Enteromorpha remote sensing image segmentation method based on channel boundary attention

Abstract

The invention provides an enteromorpha remote sensing image segmentation method based on channel boundary attention, and relates to the technical field of image processing, and the method comprises the steps: processing an enteromorpha remote sensing image through employing an improved adaptive histogram equalization spectrum enhancement algorithm, and obtaining a spectrum enhancement optimization image; based on channel boundary attention, constructing an image segmentation model by using a feature extraction module, a boundary enhancement module and a multi-scale feature fusion and decoding module; training an image segmentation model by using the spectrum enhancement optimization image to obtain a trained image segmentation model; and reasoning the enteromorpha remote sensing image by using the trained image segmentation model to obtain an enteromorpha remote sensing image segmentation result, and completing the segmentation of the enteromorpha remote sensing image. The method solves the problems that an existing depth segmentation model is insufficient in enteromorpha identification precision, fuzzy in boundary and poor in small target detection capability in a complex marine environment.

Description

TECHNICAL FIELD

[0001] The present specification relates to the technical field of image processing, in particular to a Enteromorpha prolifera remote sensing image segmentation method based on channel boundary attention. BACKGROUND

[0002] The frequent outbreak of Enteromorpha prolifera has a serious impact on the nearshore ecological environment and coastal economy, so it is of great significance to use satellite remote sensing technology to monitor the spatio-temporal distribution of Enteromorpha prolifera with high precision. Traditional Enteromorpha prolifera detection methods mostly rely on spectral threshold, vegetation index or supervised classification, etc. However, in complex marine environment, it is easy to be disturbed by water color, cloud and mist, and sea surface reflection, and it is difficult to accurately distinguish Enteromorpha prolifera from other marine targets.

[0003] Although the existing Enteromorpha prolifera remote sensing image detection method can realize the automatic identification of sea surface Enteromorpha prolifera to some extent, it still has many shortcomings. The traditional method based on spectral threshold or vegetation index is easily affected by water color change, cloud and mist shielding and sea surface reflection in complex marine environment, resulting in low recognition accuracy. In recent years, the semantic segmentation model based on deep learning (such as SegFormer) has certain advantages in feature extraction and global modeling, but when it is applied to Enteromorpha prolifera detection, there are still the following problems: first, the model has weak recognition ability for the boundary of Enteromorpha prolifera region, and often appears edge blur and target adhesion; second, there is no feature self-adaptive mechanism for multi-scale Enteromorpha prolifera target, resulting in unsatisfactory detection effect for small-scale Enteromorpha prolifera region; third, the model has limited suppression ability for background noise (such as sea waves, floating objects, etc.), which is easy to produce false detection and omission.

[0004] These problems mainly arise from the shortcomings of the existing model in feature channel weight distribution and boundary information extraction, which fails to fully consider the texture features and edge differences of Enteromorpha prolifera target. SUMMARY

[0005] In view of the above shortcomings in the prior art, the Enteromorpha prolifera remote sensing image segmentation method based on channel boundary attention provided by the present application solves the problems of insufficient recognition accuracy of Enteromorpha prolifera, fuzzy boundary and poor detection ability of small targets of the existing deep segmentation model in complex marine environment.

[0006] In order to achieve the above invention purpose, the technical scheme adopted by the present application is as follows: an Enteromorpha prolifera remote sensing image segmentation method based on channel boundary attention, comprising: S1: using an improved adaptive histogram equalization spectral enhancement algorithm to process the Enteromorpha prolifera remote sensing image to obtain a spectrum enhanced optimized image; S2: based on channel boundary attention, using a feature extraction module, a boundary enhancement module and a multi-scale feature fusion and decoding module to construct an image segmentation model; S3: training an image segmentation model by using the spectrum enhancement optimized image to obtain a trained image segmentation model; S4: performing inference on the Enteromorpha prolifera remote sensing image by using the trained image segmentation model to obtain an Enteromorpha prolifera remote sensing image segmentation result, and completing segmentation of the Enteromorpha prolifera remote sensing image.

