Mangrove forest vegetation identification method, system, equipment and medium

Abstract

The invention belongs to the technical field of mangrove forest detection and identification, and discloses a mangrove forest vegetation identification method, system and device and a medium, and the method comprises the steps: obtaining an optical image of a to-be-identified region and a corresponding complete polarization decomposition image; performing image fusion on the optical image of the to-be-identified area and the corresponding complete polarization decomposition image to obtain a to-be-identified fusion image; inputting the to-be-identified fusion image into a mangrove forest vegetation identification model for classification prediction to obtain a mangrove forest vegetation identification result; wherein the mangrove forest vegetation identification model comprises an image segmentation module and a sliding overlapping splicing prediction model which are connected in sequence, and the image segmentation module is an AttU-Net network which is constructed based on a lightweight U-Net neural network and is suitable for mangrove forest identification in a high-similarity environment. The method has higher overall performance, better detail capture capability and better robustness against class imbalance in the aspect of mangrove forest vegetation recognition.

Description

Technical Field

[0001] The present invention belongs to the technical field of mangrove detection and identification, and in particular relates to a mangrove vegetation identification method, system, equipment and medium. Background Art

[0002] Traditional methods for extracting mangrove vegetation include manual visual interpretation, index method, and image classification method based on texture and shape features. First, although manual visual interpretation is intuitive and easy to understand, experienced interpreters can achieve high interpretation accuracy with the support of professional knowledge and experience. However, this method is time-consuming and expensive, requiring a lot of human resources and time. At the same time, due to the particularity of the mangrove growth environment, conventional field surveys are difficult to meet the high temporal and spatial resolution detection requirements for mangroves. Secondly, although the index method is simple and easy to implement, for example, by calculating different indexes of remote sensing images (such as NDVI, Normalized Difference Vegetation Index), the preliminary classification of mangrove vegetation areas can be achieved. However, mangrove areas usually have complex vegetation structures and surface features, including mangrove crowns, water bodies, and mud. Since the index method mainly relies on a simple combination of spectral information, the classification method based on the index method is difficult to effectively distinguish these complex landform types.

[0003] Image classification methods based on texture and shape features mainly include machine learning methods and deep learning methods. In the study of mangrove vegetation recognition, commonly used machine learning methods include support vector machine (SVM), random forest (Random Forest) and K-Nearest Neighbors (KNN). First, the SVM method. The principle of SVM is relatively simple, does not require a lot of parameter adjustment, and has good generalization ability and recognition accuracy for small-scale data sets. However, SVM is usually classified based on manually extracted features, which limits its ability to learn complex and abstract features. At the same time, for the modeling of nonlinear relationships, the SVM method often fails to meet the expectations of the research. Secondly, the random forest method can effectively handle nonlinear relationships compared to the SVM method. At the same time, because it is an integrated learning method based on multiple decision trees, the random forest has a relatively small impact on outliers and can reduce the risk of overfitting to a certain extent. However, random forests usually learn features at a shallow level and it is difficult to automatically learn higher-level and more abstract feature representations. At the same time, it also has the same problem as the SVM method, because random forest is an integrated method based on decision trees, so the modeling of complex nonlinear relationships in images is relatively limited. Finally, the K nearest neighbor algorithm, KNN is an intuitive and easy-to-understand algorithm that does not involve complex model structure and parameter adjustment, and can effectively reduce the model building time. However, compared with the SVM and RF methods, its computational overhead is greater, because when making predictions, KNN needs to calculate the distance between the test sample and all training samples. And KNN is very sensitive to outliers, because its prediction results are greatly affected by the nearest neighbor samples, and a single outlier may have a greater impact on the prediction results. In summary, although the mangrove image classification method based on machine learning can be better applied in complex environments compared to manual visual interpretation and index methods, it also has high spatiotemporal resolution detection and big data processing capabilities. However, it still requires manual feature engineering and is limited to learning shallow features.

[0004] In order to overcome these limitations and improve the accuracy and automation of mangrove identification, more and more studies have begun to introduce deep learning technology in the field of mangrove vegetation identification. At the method level, these studies use convolutional neural networks (CNNs) to accurately extract deep features. A large number of existing technologies use improved U-Net networks and classify mangrove vegetation based on GF-2 cloudless and unobstructed optical images. All of them choose cloudless and unobstructed optical images. Therefore, at the image level, the main source of these research data is high-resolution remote sensing images, which mainly rely on optical information. Although optical images have high spatial resolution and rich color information, they are easily restricted by weather and lighting conditions. At the same time, because optical images cannot penetrate clouds, vegetation cover or underground structures, the information obtained using optical images is not complete under complex landform conditions such as mangroves. Summary of the invention

[0005] The purpose of the present invention is to provide a mangrove vegetation identification method, system, device and medium to solve the problems existing in the above-mentioned prior art.

[0006] To achieve the above object, the present invention provides a mangrove vegetation identification method, comprising:

[0007] Obtaining an optical image of the area to be identified and a corresponding fully polarized decomposed image;

[0008] Performing image fusion on the optical image of the area to be identified and the corresponding fully polarized decomposed image to obtain a fused image to be identified;

[0009] The fused image to be identified is input into a mangrove vegetation recognition model for classification prediction to obtain a mangrove vegetation recognition result; wherein the mangrove vegetation recognition model includes an image segmentation module and a sliding overlapping splicing prediction model connected in sequence, and the image segmentation module is based on a lightweight U-Net neural network architecture, which is a AttU-Net network suitable for mangrove identification in a high-similarity environment by combining a channel attention mechanism and an improved cross-entropy loss function for mangrove characteristics.

[0010] Optionally, the process of acquiring the optical image of the area to be identified specifically includes:

[0011] Obtain a panchromatic image and a corresponding multispectral image of the area to be identified, perform radiation correction, atmospheric correction, and orthorectification on the panchromatic image and the corresponding multispectral image, respectively, to obtain corrected panchromatic image and multispectral image, and fuse the corrected panchromatic image and multispectral image to obtain an optical image of the area to be identified.

[0012] Optionally, the process of acquiring the full polarization decomposition image specifically includes:

[0013] An L1A-level single-view complex image of the area to be identified is acquired, and the L1A-level single-view complex image is sequentially subjected to Pauli vector transformation, polarization coherence matrix transformation, full polarization filtering and reflection symmetric decomposition processing to obtain the full polarization decomposed image.

[0014] Optionally, the optical image of the area to be identified and the corresponding fully polarized decomposed image are fused, and the specific calculation formula is:

[0015] Based on the bilinear interpolation method, the optical image of the area to be identified and the corresponding full polarization decomposition image are resized to obtain an optical image and a corresponding full polarization decomposition image with the same size and matching texture information; the resized optical image and the corresponding full polarization decomposition image are weighted fused to obtain the fused image to be identified.

