Yellow sea enteromorpha remote sensing recognition algorithm based on edge enhanced deep learning

Abstract

The invention discloses a Yellow Sea enteromorpha remote sensing recognition algorithm based on edge enhanced deep learning, and relates to the technical field of marine environment information monitoring. The invention provides a Yellow Sea enteromorpha remote sensing recognition method based on edge enhanced deep learning, and the method can effectively improve the recognition precision of sparse broken enteromorpha and edges through the construction of a specific model, the screening of key sample features and other innovative means, improves the generalization capability and robustness of the model, breaks through the limitation of wave bands, and improves the recognition precision of the sparse broken enteromorpha and edges. And reliable technical support is provided for monitoring and ecological management of the enteromorpha prolifera in the Yellow Sea.

Description

Technical Field

[0001] The present invention relates to the technical field of marine environmental information monitoring, and particularly relates to a remote sensing identification algorithm for Enteromorpha prolifera in the Yellow Sea based on edge-enhanced deep learning. Background Art

[0002] As the largest global green tide disaster, the Enteromorpha prolifera green tide poses multiple threats to the marine ecosystem and economic society. Taking the outbreak of Enteromorpha prolifera in the Yellow Sea as an example, it not only damages aquaculture, blocks waterways, and threatens coastal tourism, but its decomposition also consumes dissolved oxygen in the water and releases hydrogen sulfide, leading to ecological imbalance in local waters.

[0003] Remote sensing monitoring of Enteromorpha prolifera, relying on large-scale and high-timeliness observation technologies that coordinate multiple platforms such as satellites and drones, can achieve precise identification of early small-scale Enteromorpha prolifera patches, dynamic inversion of coverage area, and real-time tracking of drift paths, providing key data support for disaster warning, emergency response, and resource allocation. Long-term accumulated remote sensing data also reveals the correlation between the outbreak of Enteromorpha prolifera and climate change, eutrophication, and human activities, providing a scientific basis for ecological governance and policy optimization, and helping to explore the impact of green tides on macro-ecological processes such as the carbon cycle and biodiversity, promoting the sustainable development of the marine environment.

[0004] In recent years, significant progress has been made in the remote sensing identification technology of Enteromorpha prolifera in the Yellow Sea, mainly relying on multi-source satellite data and algorithm innovation. Optical satellites play a dominant role, using the high reflectance characteristics of Enteromorpha prolifera in the near-infrared band to form a spectral difference with seawater for identification; synthetic aperture radar (SAR) data can be used as a supplement in bad weather.

[0005] Traditional remote sensing identification methods for Enteromorpha prolifera utilize the high reflectance characteristics of Enteromorpha prolifera in the near-infrared band and the spectral difference with seawater, and achieve detection through vegetation indices such as NDVI and FAI combined with threshold segmentation. However, it is easily affected by the environment and has high threshold sensitivity. Although various special indices have been developed subsequently to enhance the anti-interference ability, there are still deficiencies.

[0006] Currently, some studies have introduced artificial intelligence deep learning methods into the research of remote sensing identification of Enteromorpha prolifera to solve various defect problems encountered by traditional remote sensing identification methods for Enteromorpha prolifera. However, the inventor believes that in weak boundary scenarios such as fragmented Enteromorpha prolifera, the mixed edge area of seawater and Enteromorpha prolifera, etc., there are still problems of insufficient extraction of edge information in the deep learning network model. Specifically manifested as: (1) The CNN model (Convolutional Neural Network) relies on data-driven learning of edge features and is prone to ignoring high-frequency details (such as weak boundaries) when the training samples are limited, resulting in blurred boundaries in the segmentation results; (2) As the network deepens, the downsampling operation causes the gradual attenuation of high-order features such as edges, and it is difficult to recover fine boundaries during the decoding process; (3) The fusion of low-level features (edges) and high-level semantic information is vulnerable to background noise interference, and the cross-scale feature alignment is insufficient, affecting the boundary prediction accuracy.

