporating multi-layer feature fusion and attention modules. This fuses features from the original remote sensing imagery and water body indexes at different scales and feature levels, maximizing the prominence of water body information while suppressing other information such as vegetation, soil, and buildings, thereby improving the accuracy of water body identification. The method of this invention exhibits superior water body identification performance. Attached Figure Description

[0031] Figure 1This is a schematic diagram of the method flow of the present invention;

[0032] Figure 2 This invention relates to a dual-channel semantic segmentation network;

[0033] Figure 3 This is the feature extraction module of the present invention;

[0034] Figure 4 This is the layered feature fusion module of the present invention;

[0035] Figure 5 This is the attention module of the present invention;

[0036] Figure 6 This is the feature output module of the present invention. Detailed Implementation

[0037] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings.

[0038] This invention provides a method for water body identification based on remote sensing multispectral images using a dual-channel segmentation network, the steps of which include:

[0039] Step S1: Data preprocessing. The collected remote sensing images are preprocessed, including radiometric correction, atmospheric correction and geographic correction, to eliminate noise and errors in the images. Water bodies in the remote sensing images are manually labeled at the pixel level to obtain labels and establish a multispectral image water body dataset.

[0040] Step S2: Construct RGB image data. Since the RGB color model corresponds to human-perceived colors and can intuitively represent color information in an image, infrared (Red), green (Green), and blue (Blue) band data from remote sensing multispectral image data are selected and superimposed to form a three-channel RGB image data.

[0041] Step S3: Construct water body index data. Water body indices are indicators used for water body detection and monitoring in remote sensing image analysis. By calculating the ratios or indices between different bands, the spectral characteristics of water bodies can be highlighted, thereby distinguishing them from other ground features. Near-infrared (NIR), red, green, and shortwave infrared (SWIR) bands of remote sensing multispectral image data are selected. The Normalized Difference Vegetation Index (NDVI), Normalized AWE Index (AWElnsh), and Normalized Difference Water Index (NDWI) are calculated respectively, and these three indices are combined into a three-channel water body index dataset. The Normalized Difference Vegetation Index (NDVI) is calculated as follows:

[0042] NDVI = (NIR - Red) / (NIR + Red)

[0043] In the formula: NDVI is the normalized vegetation index, NIR is the reflectance in the near-infrared band, and Red is the reflectance in the infrared band.

[0044] The Normalized Average Water Enhancement Index (AWElnsh) is calculated as follows:

[0045] AWElnsh=4*(Green-SWIR)-(0.25*NIR+2.75*SWIR)

[0046] In the formula: AWEElnsh is the normalized water enhancement index, Green is the reflectance of the green light band, SWIR represents the reflectance of the shortwave infrared band, and NIR is the reflectance of the near-infrared band.

[0047] The Normalized Difference Water Index (NDWI) is calculated as follows:

[0048] NDWI = (NIR - SWIR) / (NIR + SWIR)

[0049] In the formula: NDWI is the normalized differential water index, NIR is the reflectance in the near-infrared band, and SWIR represents the reflectance in the short-wave infrared band.

[0050] Step S4: Construct a dual-channel semantic segmentation network, such as... Figure 2 As shown, the dual-channel semantic segmentation network includes a feature extraction module, a hierarchical feature fusion module, and a feature output module. The RGB image data generated in step S2 and the water index data generated in step S3 are used as the data inputs for the two channels, respectively. They are processed by the feature extraction module and the features of the two input channels are fused by the hierarchical feature fusion module to obtain deep features. Finally, the feature output module outputs the water segmentation results of the multispectral remote sensing image.