[0016] Optionally, the training process of the mangrove vegetation recognition model specifically includes:

[0017] Acquire a training image, wherein the training image includes a fused training image to be identified and a corresponding mangrove vegetation identification result;

[0018] An initial mangrove vegetation recognition model is constructed, the training image is input into the initial mangrove vegetation recognition model for classification prediction, and training is performed with the goal of minimizing the loss between the initial training result after classification prediction and the mangrove vegetation recognition result corresponding to the fused training image to be identified, so as to obtain the mangrove vegetation recognition model.

[0019] Optionally, the processing process of the mangrove vegetation recognition model includes:

[0020] The fused image to be identified is input into the image segmentation module for image segmentation to obtain the image to be predicted, the image to be predicted is processed through a sliding window, the image to be predicted is divided into multiple small images, classification prediction is performed on each small image, the classification prediction results are spliced, the overlapping part is selected and the maximum value is taken as the final prediction value to obtain the mangrove vegetation recognition result.

[0021] Optionally, the network structure of the image segmentation module is provided with 5 layers from top to bottom, and the first layer, the second layer, the third layer and the fourth layer are 4 groups of encoder parts and decoder parts of symmetrical structures;

[0022] The second, third, fourth and fifth layers all introduce an attention mechanism. After the convolution operation, each layer will be processed by batch normalization and calculated using the ReLU activation function; the convolution layers of the fourth and fifth layers are also connected with a Dropout layer.

[0023] A mangrove vegetation identification system, comprising:

[0024] A data acquisition module, used to obtain an optical image of the area to be identified and a corresponding fully polarized decomposed image;

[0025] An image fusion module, used for fusing the optical image of the area to be identified and the corresponding fully polarized decomposed image to obtain a fused image to be identified;

[0026] An image recognition module is used to input the fused image to be identified into a mangrove vegetation recognition model for classification prediction to obtain a mangrove vegetation recognition result; wherein the mangrove vegetation recognition model includes an image segmentation module and a sliding overlapping splicing prediction model connected in sequence, and the image segmentation module is based on a lightweight U-Net neural network architecture, and is constructed by combining a channel attention mechanism and an improved cross entropy loss function for mangrove characteristics to construct an AttU-Net network suitable for mangrove recognition in a high similarity environment.

[0027] An electronic device comprises a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program so that the electronic device executes the mangrove vegetation identification method.

[0028] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for identifying mangrove vegetation is implemented.

[0029] The technical effects of the present invention are:

[0030] The present invention provides a method for pixel-level weighted fusion of SAR and optical images. The fused image strengthens the target features, making the detection of mangroves more comprehensive and accurate. Aiming at the problem of high similarity between mangroves and other woods, the present invention is based on a lightweight U-Net convolutional neural network, and adds an attention mechanism in its feature extraction stage, so that the model pays more attention to the mangrove vegetation area in the image, and can more effectively capture the complex structure and texture features of the mangroves in the image. In the training process of the entire model, an improved ignored edge cross entropy function is used as a loss function, so that the model can process each category more balanced, and improve the recognition accuracy of mangrove vegetation. The present invention has higher overall performance, better detail capture ability and better robustness against category imbalance in identifying mangrove vegetation. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0032] Figure 1 Schematic diagram of SAR image and optical image of the study area in the embodiment of the present invention;

[0033] Figure 2 is a weighted fusion result diagram of the SAR image and the optical image in an embodiment of the present invention;

[0034] Figure 3 Schematic diagram of the AttU-Net model structure in an embodiment of the present invention;

[0035] Figure 4 Schematic diagram of the network structure of SE-Net in an embodiment of the present invention;

[0036] Figure 5 It is an experimental comparison diagram of a comparative experiment in an embodiment of the present invention;

[0037] Figure 6 is a comparison diagram of the prediction results of the ablation experiment in the embodiment of the present invention;

[0038] Figure 7 is a line graph of the accuracy and loss value of the training set and the test set based on the fused image in the embodiment of the present invention, wherein: Figure 7 (a) is the accuracy line chart of the training set and the validation set; Figure 7 (b) is a line graph of the loss values ​​of the training set and the validation set;

[0039] Figure 8 This is a comparison diagram of mangrove vegetation prediction results in the test area in an embodiment of the present invention;

[0040] Fig. 9 This is a flow chart of mangrove vegetation identification in an embodiment of the present invention. DETAILED DESCRIPTION

[0041] Various exemplary embodiments of the present invention will now be described in detail. This detailed description should not be considered as limiting the present invention, but should be understood as a more detailed description of certain aspects, features, and embodiments of the present invention.

[0042] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0043] Embodiment 1

[0044] like Figure 1 - Fig. 9As shown, a mangrove vegetation recognition method is provided in this embodiment, including: obtaining an optical image of an area to be recognized and a corresponding fully polarized decomposition image; performing image fusion on the optical image of the area to be recognized and the corresponding fully polarized decomposition image to obtain a fused image to be recognized; inputting the fused image to be recognized into a mangrove vegetation recognition model for classification prediction to obtain a mangrove vegetation recognition result; wherein the mangrove vegetation recognition model includes an image segmentation module and a sliding overlapping splicing prediction model connected in sequence, and the image segmentation module is constructed based on a U-Net network.

[0045] The fused image to be identified is input into the image segmentation module for image segmentation. After obtaining the image to be predicted, the image to be predicted is processed through a sliding window, and the image to be predicted is divided into multiple small images. Classification prediction is performed on each small image, and the classification prediction results are spliced. The maximum value of the overlapping part is selected as the final prediction value to obtain the mangrove vegetation recognition result.

[0046] As an indispensable component of the coastline, mangroves play a key role in maintaining biodiversity and mitigating climate change. Therefore, improving the accuracy of mangrove information recognition is very important for mangrove ecological protection. In view of the limited morphological information of SAR images, the large interference from noise, and the susceptibility of optical images to weather and lighting conditions, this embodiment proposes a method for pixel-level weighted fusion of SAR and optical images. The fused image strengthens the target features, making mangrove detection more comprehensive and accurate. In view of the high similarity between mangroves and other trees, this embodiment is based on the unet convolutional neural network, and adds an attention mechanism in its feature extraction stage, so that the model pays more attention to the mangrove vegetation area in the image. In order to accelerate convergence and normalize the input, a BatchNormalization (BN) layer and a Dropout layer are added after each layer of convolution. Since mangroves are a minority category in the image, this embodiment introduces an improved cross entropy loss function to improve the model's recognition ability of mangroves. Based on the fused image, an AttU-Net model suitable for mangrove recognition in a high similarity environment is constructed. Through comparative experiments, the overall accuracy of the improved U-Net model trained by the fused image in identifying the predicted area has been significantly improved. Based on the fused image, the recognition results of the AttU-Net model proposed in this embodiment are compared with its benchmark model U-Net and Dense-Net, Res-Net, and Seg-Net methods. The AttU-Net model can more effectively capture the complex structure and texture features of mangroves in the image. The average OA, F1-Score, and Kappa coefficients in the four test areas are 94.406%, 90.006%, and 84.045%, respectively, which can provide certain technical support for the detection and protection of mangrove ecosystems.