[0007] Therefore, how to solve the above technical problems is an urgent technical problem that needs to be solved by those skilled in the art at present.

[0008] The information disclosed in this background art section is only intended to enhance the overall understanding of the present invention and should not be regarded as an admission or any form of suggestion that this information constitutes prior art already known to those of ordinary skill in the art. Summary of the Invention

[0009] In view of the above technical problems, an embodiment of the present invention provides a remote sensing recognition algorithm for Enteromorpha prolifera in the Yellow Sea based on edge enhancement deep learning to solve the problems proposed in the above background art.

[0010] The present invention provides the following technical solutions: A remote sensing recognition algorithm for Enteromorpha prolifera in the Yellow Sea based on edge enhancement deep learning, comprising the following steps: Build an EeaAglaeNet Enteromorpha prolifera recognition model; the model structure includes: an encoding layer, an edge enhancement module, and a decoding layer; Obtain the band and vegetation index feature parameters in the optical remote sensing image as the initial sample data, and make the Enteromorpha prolifera distribution image label data corresponding to the initial sample data; Preprocess the initial sample data and input it into the EeaAglaeNet Enteromorpha prolifera recognition model for training, and use the SHAP analysis method to perform importance weight analysis on the performance of the initial sample data in the model training to obtain the optimal sample data; The optimal sample data includes: Band 3 (0.560μm), Band 5 (0.705μm), Band 8 (0.842μm), and FAI; Band is the band, and FAI is the phytoplankton index; Input the optimal sample data into the EeaAglaeNet Enteromorpha prolifera recognition model for training until the model is formed; the label data is the Enteromorpha prolifera distribution image and the accuracy evaluation result; Use the trained EeaAglaeNet Enteromorpha prolifera recognition model to identify the Enteromorpha prolifera distribution.

[0011] Preferably, the encoding layer adopts the U-Net encoding layer structure, and the length, width, and band of the input data are H, W, and C respectively; the convolution operation uses a 3×3 convolution kernel, the activation function is ReLU, and the downsampling operation uses a 2×2 max pooling layer; after multiple convolutions, activations, and downsamplings, a feature map with a band number of 512 and a length and width that are 1 / 8 of the original input length and width is obtained in the deepest layer.

[0012] Preferably, the input of the edge enhancement module includes three parts. One is the encoded features after dimensionality reduction by 3×3 convolution from the encoding layer; the second is the hierarchical edge information extracted by the Laplacian pyramid; the third is the prediction result downsampled to two dimensions from a higher level of the decoding layer. These three groups of features are subjected to channel concatenation and convolution operations to generate fused features; a spatial attention map is generated through a convolution layer activated by Sigmoid, and a convolutional block attention module is used to calibrate the fused features in the channel and spatial dimensions, and finally the edge enhancement features are output.

[0013] Preferably, the decoding layer adopts the upsampling path of U-Net, and the input is the edge enhancement features output by the edge feature enhancement module; the resolution of the feature map is gradually increased layer by layer through transposed convolution or interpolation operations. The convolution operation uses a 3×3 convolution kernel, the activation function is ReLU, and the upsampling operation uses a 2×2 max pooling layer; finally, the decoded feature maps of each level are output to generate the Enteromorpha recognition result corresponding to the input.

[0014] Preferably, the initial sample data includes: Band 1 (0.443μm), Band 2 (0.490μm), Band 3 (0.560μm), Band 4 (0.665μm), Band 5 (0.705μm), Band 6 (0.740μm), Band 7 (0.783μm), Band 8 (0.842μm), Band 8a (0.865μm), Band 9 (0.945μm), Band 11 (1.610μm), Band 12 (2.190μm); and, NDVI, EVI, DVI, FAI and SRG; where NDVI is the Normalized Difference Vegetation Index, EVI is the Enhanced Vegetation Index, DVI is the Difference Vegetation Index, and SRG is the Simple Green Band Ratio Vegetation Index.