[0051] Further construction steps include the feature extraction module in step S4, such as... Figure 3 As shown, in order to fully capture the spatial information of the input data and obtain data features of different scales from different levels, a network consisting of three layers—two convolutional layers with a kernel size of 3×3 and one max pooling layer with a kernel size of 3×3—is used sequentially. The output feature layers S1, S2, and S3 have feature sizes of 1 / 2, 1 / 4, and 1 / 8 of the input image, respectively. The feature layer obtained from the RGB image channels is S1. 11 S 12 S 13 The feature layer obtained from the index image channel is S. 21 S 22 S 23 ;

[0052] Further construction steps include the hierarchical feature fusion module in step S4, such as... Figure 4As shown, each feature layer S extracted by the feature extraction module 11 S 12 S 13 S 21 S 22 S 23 All of them pass through an attention module, which learns weight allocation to enable the network to pay more attention to the features of the water body area, thereby improving the accuracy of water body area identification and improving the identification accuracy from low-resolution feature maps S. 11 and S 21 The fusion process begins, and after upsampling, it merges with S. 12 and S 22 After fusion, and then upsampling, it is combined with S 13 and S 23 Fusion. The attention module, for example... Figure 5 As shown, the feature layer is subjected to max pooling and average pooling. The results of these two pooling are then concatenated and subjected to global average pooling, convolutional layers, and activation functions. The resulting feature is multiplied with the input feature, and the weight is applied to each of the original feature spaces to obtain a new feature of the attention mechanism.

[0053] Further construction steps include the feature output module in step S4, such as... Figure 6 As shown, the feature output module consists of a global average pooling layer, two convolutional layers with a kernel size of 3×3, and a deconvolutional layer with a kernel size of 2×2. The deconvolutional layer is used to upsample the feature map of the last convolutional layer to obtain the same size as the input image. Finally, pixel-by-pixel classification is performed on the upsampled feature map.

[0054] Step S5: Construct the loss function for the dual-channel semantic segmentation network. To effectively address the imbalance between positive and negative samples and the imbalance between easy and difficult classification samples, an improved Focal loss function is used, the mathematical expression of which is as follows:

[0055]

[0056] FL(P t )=(1-P t ) γ *lg(P t )*P t 2

[0057] In the formula: FL is the loss value, P is the model's estimated probability of the class; y = 1 is the true class; γ is the balance factor that adjusts the difficulty of classifying samples, γ > 0.

[0058] Step S6: Train the dual-channel semantic segmentation neural network. Use the RGB image data generated in step S2 and the water index data generated in step S3 as input to the dual-channel semantic segmentation neural network, and use the labels marked in step 1 as output to train the dual-channel semantic segmentation neural network. Use PyTorch as the deep neural network learning framework and NVIDIA GeForce RTX4060 as the GPU for model training. Use the Adam optimizer, set the initial learning rate to 0.001, and adopt the hierarchical descent learning rate method. The training times are 200 epochs (all training data are processed once).

[0059] Step S7: Identify water bodies in multispectral remote sensing satellite images. After processing in steps S2 and S3, the multispectral remote sensing satellite images to be identified are input into the neural network trained in step S6 to obtain the results of water body identification from multispectral remote sensing satellite images.