[0047] This embodiment proposes a method for fusing SAR and optical images at the pixel level. The fused image can not only retain the high resolution and color information of the optical image, but also make full use of the penetration ability of the SAR image to provide more detailed and comprehensive ground feature features, thereby enhancing the texture features and shape of the image and improving the recognition accuracy of mangroves.

[0048] In terms of research methods, this embodiment selects the U-Net neural network as the benchmark network. First, mangrove vegetation recognition is easily restricted by data samples, and the U-Net network performs well in few-sample learning. It can achieve good performance in training by using a small number of labeled samples. At the same time, the unique upsampling structure of the U-Net network enables the network to retain more spatial information, which is particularly effective for mangrove image segmentation tasks. In the recognition of mangrove vegetation, retaining spatial information is very important for capturing the details and edges of vegetation, and the upsampling structure of U-Net can accurately extract features while maintaining a certain spatial resolution.

[0049] For the mangrove vegetation recognition task, although the U-Net network has superior performance, in order to further improve the generalization of the model, reduce overfitting, enhance the focus on important features and optimize the loss function, this embodiment introduces the Dropout layer, BatchNormalization (BN) layer, attention mechanism and improved cross entropy loss function (CLoss). First, since mangrove vegetation occupies a small proportion in the image, the mangrove category is relatively small in the training data. This imbalance often causes the model to pay too much attention to other categories and ignore mangroves, increasing the risk of overfitting of background information. Therefore, this embodiment introduces the Dropout layer, which forces the model not to rely too much on specific neurons by randomly inactivating some neurons during the training process, thereby improving generalization and reducing the impact of overfitting. Secondly, in mangrove vegetation recognition, the distribution of vegetation will change with the changes in different terrains and environments. It is extremely easy to have gradient disappearance or gradient explosion during the training process, which affects the training stability of the model. Therefore, this embodiment introduces the BN layer to standardize the input of each layer, alleviate the gradient problem, accelerate convergence, and improve the training efficiency and stability of the model. Thirdly, in the recognition of mangrove vegetation, the accuracy of paying attention to the mangrove area is crucial to improving the model performance. Therefore, the attention mechanism is introduced to enable the network to focus more on the mangrove vegetation area and improve the performance of the model in the target area. Finally, consistent with the reason for adding the Dropout layer, mangrove vegetation is a minority category in the overall image, and the traditional cross entropy loss function will cause the model to over-learn the main categories when dealing with unbalanced category distribution. Therefore, this embodiment introduces an improved CLoss to weight the losses of different categories so that the model pays more attention to the recognition of mangrove vegetation. By adjusting the weights, the model can handle each category more balanced and improve the recognition accuracy of mangrove vegetation.

[0050] Thus, an AttU-Net model suitable for mangrove vegetation identification under fusion image conditions is constructed. In addition, in order to further improve the accuracy of mangrove vegetation extraction, this embodiment introduces a sliding overlapping splicing method to predict the results, which is mainly used to identify the problem that the result identification image has splicing traces and insufficient edge information. In order to better improve the accuracy of model identification of mangroves and provide certain technical support for mangrove ecological protection and management. Specifically, the sliding overlapping splicing method effectively reduces the amount of calculation of a single large image in the model and improves the calculation efficiency by dividing the large image into multiple small images for prediction. At the same time, since remote sensing images usually have huge sizes, the sliding overlapping window technology can be used to cut the large image into multiple small images, thereby avoiding the problem of insufficient memory and ensuring the feasibility of the model.

[0051] During the sliding window process, each small image will be predicted multiple times, and then these prediction results will be summarized and averaged to obtain the final prediction result of the small image. The model makes multiple inferences on image blocks at different positions, comprehensively considers multiple prediction results, improves the accuracy of prediction, and reduces the impact of random noise.

[0052] This study sets an overlap ratio to control the degree of overlap between windows. The window size is W, the overlap ratio is r, then the step size of each window can be expressed as:

[0053] Step=W(1-r) (1)

[0054] Among them, the overlap ratio is usually selected as half or one-third of the window size to improve computational efficiency while maintaining information continuity. The sliding overlapping window technology introduces flexibility and accuracy to the prediction process of the model, enabling it to better cope with the processing needs of large-scale urban water remote sensing images. Study area: Hainan Dongzhaigang National Nature Reserve is located in the northeast of Hainan Province, at the junction of Haikou City and Wenchang City. Its geographical coordinates are 110°32′~110°37′ east longitude and 19°51′~20°1′ north latitude. It is a wetland type nature reserve. The total area of ​​the reserve is 3337.6 hectares, the core area is 1635 hectares, the buffer area is 1167.1 hectares, and the experimental area is 535.5 hectares, of which the mangrove area is 1771 hectares and the tidal flat area is 1759.4 hectares. Dongzhaigang Nature Reserve has a tropical monsoon climate, with an average annual temperature of 23.8℃ (average 28.4℃ in July and 17.1℃ in January), an annual rainfall of 1,700 mm, and many typhoons in the rainy season, bringing strong winds and heavy rains. The highest sea temperature is 32.6℃, the lowest is 14.6℃, and the average is 24.5℃.

[0055] Dongzhaigang Mangrove Reserve has the characteristics of a wide variety of tree species, a large area of ​​mangroves, and a good ecological environment. The diversity of mangroves in the area can provide a wider range of samples for research, which helps to verify the applicability of the model on different mangrove vegetation types. At the same time, the large area of ​​mangroves in the area can provide enough space and data to more comprehensively understand the structure, function and dynamic changes of mangroves.

[0056] Data source: This embodiment uses the GF-3 satellite SAR image data and GF-6 satellite optical image data of Hainan Province. An experiment on high-precision extraction of mangrove vegetation was conducted with the mangrove nature reserve at the junction of Haikou City and Wenchang City in Hainan Province as the main image. Gaofen-3 is a remote sensing satellite of China's Gaofen special project. It is a 1-meter resolution radar remote sensing satellite and China's first C-band multi-polarization synthetic aperture radar (SAR) imaging satellite with a resolution of 1 meter. At the same time, Gaofen-3 has 12 imaging modes, covering traditional strip imaging mode and scanning imaging mode, as well as wave imaging mode and global observation imaging mode for ocean applications. It is the synthetic aperture radar satellite with the most imaging modes in the world. The full polarization imaging mode and capabilities of Gaofen-3 SAR images are shown in Table 1. Gaofen-6 is a low-orbit optical remote sensing satellite and China's first high-resolution satellite for precision agricultural observation. It has the characteristics of combining high resolution and wide coverage. The Gaofen-6 satellite is equipped with a 2-meter panchromatic / 8-meter multispectral high-resolution camera and a 16-meter multispectral medium-resolution wide-band camera. The 2-meter panchromatic / 8-meter multispectral camera has an observation width of 90 kilometers, and the 16-meter multispectral camera has an observation width of 800 kilometers. The satellite payload of the Gaofen-6 is shown in Table 2.