[0015] Preferably, the methods for preprocessing the optimal sample data include: radiometric calibration, atmospheric correction, resampling and land masking.

[0016] Preferably, the input feature dataset implements a three-stage segmentation strategy: first, the input features are randomly split into a training set and a test set at a ratio of 7:3, and the test set is independently used to evaluate the generalization performance of the wind field reconstruction model; 20% of the samples are secondarily extracted from the training set to construct a validation set for real-time monitoring of the convergence stability during the model training process; among them, the label data and the input feature dataset adopt the same partitioning process as the input feature data.

[0017] Preferably, the input feature dataset is standardized by using the range normalization method.

[0018] Preferably, the loss function in the model adopts a composite loss function, which is jointly composed of binary cross-entropy and Dice loss.

[0019] Preferably, the accuracy evaluation result indicators are: Accuracy, Precision, Recall, F1 Score, and IoU; Among them, Accuracy is the accuracy rate, Precision is the precision, Recall is the recall rate, F1 Score is the F1 score, and IoU is the intersection over union.

[0020] The remote sensing recognition algorithm for yellow sea enteromorpha based on edge enhancement deep learning provided by the embodiments of the present invention has the following beneficial effects: 1. The present invention introduces an edge enhancement module and designs a multi-scale pyramid structure to directly capture multi-scale edge information from the original image, making up for the blurred segmentation result boundary caused by the deficiency of CNN in weak boundary learning; 2. Retain edge details at different resolutions to avoid edge degradation; integrate a lightweight convolutional block attention module to further optimize feature correlation and improve boundary localization accuracy; 3. In addition, this method also has strong transferability and is not restricted by the band settings of remote sensing satellite data sources, thereby improving the generalization ability of the model and enhancing the robustness and accuracy of enteromorpha remote sensing recognition; 4. When the present invention selects band data as input, it also introduces a variety of vegetation indices as physical information auxiliary inputs; the vegetation index data can enhance the difference between enteromorpha and other ground objects, highlight the characteristics of enteromorpha, so as to more accurately judge the boundary position of enteromorpha, reduce the recognition error caused by factors such as clouds and seawater, and improve the accuracy and reliability of enteromorpha edge recognition. Description of the Drawings

[0021] Figure 1 It is the flow chart of the remote sensing recognition algorithm for yellow sea enteromorpha based on edge enhancement deep learning in the present invention; Figure 2 It is the framework diagram of the EeaAglaeNet enteromorpha recognition network model in the present invention; Figure 3 It is the network framework diagram of the edge enhancement module in the present invention; Figure 4 It is the result of the feature importance ranking in the present invention; Figure 5 It is the comparison result of the EeaAglaeNet model and other classical enteromorpha recognition model results. Detailed Embodiments

[0022] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the present invention.

[0023] In view of the problems mentioned in the above background technology, the embodiments of the present invention provide a remote sensing recognition algorithm for yellow sea ulva prolifera based on edge-enhanced deep learning to solve the above technical problems, and the technical solutions are as follows: The following combines the attached Figures 1-5 , and the specific implementation manners to further illustrate the present invention.

[0024] The design idea and optimization steps of a remote sensing recognition algorithm for yellow sea ulva prolifera based on edge-enhanced deep learning provided by the present invention are as follows: A remote sensing recognition method for yellow sea ulva prolifera based on edge-enhanced deep learning includes the following steps: Step 1: Obtain remote sensing satellite data containing ulva prolifera; for optical remote sensing data, obtain the ulva prolifera distribution image data information corresponding thereto.

[0025] Step 2: Perform data preprocessing on the remote sensing satellite image data according to the relevant preprocessing methods of optical remote sensing images. The preprocessing process includes radiometric calibration, atmospheric correction, resampling, and land masking processing.

[0026] Step 3: Extract the band and vegetation index characteristic parameters in the optical remote sensing image; the vegetation index results and each band data jointly form a characteristic data set.