[0007] The method has the advantages that the method provided by the application is an Enteromorpha prolifera remote sensing image segmentation method based on channel boundary attention, a channel attention mechanism is introduced to enhance the feature extraction capability, and a boundary feature enhancement module is combined to improve the target edge recognition accuracy. To verify the effectiveness of the method, multi-temporal remote sensing images of typical sea areas in the Yellow Sea are selected as experimental data, and the method is compared with a traditional SegFormer model and other deep segmentation algorithms. The experimental results show that the method has achieved significant improvement in Enteromorpha prolifera detection accuracy, boundary detail preservation, and model robustness, and provides a reliable technical means for automatic monitoring of large-scale Enteromorpha prolifera disasters. (1) The boundary information learning of shallow features is realized to achieve accurate detection of small Enteromorpha prolifera targets and improve the boundary integrity. (2) The key feature channel response is enhanced by adaptive channel re-labeling, and the background noise is suppressed, thereby improving the Enteromorpha prolifera segmentation accuracy. (3) The combination of the boundary supervision branch and the ECA module enables the shallow features to maintain global semantic consistency while strengthening the edge details, realizing the cooperative optimization of the decoder and the boundary branch. (4) A complete Enteromorpha prolifera remote sensing image segmentation framework is constructed, including multi-scale feature extraction, ECA channel re-labeling, boundary supervision, and decoding fusion process, which can adapt to various satellite remote sensing data in the Yellow Sea. (5) Through lightweight design, the method realizes comprehensive improvement of mIoU, F1, Precision, and Recall indicators without significantly increasing the computational complexity, effectively balancing accuracy and efficiency.

[0008] Further, the expression of the spectrum enhancement optimized image is: ; ; ; ; ; ; ; wherein, represents the spectrum enhancement optimized image, represents the image abscissa, represents the image ordinate, represents the final enhanced image obtained after converting the enhanced HSV back to RGB, represents the minimum possible value of the image pixel value (e.g. 0), represents the maximum possible value of the image pixel value, represents the reorganized HSV image after brightness enhancement, including channel, represents the hue channel (Hue), represents the saturation channel (Saturation), represents the enhanced value channel (Value), represents the function of converting the HSV image to the RGB color space, represents the local enhancement coefficient, represents the local enhancement coefficient baseline gain of the local enhancement coefficient, used to ensure that the basic brightness preservation ability is still maintained when there is no significant texture or contrast change (i.e. the local variance is close to zero), represents the enhancement sensitivity adjustment coefficient, represents the local standard deviation centered at (x, y) with a window size of w x w, represents the local mean value centered at (x, y) with a window size of w x w, represents a small constant to prevent the denominator from being zero, represents the brightness compensation term, represents the brightness balance factor, represents the global brightness mean value, represents the brightness component, represents the image after converting the original RGB image to the HSV color space, represents the conversion of the original multispectral remote sensing image to the HSV color space, represents the original multispectral remote sensing image.

[0009] Further, the image segmentation model comprises: a feature extraction module, configured to perform feature extraction on the spectral enhancement optimization image by using a channel boundary attention mechanism to obtain an original feature map; a channel boundary attention module, configured to perform local interaction relationship analysis on the original feature map to obtain an input feature map; a boundary enhancement module, configured to perform gradient mapping calculation on the input feature map to obtain a boundary response map; a multi-scale feature fusion and decoding module, configured to perform cascading and weighted fusion on the boundary response map based on a feature fusion path to obtain a green algae remote sensing image segmentation result.

[0010] Further, the expression of the original feature map is: ; ; ; wherein, represents the output feature of the re-calibrated th channel, represents the attention weight of the th channel, represents the th channel feature of the input feature map, represents the global description value of the th channel after global average pooling, represents the feature map height, represents the spatial position x-axis index, represents the spatial position y-axis index, represents the Sigmoid function, represents the ReLU, and represents the weight matrix.

[0011] In this way, the response to the Enteromorpha feature channel can be enhanced, and the interference of background seawater and floating objects can be suppressed, thereby improving the overall feature expression capability.

[0012] Further, the expression of the input feature map is: ; ; ; ; ; wherein, represents the input feature map, represents a three-dimensional feature tensor from the shallow layer of the encoder, H and W represent the spatial resolution, and C represents the number of channels, represents the channel description vector, represents the spatial position x-axis index, represents the spatial position y-axis index, represents the number of channels of the tensor, represents the channel weight vector, represents the Sigmoid activation function, represents a one-dimensional convolution with a length of , represents the feature after channel recalibration, represents the omitted H, W, i.e., the spatial resolution, denotes a weight coefficient of the i-th channel, denotes a weight coefficient of the i-th channel, denotes a kernel length, denotes an operation of rounding to the nearest odd integer, and denotes a hyperparameter.