[0060] The above are preferred embodiments of the present invention. Any changes made to the technical solution of the present invention that do not exceed the scope of the technical solution of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for water body identification based on remote sensing multispectral images using a dual-channel segmentation network, characterized in that, Includes the following steps: Step S1: Data preprocessing. The collected remote sensing images are preprocessed, and the water bodies in the remote sensing images are labeled at the pixel level to obtain labels and establish a multispectral image water body dataset. Step S2: Construct RGB image data by selecting the infrared (Red), green, and blue bands of the multispectral water body data and superimposing them to form a three-channel RGB image data. Step S3: Construct water body index data. Water body indices are indicators used for water body detection and monitoring in remote sensing image analysis. By calculating indices between different bands, the spectral characteristics of water bodies are highlighted, thereby distinguishing them from other ground features. Near-infrared (NIR), infrared (Red), green, and shortwave infrared (SWIR) bands of multispectral image water body data are selected. The Normalized Difference Vegetation Index (NDVI), Normalized Water Enhancement Index (AWElnsh), and Normalized Difference Water Index (NDWI) are calculated respectively, and these three indices are combined into a three-channel water body index dataset. The Normalized Vegetation Index (NDVI) is calculated as follows: In the formula: NDVI is the normalized vegetation index, NIR is the reflectance in the near-infrared band, and Red is the reflectance in the infrared band. The normalized water enhancement index (AWElnsh) is calculated as follows: In the formula: AWEElnsh is the normalized water enhancement index, Green is the reflectance of the green light band, SWIR represents the reflectance of the shortwave infrared band, and NIR is the reflectance of the near-infrared band. The Normalized Difference Water Index (NDWI) is calculated as follows: In the formula: NDWI is the normalized differential water index, NIR is the reflectance in the near-infrared band, and SWIR represents the reflectance in the short-wave infrared band. Step S4: Construct a dual-channel semantic segmentation neural network. The dual-channel semantic segmentation neural network includes a feature extraction module, a hierarchical feature fusion module, and a feature output module. The RGB image data generated in step S2 and the water index data generated in step S3 are used as the data inputs for the two channels, respectively. They are processed by the feature extraction module and the features of the two input channels are fused by the hierarchical feature fusion module to obtain deep features. The feature output module then outputs the water segmentation results of the multispectral image. Step S5: Construct the loss function of the dual-channel semantic segmentation neural network. Use the improved Focal loss function to construct the loss function of the dual-channel semantic segmentation network. Step S6: Train the dual-channel semantic segmentation neural network. Use the RGB image data generated in step S2 and the water index data generated in step S3 as the input to the dual-channel semantic segmentation neural network, and use the labels marked in step S1 as the output to train the dual-channel semantic segmentation neural network. PyTorch was used as the deep neural network learning framework, and NVIDIA GeForce RTX4060 was used as the GPU for model training. The Adam optimizer was used, the initial learning rate was set to 0.001, and a hierarchical descent learning rate method was adopted. The training time was 200 epochs. Step S7: Identify water bodies in multispectral images. After processing in steps S2 and S3, the multispectral images to be identified are input into the dual-channel semantic segmentation neural network trained in step S6 to obtain the results of multispectral image identification of water bodies.

2. The method for water body identification based on dual-channel segmentation network in remote sensing multispectral imagery according to claim 1, characterized in that, The data preprocessing in step S1 involves preprocessing the collected remote sensing images, including radiometric correction, atmospheric correction, and geographic correction, to eliminate noise and errors in the images.

3. The method for water body identification based on dual-channel segmentation network in remote sensing multispectral imagery according to claim 1, characterized in that, In step S4, the feature extraction module, in order to fully capture the spatial information of the input data and obtain data features of different scales from different levels, sequentially uses three networks consisting of two convolutional layers with a kernel size of 3×3 and a max pooling layer with a kernel size of 3×3. The output feature layers S1, S2, and S3 have feature sizes of 1 / 2, 1 / 4, and 1 / 8 of the input image, respectively. The feature layer obtained from the RGB image channels is S1. 11 S 12 S 13 The water index image channel yields a feature layer of S. 21 S 22 S 23 .

4. The method for water body identification based on dual-channel segmentation network in remote sensing multispectral imagery according to claim 3, characterized in that, In step S4, the hierarchical feature fusion module fuses the feature layers S extracted by the feature extraction module. 11 S 12 S 13 S 21 S 22 S 23 All of them pass through an attention module, which learns weight allocation to enable the network to pay more attention to the features of the water body area, thereby improving the accuracy of water body area identification and improving the identification accuracy from low-resolution feature maps S. 11 and S 21 The fusion process begins, and after upsampling, it merges with S. 12 and S 22 After fusion, and then upsampling, it is combined with S 13 and S 23 Integration.

5. The method for water body identification based on dual-channel segmentation network in remote sensing multispectral imagery according to claim 4, characterized in that, In step S4, the feature output module consists of a global average pooling layer, two convolutional layers with a kernel size of 3×3, and a deconvolutional layer with a kernel size of 2×2. The deconvolutional layer is used to upsample the feature map of the last convolutional layer to obtain the same size as the input image. Finally, pixel-by-pixel classification is performed on the upsampled feature map.

6. The method for water body identification based on dual-channel segmentation network in remote sensing multispectral imagery according to claim 1, characterized in that, In step S5, the specific method for constructing the loss function of the dual-channel semantic segmentation neural network is as follows: In order to effectively solve the problems of imbalance between positive and negative samples and imbalance between easy and difficult classification samples, an improved Focal loss function is used, the mathematical expression of which is as follows: In the formula: FL is the loss value, p is the model's estimated probability for the class; y=1 is the true class; To adjust the balance factor between easy and difficult classification samples, .

Images