[0027] Among them, the vegetation index is extracted by using band calculation, and the vegetation indices used are: NDVI, EVI, DVI, FAI, and SRG indices. The index calculation formulas are as follows:

[0028] In the formula, NIR, RED, SWIR, GREEN, are the near-infrared band reflectance, red band reflectance, short-wave infrared reflectance, green band reflectance, near-infrared wavelength, red wavelength, and short-wave infrared wavelength in the optical image.

[0029] Step 4: Use the Python programming language to construct a remote sensing recognition network for yellow sea ulva prolifera based on edge-enhanced deep learning of the remote sensing optical data set (EeaAglaeNet). The specific steps are as follows: (1) Model input characteristic data set and output feature construction.

[0030] There are many band data involved in the imaging process of remote sensing optical data and current methods of monitoring Enteromorpha prolifera using vegetation indices. The environment for Enteromorpha prolifera recognition is affected by clouds and seawater. To improve the recognition accuracy of the Enteromorpha prolifera edge, while using band data as input in the present invention, a variety of vegetation indices are also introduced as auxiliary inputs.

[0031] Taking Sentinel-2 as an example, the two input features altogether contain 17 features: band 1 (0.443μm), band 2 (0.490μm), band 3 (0.560μm), band 4 (0.665μm), band 5 (0.705μm), band 6 (0.740μm), band 7 (0.783μm), band 8 (0.842μm), band 8a (0.865μm), band 9 (0.945μm), band 11 (1.610μm), band 12 (2.190μm); and, Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), Difference Vegetation Index (DVI), Floating Algae Index (FAI), and Simple Ratio Green (SRG). Finally, the shape of the input feature dataset is constructed as (n, x, y, c); where n is the number of samples, x and y are the length and width of the sample respectively, and c is the number of features. The output features of the model are: the Enteromorpha prolifera distribution image and the accuracy evaluation result.

[0032] (2) Feature importance analysis of the model.

[0033] Based on the SHAP interpretability model framework, the importance weights of multi-dimensional variables in the Enteromorpha prolifera feature dataset (n, x, y, c) are analyzed. By quantifying the contribution degrees of each feature to the classification result and sorting them, the feature subset that drives the optimal recognition efficiency of the model is selected to maximize the improvement of the Enteromorpha prolifera detection accuracy. After screening, the features of the optimal feature subset obtained are: band3, band5, band8, FAI.

[0034] (3) Division of the training set, validation set, and test set.

[0035] Implement a three-stage segmentation strategy based on the optimized feature dataset: First, randomly split the input features into a training set and a test set at a ratio of 7:3. The test set is independently used to evaluate the generalization performance of the remote sensing identification model for Enteromorpha prolifera in the Yellow Sea. On this basis, 20% of the samples are secondarily extracted from the training set to construct a validation set for real-time monitoring of the convergence stability during model training. This scheme synchronously implements the isomorphic partitioning process of the label data to ensure strict spatio-temporal consistency of the feature-label data pairs in the training, validation, and test links.

[0036] (4)Data standardization (normalization) To address the modeling deviation problem caused by the dimensional differences of multi-dimensional features, the present invention constructs a data normalization preprocessing process. When the input parameters with different dimensions are not scaled uniformly, high-magnitude features tend to dominate the model learning direction, while the contribution of low-magnitude features is nonlinearly compressed. To eliminate the interference of dimensional differences on the modeling process, this scheme adopts the range normalization method to compress each feature value to the [0, 1] interval through linear mapping. Its mathematical expression is: (6); In the formula, and respectively represent the minimum and maximum observed values of the feature variable. This transformation realizes the balanced representation of each dimension feature while retaining the original data distribution law, providing a stable input for subsequent model construction.

[0037] (5)Model structure The Enteromorpha prolifera identification model EeaAglaeNet constructed by the present invention includes: an encoding layer (Encoder), an edge enhancement module, and a decoding layer (Decoder); among them, the network model framework of EeaAglaeNet for Enteromorpha prolifera is as Figure 2 shown.