[0013] Further, the expression of the loss function of the image segmentation model is: ; ; ; ; ; wherein, denotes a loss function of the image segmentation model, denotes a weight coefficient of the cross-entropy loss, denotes a weight coefficient of the boundary loss, denotes a weight coefficient of the shape consistency loss, denotes a cross-entropy loss, denotes a boundary loss, denotes a shape consistency loss, denotes a total number of pixels (all pixels in a training batch), denotes a true label of the i-th pixel (0: background, 1: enteromorpha), denotes a predicted probability of the i-th pixel (probability of belonging to the enteromorpha class), denotes a boundary weight coefficient of the i-th pixel, denotes a predicted boundary value of the i-th pixel (0-1), denotes a true boundary value of the i-th pixel (0 or 1), denotes an enhancement coefficient, denotes a distance of a pixel point to a true boundary, denotes a Sigmoid activation function, denotes a change amount. BRIEF DESCRIPTION OF DRAWINGS

[0014] The present specification will be further illustrated in the manner of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting, and in these embodiments, the same numbers represent the same structures, wherein: Figure 1 is an exemplary flowchart of an enteromorpha remote sensing image segmentation method based on channel boundary attention according to some embodiments of the present specification; Figure 2 is an exemplary schematic diagram of the overall architecture of an image segmentation model according to some embodiments of the present specification; Figure 3 is an exemplary schematic diagram of the embedding position of the boundary supervision branch according to some embodiments of the present specification; Figure 4 is an exemplary schematic diagram of the structure of the boundary supervision branch according to some embodiments of the present specification; Figure 5 is an exemplary schematic diagram of the structure and position of the channel attention module according to some embodiments of the present specification; Figure 6 is an exemplary schematic diagram of the 4,2,1 band pseudo-color Enteromorpha segmentation effect according to some embodiments of the present specification; Figure 7 is an exemplary schematic diagram of the channel mapping and mask according to some embodiments of the present specification; Figure 8 is an exemplary schematic diagram of the mIoU change curve of the original network and the improved network according to some embodiments of the present specification; Figure 9 is an exemplary schematic diagram of the Precision change curve of the original network and the improved network according to some embodiments of the present specification; Figure 10 is an exemplary schematic diagram of the detection effect of different networks according to some embodiments of the present specification. DETAILED DESCRIPTION

[0015] The specific embodiments of the present application are described below to facilitate the understanding of the present application for those skilled in the art, but it should be clear that the present application is not limited to the scope of the specific embodiments, and for those skilled in the art, it is obvious that various changes are within the spirit and scope of the present application defined and determined by the appended claims, and all the inventions utilizing the concept of the present application are within the scope of protection.

[0016] EMBODIMENT Figure 1 is an exemplary flowchart of an Enteromorpha remote sensing image segmentation method based on channel boundary attention according to some embodiments of the present specification. As shown in Figure 1 , the flow includes the following steps. In some embodiments, the flow can be executed by a processor.

[0017] S1: Use the improved adaptive histogram equalization spectral enhancement algorithm to process the Enteromorpha remote sensing image to obtain a spectral enhancement optimized image.

[0018] The remote sensing image of Enteromorpha is an original multispectral Enteromorpha remote sensing image with problems of sea surface reflection interference, uneven brightness, and fuzzy target boundary.

[0019] The spectrum enhancement optimization image is an enhanced optimization image in an HSV color space.

[0020] In some embodiments, the processor can introduce a local variance adjustment term and a brightness balance coefficient on the basis of a traditional CLAHE for different sea area illumination and reflection characteristics; the local enhancement coefficient and the brightness compensation term are adaptively defined; when the local area texture is complex (such as the Enteromorpha edge or the floating belt), the variance is large, and the algorithm automatically increases the enhancement strength; and in the low-variance area (such as the sea water background), the enhancement strength is reduced to realize dynamic contrast control; the enhanced brightness channel and the original hue and saturation channels are recombined, and are restored to the RGB space through inverse transformation; finally, the image is normalized to match the neural network input requirements to obtain a spectrum enhancement optimization image. The spectrum enhancement optimization image can be expanded by using data enhancement means such as rotation, mirroring, color disturbance, and illumination change to expand the sample set, and the training set, the verification set, and the test set are divided according to a 7:2:1 ratio.