[0057] Table 1 Full polarization imaging modes and capabilities of GF-3 SAR images

[0058]

[0059] Table 2 Satellite payload of GF-6

[0060]

[0061] Data preprocessing: In this embodiment, the panchromatic image and multispectral image of the GF-6 mangrove study area are subjected to radiation correction, atmospheric correction, and orthorectification. After that, the corrected panchromatic image and multispectral image are fused by software to obtain an optical image with higher temporal and spatial resolution. The fully polarized SAR incoherent polarization decomposition product is input with the L1A-level single-view complex image (SLC) standard product of the three fully polarized observation modes (fully polarized strip 1, fully polarized strip 2, and wave mode) of the Gaofen-3 satellite and the 1-meter C-SAR satellite, and is obtained through Pauli vector transformation, polarization coherence matrix transformation, full polarization filtering, and reflection symmetric decomposition processing steps. Afterwards, the optical image of GF-6 is used as the reference image, and the result of the fully polarized SAR target decomposition is used as the image to be registered for geographic registration. Finally, the preprocessed optical image and the polarized SAR decomposition result are cropped according to the same area of ​​interest to obtain the optical and fully polarized decomposition images of the same area.

[0062] In this embodiment, the preprocessed optical image and SAR image are weightedly fused at the pixel level through code. The fused image can provide richer and more comprehensive surface information. This embodiment generates a series of fused images, each of which represents a different proportion of the SAR image and the optical image. Corresponding to the fused image in the figure, let the weights of the SAR and optical images be a and b, respectively, satisfying a+b=1. Here, a represents the weight of the SAR image, and b represents the weight of the optical image. By adjusting the proportion of the SAR image and the optical image in the fused image, the study is divided into 11 fused images with different proportions, such as Figure 2 shown.

[0063] The designed weighted fusion module involves two steps. First, due to the different pixel sizes, the two images of the same region of interest will have texture and size mismatch problems after cropping. Therefore, to address this problem, the two images are resized. The bilinear interpolation method is used to calculate the value of the new pixel according to the size and pixel layout of the optical image, so that the adjusted SAR image and the optical image are completely matched in size, thereby aligning the texture information of the objects in the two images. The formula is as follows:

[0064]

[0065] Where dst(x,y) represents the pixel value at coordinate (x,y) in the optical image. src(i,j) represents one of the four pixels closest to (x,y) in the SAR image. α and β represent the fractional part between the coordinate (x,y) and the nearest integer coordinate i,j, respectively.

[0066] Finally, two images with the same size and matching texture information are weighted fused using the following formula:

[0067] dst(x,y)=src1(x,y)·α+src2(x,y)·β+γ (3)

[0068] Among them, dst(x,y) represents the pixel value at coordinate (x,y) in the output image, src1(x,y) and src2(x,y) represent the pixel values ​​at the corresponding positions in the optical image and SAR image respectively. α and β represent the weights of the two images, i.e. a and b respectively. γ is used to adjust the brightness of the fused image.

[0069] AttU-Net neural network: The network AttU-Net neural network designed in this embodiment is based on the U-Net network framework. The U-Net network is a convolutional neural network for image segmentation, which has two parts: an encoder and a decoder. The encoder extracts high-level features of the image through downsampling operations, and the decoder restores the resolution through upsampling operations. This structure can retain high-resolution information and can more effectively adapt to the complex structure and texture of mangrove vegetation.

[0070] However, when conducting research on mangrove vegetation identification, problems that need to be dealt with include the relative scarcity of training data and the more complex background environment. In this context, the U-Net network itself has some potential problems, which are mainly reflected in the following three aspects. First, the problem of overfitting. As a deep neural network, U-Net is often prone to overfitting when there is insufficient training data. Secondly, as the network deepens, the training becomes unstable, resulting in difficulties in convergence, especially when dealing with complex and diverse mangrove vegetation. Finally, due to the excessive similarity between mangroves and other trees, for mangrove images under different environmental conditions, U-Net's sensitivity to input data varies greatly, resulting in poor performance in mangrove scenes with large changes.

[0071] In response to these three problems, this embodiment improves the recognition performance of mangrove vegetation by adding a Dropout layer, a Batch Normalization (BN) layer, and an attention mechanism. First, the Dropout layer randomly discards neurons during the training process to reduce the dependency between neurons, thereby reducing the risk of overfitting and improving the generalization performance of the model. Secondly, the BN layer. Mangrove vegetation will have some differences under different conditions such as light and humidity. The introduction of the BN layer helps to normalize the activation values ​​of the intermediate layer, improve the robustness of the network, and make it more adaptable to the vegetation characteristics of different mangrove environments. Finally, the attention mechanism. Mangrove vegetation usually has complex structures and changes. The introduction of the attention mechanism can make the network more targeted and focus on areas that are more important for mangrove vegetation recognition, thereby improving the network's perception of vegetation structure and texture, and improving the recognition accuracy of mangrove vegetation.

[0072] The AttU-Net model structure is as follows Figure 3As shown in (where S is the transition layer processed by the attention mechanism module, and D is the transition layer of the Dropout operation). The specific processing steps are as follows: First, downsample the first layer, use the 2D convolution layer and batch normalization layer to process the input data, then perform the maximum pooling operation and apply the attention mechanism module (Squeeze-and-Excitation module) to perform channel attention mechanism processing. In the second downsampled layer, define a new 2D convolution layer, after two convolutions and batch normalization processing, perform the maximum pooling operation and Squeeze-and-Excitation module processing. Similarly, in the third downsampled layer, the output of the previous layer is convoluted, normalized, pooled, and processed by the Squeeze-and-Excitation module. Repeat the similar process for the fourth downsampled layer, but add the Dropout layer after the convolution to prevent overfitting. The fifth downsampled layer is similar to the fourth layer, and the convolution, normalization, and Dropout operations continue. The first upsampling layer uses convolution and upsampling functions to enlarge the image, and concatenates the result with the output of the fourth downsampling layer, followed by convolution and normalization. The second upsampling layer continues to upsample, concatenates with the output of the third downsampling layer, and performs two convolutions. The third upsampling layer continues to upsample, concatenates with the output of the second downsampling layer, and performs two convolutions. The fourth upsampling layer continues to upsample, concatenates with the output of the first downsampling layer, and performs two convolutions, and finally performs another 2D convolution. The fifth upsampling layer performs the last convolution and outputs the result of the filter. The activation function selects softmax to indicate the probability that each pixel belongs to each category. Then, use the Model function to define the Keras model, configure the optimizer, loss function, and evaluation criteria of the model, use the Adam optimizer, the loss function is the improved classification cross entropy function CLoss, and the evaluation criterion is the accuracy. Finally, return the model.