[0038] First, the model retains the basic structure of the U-Net network, that is, it retains the structures of the encoding layer and the decoding layer. And it retains the practice of introducing the data of the encoding layer during the decoding process to reduce data loss. Assume that the length, width, and number of bands of the input data are H, W, and C respectively. In the encoding module, the convolution operation uses a 3*3 convolution, the activation function is ReLu, and the downsampling operation uses a 2*2 max pooling layer. The deepest layer obtains a band number of 512, and the length and width are 1 / 8 of the original input length and width respectively.

[0039] In the Edge-Enhanced Attention (EEA) module, the main functions are as follows: fusing the encoded features, high-frequency edge information, and high-level prediction features among multiple scales to enhance the ability to retain boundary details. The inputs of this module are: ① the encoded features after dimensionality reduction by a 3×3 convolution from the encoding layer; ② the hierarchical edge information extracted by the Laplacian pyramid; ③ the prediction results downsampled to two dimensions from a higher level of the decoding layer. Channel concatenation and convolution operations are performed on the above three groups of features to generate fused features. A convolutional layer activated by Sigmoid generates a spatial attention map (Attention MASK) to suppress noise and focus on key edge regions. The Convolutional Block Attention Module (CBAM) is used to calibrate the fused features in the channel and spatial dimensions, and the edge-enhanced features are output. The network framework of this module is as shown in Figure 3 shown.

[0040] In the decoding module, a U-Net style upsampling path is adopted, and the resolution of the feature map is gradually increased layer by layer through transposed convolution or interpolation operations. The convolution operation uses a 3*3 convolution, the activation function is ReLu, and the upsampling operation uses a 2*2 max pooling layer. The input is the edge-enhanced features output by the Edge Feature Enhancement Module (EEA), and the output is the decoding feature maps of each level, and finally the recognition result of the input Enteromorpha is generated.

[0041] In terms of the design of the loss function, a composite loss function is adopted, which is jointly composed of Binary Cross-Entropy (BCE) and Dice loss. Among them, Binary Cross-Entropy is used to quantify the difference in the probability distribution between the predicted value and the true label, and Dice loss enhances the sensitivity of the model to the shape of the target region by calculating the proportion of the overlapping area between the predicted mask and the true mask. After weighted superposition of the two, the formula is expressed as: (7); In terms of the optimization strategy, the Adaptive Moment Estimation (Adam) algorithm is adopted, which combines the dynamic adjustment mechanisms of the first moment and the second moment to effectively balance the direction and step size of gradient update. The initial learning rate is set to 0.0001, and the parameter update amplitude is restricted by gradient clipping (ClipValue = 0.5) to further ensure the stability of the training process.

[0042] (4) Indexes of the accuracy evaluation results of the model.

[0043] The selected accuracy evaluation result indicators for this invention are: Accuracy, Precision, Recall, F1 Score, and Intersection over Union (IoU). Accuracy can measure the classification correctness of the model on all samples, and the calculation formula is as follows: (8); Precision can measure the proportion of actual positive classes among the positive samples predicted by the model, and the calculation formula is as follows: (9); Recall can measure the detection ability of the model for Enteromorpha prolifera targets, and the calculation formula is as follows: (10); F1 Score is the harmonic mean of precision and recall, and the calculation formula is as follows: (11); Intersection over Union (IoU) can measure the overlapping degree between the Enteromorpha prolifera prediction result and the true label, and show the consistency between the Enteromorpha prolifera prediction result and the actual situation. The calculation formula is as follows: (12); Among them, TP is the true positive example, TN is the true negative example, FP is the false positive example, and FN is the false negative example. Specific embodiments

[0044] In this embodiment, the multi-spectral image data of the Sentinel-2A / B satellite is used as the verification object to demonstrate the remote sensing monitoring method of Enteromorpha prolifera in the Yellow Sea based on the edge enhancement deep learning framework. Users need to deploy a Python 3.6.13 running environment on the local terminal and integrate the TensorFlow 1.15.0 and Keras 2.3.1 software libraries. Among them, TensorFlow is an open-source deep learning framework that supports distributed computing and provides the underlying interfaces for building, training, and deploying neural network models. Keras is a high-level API based on TensorFlow that simplifies the rapid construction and debugging process of complex models through modular design.