[0021] In some embodiments, the expression of the spectrum enhancement optimization image is: ; ; ; ; ; ; ; wherein, represents the spectrum enhancement optimization image, represents an image abscissa, represents an image ordinate, represents a final enhanced image obtained after the enhanced HSV is converted back to RGB, represents the minimum possible value of an image pixel value (for example, 0), represents the maximum possible value of an image pixel value, represents an HSV image recombined after brightness enhancement, including a channel, represents a hue channel (Hue), represents a saturation channel (Saturation), represents an enhanced brightness channel (Value), denotes a function to convert the HSV image to the RGB color space, denotes a local enhancement coefficient, denotes a local enhancement coefficient a baseline gain for ensuring that a basic brightness preservation capability is maintained when there is no significant texture or contrast variation (i.e. local variance close to zero), denotes an enhancement sensitivity adjustment coefficient, denotes a local standard deviation centered at (x, y) with window size w x w, denotes a local mean centered at (x, y) with window size w x w, denotes a small constant to prevent the denominator from being zero, denotes a brightness compensation term, denotes a brightness balancing factor, denotes a global brightness mean, denotes a brightness component, denotes an image after converting the original RGB image to the HSV color space, denotes converting the original multispectral remote sensing image to the HSV color space, denotes the original multispectral remote sensing image.

[0022] S2: based on the channel boundary attention, using the feature extraction module, the boundary enhancement module and the multi-scale feature fusion and decoding module, an image segmentation model is constructed.

[0023] The image segmentation model is a neural network model for segmenting Enteromorpha remote sensing images.

[0024] In some embodiments, as shown in Figure 2 the image segmentation model comprises: a feature extraction module (SegFormer) for performing feature extraction on the spectral enhancement optimization image using a channel boundary attention mechanism to obtain an original feature map; a channel boundary attention module for performing local interaction relationship analysis on the original feature map to obtain an input feature map; a boundary enhancement module for performing gradient mapping calculation on the input feature map to obtain a boundary response map; and a multi-scale feature fusion and decoding module for performing cascade and weighted fusion on the boundary response map based on a feature fusion path to obtain an Enteromorpha remote sensing image segmentation result.

[0025] The feature extraction module (SegFormer) is a semantic segmentation framework that combines a Transformer encoder and a lightweight MLP decoder. Its core idea is to leverage the global modeling capabilities of the Transformer to obtain long-distance dependency information. The original SegFormer responds to features from different channels with equal degree, making it difficult to effectively highlight the saliency of the Ulva prolifera region. This invention introduces a channel attention mechanism in the encoder part. By performing global average pooling on the feature map, it extracts the global descriptive vector of each channel, and then generates channel weight coefficients through two fully connected layers and a sigmoid activation function, thereby achieving adaptive recalibration of channel features.

[0026] In some embodiments, the expression for the original feature map is: ; ; ; in, Indicates the recalibrated first Output characteristics of each channel Indicates the first Attention weights for each channel, This represents the c-th channel feature of the input feature map. Indicates the first The global descriptor value obtained by global average pooling of the channel. Indicates the feature map height. Indicates the width of the feature map. Indicates the x-axis index of spatial location. Indicates the y-axis index of the spatial location. This represents the Sigmoid function. Represents ReLU. and This represents the weight matrix.

[0027] In this way, the response to the feature channels of *Ulva prolifera* can be enhanced, while the interference from background seawater and floating objects can be suppressed, thereby improving the overall feature expression capability.

[0028] In convolutional neural networks, shallow features primarily preserve low-level texture information of the image, such as edges, corners, and local gradients, while deep features gradually abstract into high-level semantics. If classification relies solely on deep semantic information, boundary details are easily overlooked.

[0029] In some embodiments, the Efficient Channel Attention module is different from the main semantic segmentation head that emphasizes the consistency of class prediction regions, as shown in Figure 3 The boundary branch explicitly learns to distinguish edge and non-edge pixels, thus enforcing structural constraints during the segmentation process.

[0030] In some embodiments, as shown in Figure 4 The boundary supervision branch is composed of a lightweight double-layer convolution classifier. First, a 3x3 convolution extracts edge-oriented features from the decoder feature map. Compared with fixed gradient operators (e.g., Sobel or Laplacian), the learnable convolution kernel can adaptively approximate a set of directional derivatives, thus capturing strong local intensity transitions corresponding to object boundaries. The output is normalized by batch normalization (BN) to stabilize the feature distribution, and then a ReLU activation function is used to suppress weak noise responses and preserve high-amplitude activations typical of edge regions. Subsequently, a 1x1 convolution acts as a pixel-wise linear classifier, integrating multiple edge evidences into a logit score for boundary probability. Then, the original logit map is upsampled by bilinear interpolation to match the original image resolution, ensuring alignment with the true boundary map.

[0031] The encoder shallow features of SegFormer (C1: 128x128x64, C2: 64x64x128) retain rich edge and texture information, but also mix noise and background interference. We introduce ECA (Efficient Channel Attention) to re-scale the channels of C1 and C2 without changing the backbone and decoding mainline structure, to highlight the discriminative channels related to "green algae boundaries", and feed the same re-scaled shallow features into: (i) the main decoding path (keeping the original decoder and semantic segmentation head unchanged); (ii) the edge supervision branch (Boundary Branch).