[0073] In addition, because mangrove vegetation is a minority category in the overall image, the traditional cross entropy loss function will cause the model to over-learn the main category when dealing with unbalanced category distribution. At the same time, because mangrove vegetation usually exists in marginal areas such as the seaside and the shore, these areas often contain mangroves and other types of landforms at the same time. Since edge pixels are between two or more categories, their true labels have great uncertainty. Therefore, in order to make the model pay more attention to the recognition of mangrove vegetation and prevent errors at the edges from being transmitted to the entire network through back propagation, affecting the convergence of the model. The AttU-Net network proposed in this embodiment adopts an improved edge-ignoring cross entropy function as the loss function, and improves it based on the classification cross entropy function (CELoss). The parameter r is added to the denominator to adjust the size of the prediction area, and the weight ω is added to the numerator. jBy adjusting the weights, the model can process each category more balanced and improve the recognition accuracy of mangrove vegetation. The final improved cross entropy function ignoring the edge is recorded as CLoss. The formula is as follows:

[0074]

[0075] Among them, N is the number of samples, C is the number of categories, and y ij Indicates whether sample i belongs to category j, p ij is the probability that sample i is predicted as category j, and r is the ratio of the number of pixels in the selected area to the number of pixels in the entire image, thereby achieving the purpose of ignoring edge information. j is the weight of category j, which is used to adjust the importance of category j.

[0076] Dropout layer: The main function of Dropout is to reduce overfitting and increase the generalization ability of the model. By randomly "closing" some neurons, the network is forced to learn more robust features. Therefore, Dropout is a simple and effective regularization method that helps to improve the generalization performance of the model. In mangrove vegetation recognition, due to the presence of noise in the data and complex environmental changes, Dropout can effectively prevent the model from overfitting the training data and improve the adaptability of the model to different mangrove scenarios.

[0077] During training, Dropout sets the output of a neuron to zero with probability p by randomly "turning off" the neuron in each training iteration. The formula for Dropout can be expressed as:

[0078]

[0079] Where x is the input, ⊙ represents element-wise multiplication, and mask is a binary mask of the same size as the input, with each element set to zero with probability p.

[0080] BatchNormalization (BN) layer: The BN layer normalizes each feature so that its mean is close to 0 and its variance is close to 1. By introducing learnable scaling and translation parameters, the model can learn to adapt to features of different distributions. The BN layer can help the network better adapt to different input distributions. In this embodiment, the BN layer is applied before the activation function, which helps to prevent gradient explosion or gradient vanishing problems, reduce the training time of the network, and improve the generalization ability of the model under limited samples.

[0081] The formula for Batch Normalization can be expressed as:

[0082]

[0083] Where x is the input feature, μ and σ are the mean and standard deviation of the input feature, respectively, and γ and β are learnable scaling and translation parameters.

[0084] SE-Net (Squeeze-and-Excitation Network) attention mechanism: SE-Net is a deep neural network based on the attention mechanism, which aims to improve the model's attention to important features in the input. Its core idea is to emphasize the importance of each channel in the network by learning adaptive weights, thereby improving the model's representation ability.

[0085] It mainly includes two operations, namely the Squeeze operation and the Excitation operation. The Squeeze stage is responsible for global pooling, compressing the information of each channel of the input feature map. Global average pooling is used to obtain the global information of each channel. The Excitation stage introduces two fully connected layers (FC layers) to learn the weights between channels. These two FC layers are used for dimensionality reduction (reducing the number of channels) and dimensionality increase (restoring the number of channels), respectively, and then generate weights through the Sigmoid function. These weights are applied to the input feature map to obtain a weighted feature map. The network structure of SE-Net is as follows Figure 4 Shown

[0086] Among them, F tr It is a traditional convolutional feature extraction structure, with X and U as its input and output. H'×W'×C' , U∈R H×W×C . That is, the input feature map is U∈R H×W×C , where H and W are the height and width of the feature map, respectively, and C is the number of channels. The squeeze stage obtains the global descriptor z∈R by global average pooling. C , representing the global information of each channel. The Excitation stage introduces two FC layers to map z to a weight vector s∈R C , where s i Represents the weight of the i-th channel. Finally, the weighted feature map Y∈R is obtained by multiplying the weight vector s by the original input feature map U element by element. H×W×C , the formula is as follows:

[0087] Y i =s i ×X i (7)

[0088] Channel weights i The calculation formula is:

[0089] s i =σ(W2δ(W1z i )) (8)

[0090] Among them, δ represents the ReLU activation function, σ represents the Sigmoid activation function, and W1 and W2 are the weight matrices of the FC layer.

[0091] Compared with other attention mechanisms, SE-Net focuses more on adjusting channel weights, emphasizing the importance of each channel, and is suitable for tasks that pay more attention to channel information. However, mangrove vegetation recognition usually requires emphasizing the information of specific channels such as the color and shape of vegetation, which often have obvious differences in different channels. Therefore, SE-Net can better capture the information of these important channels by adjusting channel weights.

[0092] The specific steps of the method in this embodiment are as follows:

[0093] (1) Data processing: Based on the fused image, this embodiment uses the visual interpretation method combined with the vector data file of field measurement for manual annotation, so as to obtain real and accurate sample labels. By sliding cropping, a sample data set of 256 pixels × 256 pixels is generated. Then, the obtained sample data set is divided into 3000 training sets and 1000 validation sets for data enhancement, for a total of 4000 sample data sets. Among them, the image map and label Figure 1 One to one correspondence.

[0094] (2) Constructing the AttU-Net model: In order to improve the recognition accuracy of the model, make the model converge faster, and reduce the risk of overfitting of background information, this embodiment improves the U-Net network by adding a Dropout layer, a BN layer, and an attention mechanism to improve the recognition performance of mangrove vegetation.

[0095] (3) Training the AttU-Net model: This embodiment trains the AttU-Net model based on a sample set of 11 fused images. Under the same parameter conditions, the fused image with the best accuracy evaluation is selected as the main research area. Then, by adjusting the parameters, the optimal mangrove vegetation recognition model in the research area is obtained.

[0096] (4) Sliding splicing prediction: This embodiment introduces a sliding overlapping splicing method to construct a prediction model. Its purpose is to effectively eliminate the problem of splicing traces and enrich the edge information of the predicted image. The test set is input into the prediction model to obtain the prediction result map of the network identifying mangroves.