[0045] Step 1: Input feature contribution degree analysis.

[0046] During the model training process, the number and combination of input features directly affect the training efficiency and prediction performance. Too few parameters are likely to cause underfitting problems, making it difficult for the model to capture the deep correlation rules in the data; too many parameters may exacerbate the overfitting risk, increase the computational complexity, and prolong the inference latency. To optimize the input configuration of the EeaAglaeNet model, this solution uses SHAP (Shapley Additive Explanations) values to evaluate the contribution of multi-source features and screen out the feature sets that have significant positive or negative effects on the model output.

[0047] There are two major categories of input features for the SHAP method, a total of 17 kinds. The first category is multi-spectral band feature data: band 1 (0.443μm), band 2 (0.490μm), band 3 (0.560μm), band 4 (0.665μm), band 5 (0.705μm), band 6 (0.740μm), band 7 (0.783μm), band 8 (0.842μm), band 8a (0.865μm), band 9 (0.945μm), band 11 (1.610μm), and band 12 (2.190μm). The second category is vegetation index feature data: NDVI, EVI, DVI, FAI, and SRG vegetation indices. The contribution of the input features is as Figure 4 shown.

[0048] Integrate the above input features to construct an optimized feature dataset with a data dimension of (n, 128, 128, 17). Divide the dataset according to a preset ratio: 70% of the total samples are used for model training (20% of which are used as the validation set to tune the hyperparameters), and the remaining 30% are used as an independent test set to ensure the objectivity of the model performance evaluation. For the performance of the model in the test set. Based on the quantitative evaluation results of the feature contribution, key spectral bands and vegetation index combinations are screened out, including Band3 (0.560μm), Band5 (0.705μm), Band8 (0.842μm), and FAI. Integrate the above features to construct an optimized feature dataset.

[0049] Step 2: Accuracy analysis. Use the optimized feature dataset screened in Step 1 and input it into the remote sensing identification model of yellow sea ulva (EeaAglaeNet) based on edge enhancement deep learning of the present invention. The accuracy in the result accuracy evaluation index is 0.9762, and the precision value reaches 0.9530. By improving the loss function and enhancing the operation of the edge extraction module, it can be found that the accuracy of the model has been greatly improved.

[0050] Step 3: Comparison with other classical Enteromorpha prolifera recognition models. To verify the effectiveness of EeaAglaeNet in the remote sensing recognition of the edges of Enteromorpha prolifera in the Yellow Sea and fragmented Enteromorpha prolifera in this implementation case, the inversion results were compared with the inversion results of the classical neural network model U-Net and the lightweight Enteromorpha prolifera recognition model AglaeNet model. Among them, the comparison results of the test set are shown in Table 1. Compared with other classical inversion models, the accuracy of the EeaAglaeNet model proposed in the present invention has increased by 4.02%.

[0051] As Figure 5 shown, the extraction effect of the model proposed in the present invention in terms of the edges of Enteromorpha prolifera and fragmented Enteromorpha prolifera has been greatly improved. By comparing the results of different models, it can be known that the U-Net deep learning network framework has good recognition ability in the field of Enteromorpha prolifera recognition in the Yellow Sea. The traditional U-Net deep learning model can achieve an overall recognition accuracy of 0.9128. However, in the lightweight deep learning network model AlgaeNet, since its design is a deep learning model for the red, green, and blue three bands, its performance on the multi-feature data set of remote sensing reflectance and vegetation index needs to be improved.