[0032] Before the shallow features enter the decoder and boundary branch, they are re-scaled by the ECA (Efficient Channel Attention) module. Unlike traditional SE attention, ECA does not introduce a fully connected layer or channel dimension reduction, but uses a lightweight one-dimensional convolution to directly capture the local interaction between channels, thus enhancing the discriminability of features while maintaining the efficiency of the model.

[0033] In some embodiments, the expression of the input feature map is: ; ; ; ; ; wherein, represents the input feature map, represents a three-dimensional feature tensor from the shallow layer of the encoder, H and W represent the spatial resolution, and C represents the number of channels, represents a channel description vector, represents a spatial position x-axis index, represents a spatial position y-axis index, represents the number of channels of the tensor, represents a channel weight vector, represents a Sigmoid activation function, represents a one-dimensional convolution with a length of , represents a channel-reweighted feature, represents omitted H, W, i.e., the spatial resolution, represents a weight coefficient of the th channel, represents a kernel length, represents an operation of rounding to the nearest odd integer, and represent hyperparameters. and are hyperparameters, and take odd represents rounding to the nearest odd integer. Taking the shallow feature in this paper as an example, the number of channels of is calculated to be = 6 k ≈ 5; the number of channels of is calculated to be = 7 k ≈ 5. Therefore, one-dimensional convolution of is used in both layers. This design significantly improves the response of edge and texture-related channels in the shallow feature while keeping the computational complexity extremely low, providing clearer representation for the subsequent boundary supervision branch and decoder.

[0034] In some embodiments, as Figure 5As shown, the channel boundary attention module optimizes the boundary true value using the binary cross-entropy loss with logits (BCEWithLogitsLoss) during the training process, and the gradient of the boundary branch flows back to the shallow features C1 and C2. Since the decoder fuses these shallow features, this gradient constraint indirectly enhances the edge sensitivity of the backbone branch. The boundary branch optimizes the shallow features so that the decoder can better retain high-frequency structural information when performing semantic segmentation, thereby improving the accuracy of boundary prediction. This fusion method not only enhances the edge sensitivity of semantic segmentation, but also improves the purity of boundary prediction, while avoiding the additional overhead caused by repeatedly stacking complex attention operators in the boundary branch.

[0035] Due to the obvious spectral transition zone at the junction between the edge of Enteromorpha and seawater, the prediction of the traditional semantic segmentation network in the edge region is discontinuous and fuzzy. The present application introduces a boundary feature enhancement module (BEM) to explicitly guide the model to learn edge features.

[0036] The boundary enhancement module (Boundary Enhancement Module) generates a boundary response map by calculating the gradient map of the feature map .

[0037] In some embodiments, when training the image segmentation model, a boundary constraint loss function (Boundary Loss) is designed for the boundary enhancement module:

[0038] wherein is the true boundary map, is the predicted boundary map, and Dice represents the similarity coefficient function. The loss is combined in the total loss with a weight λ:

[0039] Through joint optimization, the model enhances the boundary perception ability while maintaining the overall segmentation accuracy, making the Enteromorpha edge recognition clearer.

[0040] In some embodiments, the processor can utilize a multi-scale feature fusion and decoding module (Multi-scaleFusion) to cascade and weight fuse the shallow spatial details and deep semantic features through a top-down feature fusion path to obtain the Enteromorpha remote sensing image segmentation result.

[0041] The Enteromorpha remote sensing image segmentation result is an image or data table expression form obtained by inputting the remote sensing image into the image segmentation model for pixel-level classification, and used to represent the spatial distribution of Enteromorpha. The segmentation result can include but is not limited to: Enteromorpha region binary mask image, Enteromorpha probability map of each pixel, Enteromorpha boundary annotation image, Enteromorpha contour or polygon vector data extracted from the segmentation result, and Enteromorpha coverage area, spatial range or other distribution parameters calculated therefrom.

[0042] Using this fusion strategy can maintain local edge details while improving the model's detection ability for small-scale Enteromorpha targets.

[0043] S3: training the image segmentation model using the spectrum-enhanced optimized image to obtain a trained image segmentation model.

[0044] In some embodiments, the image segmentation model can be trained by a plurality of labeled training samples. For example, a plurality of labeled training samples can be input into an initial image segmentation model, a loss function can be constructed by the label and the result of the initial image segmentation model, and the parameters of the initial image segmentation model can be iteratively updated by gradient descent or other methods based on the loss function. When the preset condition is met, the model training is completed, and a trained image segmentation model is obtained. The preset condition can be loss function convergence, iteration number reaching a threshold, etc.