[0097] (5) Accuracy evaluation: This embodiment uses three evaluation factors to evaluate the classification results of mangrove vegetation, namely, F1-score, overall accuracy (OA) and kappa coefficient.

[0098] Experimental accuracy evaluation: The classification problem involved in this embodiment is a binary classification problem, that is, dividing the image into mangrove vegetation areas and non-mangrove vegetation areas. In the binary classification problem, the confusion matrix is ​​a table used to evaluate the performance of the classification model, which classifies the relationship between the model's prediction results and the actual categories in detail. The four main elements of the confusion matrix are true positive examples (True Positive, TP), true negative examples (True Negative, TN), false positive examples (False Positive, FP) and false negative examples (False Negative, FN). The layout of the confusion matrix is ​​shown in Table 3:

[0099] Table 3 Layout of the confusion matrix

[0100]

[0101] Among them, the True Positive (TP) is when the model correctly identifies mangrove vegetation as vegetation. The True Negative (TN) is when the model correctly identifies non-vegetation areas as non-vegetation. The False Positive (FP) is when the model incorrectly identifies non-vegetation areas as vegetation. The False Negative (FN) is when the model incorrectly identifies mangrove vegetation as non-vegetation.

[0102] Based on the confusion matrix, this embodiment uses three evaluation factors to evaluate the classification results of mangrove vegetation. They are F1-score, overall accuracy (OA) and kappa coefficient. Among them, F1-Score is an indicator that comprehensively considers the precision (Precision) and recall (Recall) of the model. It provides a single metric by balancing the performance of the model on positive and negative examples. The range of F1-Score is between 0 and 1. The closer to 1, the better the balance between the precision and recall of the model. OA is a simple and intuitive evaluation indicator, which indicates the proportion of samples correctly classified by the model in all categories to the total number of samples. The Kappa coefficient is a statistic used to measure the performance of the classification model. The range of the Kappa coefficient is between -1 and 1. The closer to 1, the better the performance of the model. Unlike OA, the Kappa coefficient is more robust to unbalanced categories and random guessing. The formulas for F1-score, overall accuracy (OA) and kappa coefficient are as follows:

[0103]

[0104] Among them, p0 and p e The expression is as follows:

[0105]

[0106] Comparative experiment based on fused image: In order to verify the effectiveness of the fused image proposed in this embodiment for mangrove vegetation recognition, this embodiment designed the following comparative experiment. The experiment was conducted based on 11 fused images with different weighting ratios, and the weighting ratio with the best performance in the mangrove vegetation recognition task was selected. The prediction area of ​​the comparative experiment selected a densely distributed mangrove vegetation area of ​​500x500 pixels. The accuracy evaluation indicators used were F1-Score, OA and Kappa coefficients. The specific accuracy evaluation results are shown in Table 4.

[0107] Table 4 Accuracy evaluation results of comparative experiments based on fusion images

[0108]

[0109] The experimental comparison diagram of the comparative experiment is as follows Figure 5 As shown, it includes the predicted images of the AttU-Net model trained based on fused images of different scales, the test images of the same area of ​​fused images of different scales, and the true label images of the area (the white area in the image indicates the area that the model considers to be mangrove vegetation, and the black area indicates the area that the model considers to be non-mangrove vegetation).

[0110] From the experimental comparison chart, when the proportion of optical images is high, the experimental comparison chart shows that the recognition error areas are relatively concentrated. This shows that when the model processes optical images, it is more susceptible to local details, and it is easy to confuse objects of similar colors or textures, resulting in the aggregation of erroneous areas. On the contrary, when the proportion of SAR images is high, the experimental comparison chart shows that the erroneous areas are scattered and dotted. Because SAR images have a stronger ability to reveal the structure of objects, it is easier for the model to produce obvious boundaries between objects during the segmentation process, but it is also easy to produce point-like misclassification. Combined with the accuracy evaluation results of the comparative experiment, this embodiment selects the fusion ratio image with the best accuracy evaluation and the best visual effect, that is, the fusion image with a:b=2:8 is selected as the main research object. Obviously, by introducing information from a smaller proportion of SAR images, the model may better capture the boundaries and structures of objects, avoiding the problem of excessive concentration of recognition error areas in optical images.

[0111] Ablation experiment: Through comparative experiments, this embodiment selects the fusion image with a:b=2:8 as the research image for the ablation experiment in this section. In order to better verify the effectiveness of the three modules introduced in this embodiment, ablation experiments were performed on the SE-Net layer, Dropout layer, and BN layer respectively. The prediction area of ​​the experiment selected a 500 pixel x 500 pixel area with dense distribution of mangrove vegetation. The accuracy evaluation indicators used F1-Score, OA, and Kappa coefficients. The original U-Net network is the benchmark model of the experiment. The accuracy evaluation results of the experiment are shown in Table 5. If "√" is marked in the table, it proves that the model has added the module. If "√" is not marked, it proves that the model has not added the module.

[0112] Table 5 Accuracy evaluation results of the ablation experiment test area

[0113]

[0114]

[0115] The comparison of the prediction results of the ablation experiment is shown in the figure below. Figure 6 As shown, it includes the test image, the real label image of the mangrove vegetation in the image, and the predicted image.

[0116] First of all, according to the accuracy evaluation results of the ablation experiment test area, it is not difficult to see that the OA, F1-score and Kappa coefficient of the model after adding the attention mechanism or the Dropout layer alone have been significantly improved compared to the baseline model. This proves that by introducing the attention mechanism, the model pays more attention to the detailed information of the texture, structure and specific areas, which helps to improve the recognition of mangrove vegetation. And by randomly discarding some neurons with a certain probability during the training process, the problem of overfitting of background information can be effectively reduced, and the generalization ability of the model can be increased. However, after adding the BN layer alone, the OA, F1-score and Kappa coefficient decreased significantly compared to the baseline model. At the same time, combined with Figure 6 As shown in the No. 4 sub-figure in the figure, adding the BN layer alone will introduce some noise in the mangrove vegetation recognition task, causing the model to overfit the background information in the training set, thereby identifying some mangrove vegetation areas as background, which reduces the generalization ability of the model. Figure 6 In the No. 4 sub-image in the figure, there is a segmentation error in the upper left corner that does not appear in any other prediction results. This is because the U-Net model itself is not complex enough. After adding the BN layer, the model's expressive power is insufficient, which reduces the performance of the model.