[0052] The EeaAglaeNet model proposed in the present invention can more effectively identify fragmented Enteromorpha prolifera and the edges of Enteromorpha prolifera, reduce the missed detection rate of Enteromorpha prolifera, and improve the accuracy and precision of Enteromorpha prolifera recognition. Figure 5 As shown, the results of Enteromorpha prolifera recognized by each model have been circled and marked with red boxes in the figure; from the circled results, it can be seen that the fragmented Enteromorpha prolifera and the edges of Enteromorpha prolifera that were not recognized by the traditional model method can have good recognition effects in the EeaAglaeNet model.

[0053] Table 1 List of model accuracy evaluations Precision IoU Recall F1 Score U-Net 0.9128 0.7785 0.7785 0.8301 AglaeNet 0.8565 0.6582 0.7399 0.7940 EeaAglaeNet 0.9530 0.7016 0.7268 0.8246 In summary, the present invention proposes a method for remote sensing recognition of Enteromorpha prolifera in the Yellow Sea based on the fusion of edge enhancement and multi-scale attention, which can more effectively identify fragmented Enteromorpha prolifera in the Yellow Sea; the present invention has significant advantages in edge extraction, multi-scale fusion, and data adaptability; by introducing an edge enhancement mechanism and a multi-scale pyramid structure, the recognition accuracy of the model for fragmented Enteromorpha prolifera and its boundaries has been effectively improved; combined with the channel-spatial attention mechanism, the correlation between features has been enhanced, background noise interference has been suppressed, and the recognition effect under complex sea conditions has been improved.

[0054] Meanwhile, based on feature importance analysis, the present invention constructs a band-independent data input scheme, breaking through the dependence on specific remote sensing bands and realizing the compatible migration of multi-source remote sensing data (such as optical and SAR). In view of the characteristics of Enteromorpha prolifera in the Yellow Sea, a new loss function is designed to further improve the recognition accuracy of the edge area and the robustness of the model, realizing the fine extraction of fragmented Enteromorpha prolifera and the edge of Enteromorpha prolifera, and providing high-reliability technical support for disaster emergency response and ecological governance decision-making.

[0055] Although the specific implementation manners of the present invention have been described above in conjunction with the accompanying drawings, they are not limitations on the protection scope of the present invention. Those skilled in the art should understand that, based on the technical solutions of the present invention, various modifications or deformations that can be made by those skilled in the art without creative efforts are still within the protection scope of the present invention.

Claims

1. A remote sensing recognition algorithm for Enteromorpha in the Yellow Sea based on edge-enhanced deep learning, characterized in that: The following steps are involved: Build the EeaAglaeNet Enteromorpha recognition model; the model structure includes: encoding layer, edge enhancement module and decoding layer; Acquire the band and vegetation index characteristic parameters in the optical remote sensing image as initial sample data, and generate enteromorpha distribution image label data corresponding to the initial sample data; After preprocessing, the initial sample data is input into the EeaAglaeNet Enteromorpha recognition model for training. The SHAP analysis method is used to analyze the importance weight of the performance of the initial sample data in the model training to obtain the optimal sample data. The optimal sample data include: Band 3 (0.560μm), Band 5 (0.705μm), Band 8 (0.842μm) and FAI; Band is the band, and FAI is the phytoplankton index; The optimal sample data is input into the EeaAglaeNet Enteromorpha recognition model for training until the model is formed; the label data is the Enteromorpha distribution image and the accuracy evaluation result; The trained EeaAglaeNet enteromorpha recognition model is used to identify the distribution of enteromorpha.

2. The edge-enhanced deep learning-based remote sensing recognition algorithm for Enteromorpha in the Yellow Sea according to claim 1, characterized in that: The encoding layer adopts the U-Net encoding layer structure, and the length, width and band of the input data are H, W and C respectively; the convolution operation uses a 3×3 convolution kernel, the activation function is ReLU, and the downsampling operation uses a 2×2 maximum pooling layer; after multiple convolutions, activations and downsampling, the deepest layer obtains a feature map with 512 bands and a length and width that are 1 / 8 of the original input length and width.