[0045] In some embodiments, the training sample can include a historical spectrum-enhanced optimized image. The label can correspond to the actual Enteromorpha remote sensing image segmentation. The label can be manually annotated.

[0046] In some embodiments, the expression of the loss function of the image segmentation model is: ; ; ; ; ; wherein, represents the loss function of the image segmentation model, represents the weight coefficient of the cross-entropy loss, represents the weight coefficient of the boundary loss, represents the weight coefficient of the morphological consistency loss, represents the cross-entropy loss, represents the boundary loss, represents the morphological consistency loss, represents the total number of pixels (all pixels in the training batch), y i represents a real label (0: background, 1: enteromorpha) of the i th pixel, pi represents a predicted probability (probability of belonging to the enteromorpha category) of the i th pixel, wi represents a boundary weight coefficient of the i th pixel, bi represents a predicted boundary value (0-1) of the i th pixel, yi represents a real boundary value (0 or 1) of the i th pixel, ki represents an enhancement coefficient, di represents a distance of a pixel point to a real boundary, Sigmoid represents a Sigmoid activation function, delta represents a change amount.

[0047] In some embodiments, in a test stage, the image segmentation model outputs an enteromorpha distribution probability map, and a detection result is generated through threshold segmentation. Experiments show that the improved loss function of the present application has a significant improvement in boundary continuity and small target detection accuracy, verifying the effectiveness and innovation of the method.

[0048] In some embodiments, the processor can select multi-temporal HY-1C, HY-1D and HY-1E images of the Yellow Sea and the coastal waters of Shandong as experimental areas, and compare them with the original SegFormer, U-Net and DeepLabv3+ models respectively. The experimental results show that the method of the present application improves the edge recognition accuracy (Boundary IoU) by about 5.8%, the small target detection accuracy by about 7.3%, and the overall detection F1-score by 4.6%, proving the stability and practicality of the method in complex marine environments.

[0049] S4: using the trained image segmentation model to infer the enteromorpha remote sensing image to obtain an enteromorpha remote sensing image segmentation result, and completing segmentation of the enteromorpha remote sensing image.

[0050] In some embodiments, the present application realizes accurate extraction and boundary recovery of Enteromorpha by embedding an efficient channel attention (ECA) module in the shallow layer of the encoder and adding a boundary supervision branch in the decoder. This method can adaptively strengthen key channel features, enhance the detection ability of small or scattered Enteromorpha clusters, and explicitly learn edge information to improve the Enteromorpha boundary blur and breakage problem. Under the premise of not significantly increasing the computational complexity, the present method improves the global semantic consistency and local boundary sensitivity. The experimental results show that compared with U-Net, HRNet, PSPNet and the original SegFormer network, the mIoU, F1, Precision and Recall of the present method are increased by 14.34%, 10.53%, 5.16% and 13.21% respectively, which can more comprehensively and accurately extract the Enteromorpha coverage area, and has higher robustness and application value in complex marine background.

[0051] To realize high-precision extraction of Enteromorpha in remote sensing images, the present application provides an ECAB-SegFormer model. The input remote sensing image is first extracted by the SegFormer encoder to obtain multi-scale features. The shallow layer features C1 and C2 are re-labeled by the efficient channel attention (ECA) module to enhance the feature response related to the Enteromorpha edge. Then, the re-labeled features are input into the decoder and the boundary supervision branch respectively for semantic segmentation and boundary information learning, and finally the Enteromorpha segmentation mask is generated by fusion. This process can maintain global semantic consistency and enhance edge sensitivity, thereby improving the recognition ability of small or scattered Enteromorpha regions.

[0052] Figure 6 A typical remote sensing image Enteromorpha segmentation result is shown in the figure, in which green represents the correctly identified Enteromorpha region, red represents the boundary region, and blue represents the background region. As can be seen from the figure, the present embodiment can identify small or scattered Enteromorpha clusters completely and effectively suppress misidentification caused by the marine background, realizing high-precision Enteromorpha segmentation.