[0117] In the models with both modules added at the same time, it can be seen that compared with the models with only the attention mechanism or the Dropout layer added (No.2, No.3), the model accuracy has decreased (No.5, No.6, No.7). Compared with the model with only the BN layer added (No.4), the model accuracy has improved (No.6, No.7). Combining the images, it can be seen that adding the attention mechanism and the Dropout layer at the same time introduces the complexity of the model, making the feature distribution more dynamic and elusive. This ultimately leads to the problem of overfitting of the background information in the model, such as Figure 6 In the No.5 image, the area of ​​the lower right corner where the recognition error is detected is significantly larger. On the basis of the model that has already added the BN layer, adding the attention mechanism and the Dropout layer has improved (No.6, No.7). The model that added the last three modules at the same time has the best F1-Score, OA and Kappa coefficient compared to any other model. Figure 6 In the No. 8 sub-image, the area with the smallest recognition error is the smallest, and there is no other obvious case of identifying the mangrove area as a non-mangrove area. Therefore, although the addition of the attention mechanism and the Dropout layer introduces the complexity of the model, making the feature distribution more dynamic and elusive. However, the BN layer helps to normalize the distribution of features in such a complex model, making the model easier to converge and improving the accuracy of model recognition. In summary, adding SE-Net, Dropout, and BN modules at the same time can significantly improve the model's water body recognition ability.

[0118] Comparative experiment: In order to more comprehensively verify the performance of the model and compare it with its benchmark model U-Net and other mainstream Seg-Net, Dense-Net, and Res-Net deep learning networks, this embodiment selects two areas outside the sample area for prediction, namely, areas with more mangrove vegetation and areas with less vegetation. The area size is 500 pixels × 500 pixels. The parameter settings of the model are shown in Table 6.

[0119] First, the area with more mangrove vegetation covers a variety of different types of mangrove plants and complex terrain. The model here not only needs to have the ability to better handle details and textures, but also needs to have rich information extraction capabilities and accuracy, which is conducive to comparing the performance of the model in the mangrove diversity scene. Secondly, the area with less vegetation selected in this embodiment involves land features such as water areas and bare land. The model needs to better identify and classify these different types of land features, which is conducive to comparing the generalization ability of the model.

[0120] Finally, by selecting these two types of areas, we can compare the performance of the model in different scenarios and thus comprehensively evaluate its performance. Areas with more vegetation are used to compare the complexity and richness of the model, while areas with less vegetation are used to compare the discrimination and classification accuracy of the model.

[0121] Table 6 Training parameter settings

[0122] parameter Specific settings Batch size 16 Learning Rate 1e-4 Number of training sessions 65 Optimizer Adam

[0123] The accuracy and loss of the training set and test set based on the fused image of the model in this embodiment are as follows: Figure 7 As shown. The line graph on the left represents the accuracy of the training set and the validation set, the horizontal axis is the number of iterations, and the vertical axis is the accuracy. The line graph on the right represents the loss value of the training set and the validation set, the horizontal axis is the number of iterations, and the vertical axis is the loss value.

[0124] The prediction results of the training model network AttU-Net of this embodiment are compared with its benchmark model U-Net and other three mainstream deep learning networks Seg-Net, Dense-Net, and Res-Net for the two selected test areas. The evaluation indicators are F1-Score, OA and Kappa coefficient. The prediction result comparison chart is shown in Figure 2. Figure 8 The accuracy evaluation results of test areas 1-4 are shown in Table 7.

[0125] Depend on Figure 8 It can be seen that the U-Net network presents better visual effects than the other three mainstream deep learning networks Seg-Net, Dense-Net, and Res-Net. The AttU-Net network proposed in this embodiment not only inherits the characteristics of the U-Net network structure itself, such as being able to retain high-resolution information, but also has the best visual results, is more accurate in identifying mangrove vegetation, and has a better ability to identify mangroves and non-mangrove areas. It can more effectively adapt to the complex structure and texture of mangrove vegetation.

[0126] Table 7 Accuracy evaluation results of test areas 1-4

[0127]

[0128] According to the accuracy evaluation results in Table 7, it can be seen that, whether in test area 1 or test areas 2-4, the accuracy evaluation indicators of the AttU-Net model proposed in this embodiment are optimal compared with other models.

[0129] In test area 1 where there are fewer mangrove areas, the F1 score and kappa coefficient of AttU-Net proposed in this embodiment have a cliff-like improvement over its baseline network and the other three networks. Compared with the baseline network U-Net, the F1 score and kappa coefficient of AttU-Net have increased by 6.425% and 7.08% respectively. Its F1 score is 12.906% higher than that of Res-Net, the model with the best F1 score among the other three models. Its kappa coefficient is 13.903% higher than that of Dense-Net, the model with the best kappa coefficient among the other three models. Therefore, when dealing with fewer mangrove areas, AttU-Net performs better than other networks.

[0130] In the test area 2 with more mangroves, although the overall accuracy and F1 score of AttU-Net are not significantly different from other models, the kappa coefficient is significantly improved compared with other networks, reaching 88.959%. This shows that although other networks have achieved higher accuracy and comprehensive performance in the recognition of mangrove vegetation as a whole, they perform poorly in considering the consistency and randomness of classification in the mangrove vegetation recognition task. This is more obvious, such as Figure 8 The prediction results of Seg-Net in the test area 2 show obvious defects, especially in the details of the background information. The recognition performance of the river channel is poor, or even not reflected at all. Therefore, although its OA can reach 92.363% and the F1 score can reach 95.753%, its kappa coefficient is only 58.229%, which shows that it is difficult to ensure the classification consistency of the model in different categories in the face of category imbalance. The kappa coefficient of the AttU-Net model proposed in this embodiment is the best. Compared with the baseline network U-Net, the kappa coefficient of AttU-Net has increased by 9.265%. Its kappa coefficient is 8.034% higher than that of Dense-Net, the model with the best kappa coefficient among the other three models.

[0131] In test area 3, which is disturbed by farmland and houses, the performance of all models has declined, and the overall accuracy cannot reach the average level of the first two study areas. This shows that human interference can easily cause changes in the texture, shape, and color of mangroves, making it difficult for the model to accurately identify mangrove vegetation. From the figure, the texture and shape of the mangrove area adjacent to the house have changed significantly, and the color has also become lighter. Compared with the best performing baseline network Unet in the comparison model, the AttU-Net model proposed in this embodiment has been improved, with OA, F1-Score and kappa increased by 0.399%, 1.707% and 1.842% respectively. It is verified that the improvement method proposed in this embodiment is effective.

[0132] In test area 4, where there is more serious interference from farmland and houses, the overall accuracy of the four models cannot exceed 90%. Except for the Seg-Net network, the performance of other models has declined seriously. However, the AttU-Net model proposed in this embodiment still has a relatively high performance. The kappa coefficient is 7.377% higher than that of the U-Net model with the best kappa coefficient among the other three models, the F1-Score is 3.243% higher than that of Seg-Net, and the OA is 3.321% higher than that of Unet. In this test area, the performance of the Seg-Net model is greatly improved compared with that of test area 3. Combined with test areas 1-2, it is not difficult to see that the Seg-Net model is particularly sensitive to areas with green color and tortuous texture morphology. The larger the proportion of mangrove area in the test area, the better the recognition effect, and the smaller the proportion, the worse the recognition effect.