3. The edge-enhanced deep learning-based remote sensing recognition algorithm for Enteromorpha in the Yellow Sea according to claim 1, characterized in that: The input of the edge enhancement module consists of three parts: the first is the coding features from the coding layer after 3×3 convolutional dimension reduction; The second is the hierarchical edge information extracted through the Laplacian pyramid; the third is the prediction result from the higher level of the decoding layer downsampled to two dimensions; The three sets of features are channel-joined and convolved to generate fused features. The spatial attention map is generated through the Sigmoid-activated convolution layer, and the convolution block attention module is used to calibrate the channel and spatial dimensions of the fused features, and finally the edge enhancement features are output.

4. The edge-enhanced deep learning-based remote sensing recognition algorithm for Enteromorpha in the Yellow Sea according to claim 1, characterized in that: The decoding layer adopts the U-Net upsampling path, and the input is the edge enhancement feature output by the edge feature enhancement module; the feature map resolution is improved layer by layer through transposed convolution or interpolation operations. The convolution operation uses a 3×3 convolution kernel, the activation function is ReLU, and the upsampling operation uses a 2×2 maximum pooling layer; finally, the decoding feature maps of each level are output to generate the Enteromorpha recognition results corresponding to the input.

5. The edge-enhanced deep learning-based remote sensing recognition algorithm for Enteromorpha in the Yellow Sea according to claim 1, characterized in that: The initial sample data includes: Band 1 (0.443μm), Band 2 (0.490μm), Band 3 (0.560μm), Band 4 (0.665μm), Band 5 (0.705μm), Band 6 (0.740μm), Band 7 (0.783μm), Band 8 (0.842μm), Band 8a (0.865μm), Band 9 (0.945μm), Band 11 (1.610μm), Band 12 (2.190μm); As well as NDVI, EVI, DVI, FAI and SRG; among them, NDVI is the Normalized Difference Vegetation Index, EVI is the Enhanced Vegetation Index, DVI is the Difference Vegetation Index, and SRG is the Simple Green Band Ratio Vegetation Index.

6. The edge-enhanced deep learning-based remote sensing recognition algorithm for Enteromorpha in the Yellow Sea according to claim 1, characterized in that: The optimal sample data preprocessing methods include: radiation calibration, atmospheric correction, resampling and land mask.

7. The edge-enhanced deep learning-based remote sensing recognition algorithm for Enteromorpha in the Yellow Sea according to claim 1, characterized in that: A three-stage segmentation strategy was implemented for the input feature dataset: first, the input features were randomly split into a training set and a test set with a ratio of 7:

3. The test set was independently used to evaluate the generalization performance of the remote sensing recognition model for Enteromorpha in the Yellow Sea; 20% samples are extracted from the training set for the validation set, which is used to monitor the convergence stability of the model training process in real time. The label data and the input feature data set adopt the same partitioning process as the input feature data.

8. The edge-enhanced deep learning-based remote sensing recognition algorithm for Enteromorpha in the Yellow Sea according to claim 1, characterized in that: The input feature data set is standardized using the range normalization method.

9. The edge-enhanced deep learning-based remote sensing recognition algorithm for Enteromorpha in the Yellow Sea according to claim 1, characterized in that: The loss function in the model adopts a composite loss function, which is composed of binary cross entropy and Dice loss.

10. The edge-enhanced deep learning-based remote sensing recognition algorithm for Enteromorpha in the Yellow Sea according to claim 1, characterized in that: The accuracy evaluation result indicators are: Accuracy, Precision, Recall, F1 Score and IoU; Among them, Accuracy is the accuracy, Precision is the accuracy, Recall is the recall rate, F1 Score is the F1 score, and IoU is the intersection over union ratio.

Images