[0053] In some embodiments, the remote sensing data used by the present application comes from the HY-1 series of ocean color satellites independently developed by China, mainly including HY-1C, HY-1D and HY-1E. Among them, HY-1C satellite was launched in 2018, carrying Coastal Zone Imager (CZI, resolution about 50 m) and Medium Resolution Ocean Observation Imaging Spectrometer (PMRIS, resolution about 1 km), which can realize continuous observation of large-scale water color parameters; HY-1D satellite was launched in 2020, and its sensor performance is optimized in terms of spectral resolution and radiation accuracy; HY-1E is a subsequent upgraded model, which further improves the spatial and temporal resolution and quantitative ability. The data collection range of these satellites covers the key sea areas of China's near sea, especially the Yellow Sea, providing reliable data support for the long-term and continuous monitoring of Enteromorpha green tide. Compared with the commonly used remote sensing satellites in the world (such as Landsat series, Sentinel-2, MODIS, etc.), although the HY-1 series of satellites are mainly designed for ocean color observation, they are not inferior in terms of spatial resolution, coverage range and water color information sensitivity, and can provide more targeted monitoring capability for marine remote sensing scientific research, especially suitable for the monitoring and research of marine emergencies such as Enteromorpha green tide.

[0054] In some embodiments, as shown in Figure 7 The dataset consists of original remote sensing images and corresponding Enteromorpha mask label vector files, hereinafter referred to as "HYU dataset". The original data is in GeoTIFF format, which has rich multi-channel information that can fully represent the spectral characteristics of water bodies. The true boundary of the Enteromorpha coverage area is represented by a shp vector file generated by manually correcting the vector results obtained by inversion interpretation, and rasterized into a mask annotation map aligned with the image.

[0055] In some embodiments, considering that deep learning models are usually based on RGB image input, in order to reduce the interference of marine noise and enhance the contrast of Enteromorpha, the present study selects 4, 2, 1 three channel bands from the original channels, and maps them to the R, G, B three channels of the PNG image, which not only guarantees the image feature expression ability, but also is compatible with common segmentation network structure. The dataset division is shown in Table 1.

[0056] Table 1 Dataset division table

[0057] In some embodiments, all experiments of the present application are performed on a server equipped with an NVIDIA GeForce RTX 4090 graphics card, the operating system is Ubuntu 20.04, the deep learning framework is PyTorch 1.13, and the CUDA version is 11.7. Data preprocessing and visualization operations use Python 3.8 and combine commonly used scientific computing libraries such as NumPy, OpenCV, and Matplotlib. Among them, the training configuration of the HYU dataset is shown in Table 2.

[0058] Table 2 Training configuration table of HYU dataset

[0059] In some embodiments, to verify the performance of the ECAB-SegFormer model proposed in the present application in the segmentation of Enteromorpha remote sensing images, the present embodiment performs experiments on the HYU dataset, which is divided into two parts: ablation experiments and comparative experiments with other common segmentation models. The input remote sensing image is extracted by the SegFormer encoder to obtain multi-scale features, among which the shallow features are re-labeled by the efficient channel attention (ECA) module, and then input into the decoder and the boundary supervision branch for semantic segmentation and boundary learning, respectively, to finally generate the Enteromorpha segmentation mask.

[0060] The original network model and the network model using different mechanisms are tested on the HYU dataset, and the evaluation indicators are compared as shown in Table 3.

[0061] Table 3 Comparison table of experimental accuracy of improved network and original network

[0062] On the HYU dataset, from the comparison results of the segmentation performance of the baseline network and the improved network variants, it can be seen that whether only the boundary supervision branch is introduced or only the ECA channel attention module is embedded, the segmentation accuracy is improved to varying degrees. Specifically, the boundary supervision branch significantly improves the recall rate (from 69.84% to 75.67%), indicating that it can effectively enhance the boundary positioning ability and restore the fine-grained structural information; among them, the ECA module is embedded in the C1-C4 feature layer, and from the data, it can be seen that it improves the precision rate more significantly (from 91.75% to 92.84%), indicating that the channel attention mechanism can strengthen feature selection and suppress irrelevant responses.

[0063] When combined in the improved network proposed in this study, each indicator is optimal, with mIoU, F1, precision, and recall increasing by 6.84%, 4.76%, 2.37%, and 6.22% respectively compared with the baseline. These results show that the integration of boundary perception learning and channel attention mechanism not only complements the encoder-decoder backbone network, but also improves the global semantic consistency and local boundary accuracy at the same time, thereby improving the segmentation accuracy of the entire model.