[0133] In summary, by combining the comparison chart of mangrove vegetation prediction results in the test area and the accuracy evaluation results of test areas 1-4, AttU-Net shows higher overall performance, better detail capture ability and better robustness against category imbalance in the mangrove vegetation recognition task. Therefore, AttU-Net is an effective model suitable for the study of high-precision recognition of mangrove vegetation by fusion images, and can play a certain positive role in the field of mangrove vegetation detection and protection.

[0134] In order to extract mangrove vegetation information more accurately, this embodiment proposes a method for pixel-level weighted fusion of SAR images and optical images at the image level. At the method level, an AttU-Net model is established to accurately identify mangrove vegetation information. In order to verify the effectiveness of the fused image, this embodiment is based on the AttU-Net model, and the fused image training model is based on different weighted ratios. Finally, the fused image with a weighted ratio of 2:8 is finally selected through comparative experiments. In order to verify the effectiveness of the AttU-Net model, it is compared with the benchmark model U-Net and other three mainstream deep learning networks Seg-Net, Dense-Net, and Res-Net on the prediction results of the selected two test areas. The results show that the model has higher overall performance, better detail capture ability, and better robustness against category imbalance in identifying mangrove vegetation. The average OA, F1-Score and Kappa coefficients in the four test areas are 94.406%, 90.006% and 84.045%. Obviously, the AttU-Net model can play a certain positive role in the field of mangrove vegetation detection and protection.

[0135] A mangrove vegetation identification system, comprising:

[0136] A data acquisition module, used to obtain an optical image of the area to be identified and a corresponding fully polarized decomposed image;

[0137] An image fusion module, used for fusing the optical image of the area to be identified and the corresponding fully polarized decomposed image to obtain a fused image to be identified;

[0138] An image recognition module is used to input the fused image to be identified into a mangrove vegetation recognition model for classification prediction to obtain a mangrove vegetation recognition result; wherein the mangrove vegetation recognition model includes an image segmentation module and a sliding overlapping splicing prediction model connected in sequence, and the image segmentation module is constructed based on a U-Net network.

[0139] An electronic device comprises a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program so that the electronic device executes the mangrove vegetation identification method.

[0140] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for identifying mangrove vegetation is implemented.

[0141] The above is only a preferred specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for identifying mangrove vegetation, characterized in that: include: Obtaining an optical image of the area to be identified and a corresponding fully polarized decomposed image; Performing image fusion on the optical image of the area to be identified and the corresponding fully polarized decomposed image to obtain a fused image to be identified; The fused image to be identified is input into a mangrove vegetation identification model for classification prediction to obtain a mangrove vegetation identification result; wherein the mangrove vegetation identification model includes an image segmentation module and a sliding overlapping splicing prediction model connected in sequence, and the image segmentation module is constructed based on a U-Net network.

2. A method for identifying mangrove vegetation according to claim 1, characterized in that: The process of acquiring the optical image of the area to be identified specifically includes: Obtain a panchromatic image and a corresponding multispectral image of the area to be identified, perform radiation correction, atmospheric correction, and orthorectification on the panchromatic image and the corresponding multispectral image, respectively, to obtain corrected panchromatic image and multispectral image, and fuse the corrected panchromatic image and multispectral image to obtain an optical image of the area to be identified.

3. A method for identifying mangrove vegetation according to claim 1, characterized in that: The process of acquiring the fully polarized decomposed image specifically includes: An L1A-level single-view complex image of the area to be identified is acquired, and the L1A-level single-view complex image is sequentially subjected to Pauli vector transformation, polarization coherence matrix transformation, full polarization filtering and reflection symmetric decomposition processing to obtain the full polarization decomposed image.

4. A method for identifying mangrove vegetation according to claim 1, characterized in that: The optical image of the area to be identified and the corresponding fully polarized decomposed image are fused, and the specific calculation formula is: Based on the bilinear interpolation method, the optical image of the area to be identified and the corresponding full polarization decomposition image are resized to obtain an optical image and a corresponding full polarization decomposition image with the same size and matching texture information; the resized optical image and the corresponding full polarization decomposition image are weighted fused to obtain the fused image to be identified.

5. A method for identifying mangrove vegetation according to claim 1, characterized in that: The training process of the mangrove vegetation recognition model specifically includes: Acquire a training image, wherein the training image includes a fused training image to be identified and a corresponding mangrove vegetation identification result; An initial mangrove vegetation recognition model is constructed, the training image is input into the initial mangrove vegetation recognition model for classification prediction, and training is performed with the goal of minimizing the loss between the initial training result after classification prediction and the mangrove vegetation recognition result corresponding to the fused training image to be identified, so as to obtain the mangrove vegetation recognition model.

6. A method for identifying mangrove vegetation according to claim 1, characterized in that: The processing process of the mangrove vegetation identification model includes: The fused image to be identified is input into the image segmentation module for image segmentation to obtain the image to be predicted, the image to be predicted is processed through a sliding window, the image to be predicted is divided into multiple small images, classification prediction is performed on each small image, the classification prediction results are spliced, the overlapping part is selected and the maximum value is taken as the final prediction value to obtain the mangrove vegetation recognition result.

7. A method for identifying mangrove vegetation according to claim 1, characterized in that: The network structure of the image segmentation module is arranged with 5 layers from top to bottom, and the first layer, the second layer, the third layer and the fourth layer are 4 groups of encoder parts and decoder parts with symmetrical structures; The second, third, fourth and fifth layers all introduce an attention mechanism. After the convolution operation, each layer will be processed by batch normalization and calculated using the ReLU activation function; the convolution layers of the fourth and fifth layers are also connected with a Dropout layer.

8. A mangrove vegetation identification system, characterized in that: include: A data acquisition module, used to obtain an optical image of the area to be identified and a corresponding fully polarized decomposed image; An image fusion module, used for fusing the optical image of the area to be identified and the corresponding fully polarized decomposed image to obtain a fused image to be identified; An image recognition module is used to input the fused image to be identified into a mangrove vegetation recognition model for classification prediction to obtain a mangrove vegetation recognition result; wherein the mangrove vegetation recognition model includes an image segmentation module and a sliding overlapping splicing prediction model connected in sequence, and the image segmentation module is constructed based on a U-Net network.

9. An electronic device, characterized in that: It comprises a memory and a processor, the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute a mangrove vegetation identification method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that: It stores a computer program, which, when executed by a processor, implements a mangrove vegetation identification method as described in any one of claims 1 to 7.

Images