[0064] The changes in segmentation accuracy of the original and improved segformer network models during training are shown in Figure 8 , Figure 9 During training on the HYU dataset, the segmentation accuracy of the original SegFormer network model gradually increased with the number of iterations, with rapid increases in the early stage, gradual stabilization in the middle and late stages, but overall there were still fluctuations, especially the boundary-related mIoU converged slowly, indicating that the model had deficiencies in identifying green algae boundaries. In contrast, after adding the boundary supervision branch and ECA channel attention module to the improved network in this study, each indicator was significantly improved, with faster convergence speed in the middle of training, and the mIoU curve was significantly higher than that of the original network, indicating that the boundary supervision branch effectively enhanced the learning ability of boundary pixels, while Precision and F1 were also improved, reflecting that the channel attention mechanism strengthened feature selection and suppressed irrelevant responses. The final convergence value showed that the mIoU of the improved network reached 72.61%, and the Precision was 94.12%, significantly better than the original network. This result shows that the integration of boundary perception learning and channel attention mechanism not only improves the overall segmentation accuracy, but also enhances the stability and convergence efficiency of the model training process, making the boundary recognition of green algae remote sensing images clearer and the details more complete.

[0065] To further explore the role of the ECA module at different network levels, this study designed multiple groups of ablation experiments (A1-D2) with embedded positions, where the shallow layers correspond to the first two layers [0,1] of the encoder, and the high layers correspond to the last two layers [2,3], combined with the boundary supervision branch. The evaluation indicator comparison is shown in Table 4.

[0066] Table 4 Ablation experiments of different ECA activation layers

[0067] The experimental results are shown in the table: without using boundary supervision, the shallow embedding ECA (A1) mainly improves the recall rate of the model (69.84%→74.14%), which verifies that the shallow channel re-labeling can strengthen the response of edge and texture details, but due to the limited inhibition of false detection ability, the overall precision rate is slightly improved (91.75%→93.01%). In contrast, the high-level embedding ECA (A2) brings more significant precision rate improvement (94.49%), which proves that the high-level semantic attention has advantages in reducing background interference and enhancing class discrimination, but the improvement of recall rate is relatively limited (72.83%).

[0068] After introducing boundary supervision, shallow activation ECA+boundary supervision (B1) significantly improves the recall rate (75.67%), making the model better recover the boundary and detail structure; while high-level+boundary supervision (B2) takes into account the higher precision rate (93.90%) and better overall consistency (mIoU=71.88%, F1=83.64%).

[0069] Further, when the ECA module is activated at a specific feature layer and combined with boundary supervision (D2), the model performance is further improved. Shallow fusion (D1) maintains an advantage in recall rate (74.68%), but the precision rate is relatively low (92.34%); in contrast, high-level fusion (D2, This study) achieves the best mIoU (72.61%) and F1 (84.14%), while the Recall also reaches the highest level (76.06%). This result shows that under the constraint of boundary supervision, activating ECA at a specific layer can form a complement between global semantic modeling and local boundary detail recovery, significantly improving the overall segmentation performance.

[0070] In summary, the ECA module is more beneficial to edge detail enhancement in the shallow layer, while it is better at improving the precision rate in the high layer; its best effect appears when combined with boundary supervision, especially specific feature layer+boundary supervision fusion (D2), which can balance between global consistency and boundary sensitivity, providing a reliable strategy for subsequent model optimization.

[0071] In this experiment, the label information of the original data set is visualized on the image and compared with the experimental results. Several classical and mainstream deep learning semantic segmentation algorithms and common improvements of SegFormer are used to verify the effectiveness of the improved algorithm in this paper. The high-resolution network (HRnet), PSPNet, U-Net, and the original SegFormer network and the SegFormer-ASPP network with the introduction of the Atrous Spatial Pyramid Pooling (ASPP) module are selected for comparison with the algorithm in this paper, and the comparison results are shown in Table 5.

[0072] Table 5 Precision comparison table of different methods

[0073] The overall experimental precision of the improved method on the HYU dataset is better than that of the traditional convolutional network (HRNet, PSPNet, U-Net) and the original SegFormer model (Table 5). Specifically, in this experiment, the mIoU, F1, Precision, and Recall of the baseline SegFormer are 65.81%, 79.38%, 91.75%, and 69.84%, respectively, while the corresponding indicators of the improved method in this paper are 72.61%, 84.14%, 94.12%, and 76.06%, respectively, and the overall performance is comprehensively improved. Compared with the contrast network, the improved method has the highest improvement in mIoU, up to 14.34% (compared with U-Net), and the highest improvement in Precision, up to 5.16% (compared with PSPNet), showing the significant advantage of the improved strategy in balancing global semantics and boundary details.

[0074] Under different network tests, it can be found that U-Net and PSPNet have limitations in boundary detail recovery and large-scale background suppression, with relatively low F1 and Recall, while HRNet captures boundary details well, but the overall mIoU and F1 level is still lower than SegFormer. The improved method effectively enhances the response of edge-sensitive channels by embedding ECA modules in the shallow layers of the SegFormer decoder input layer and combining boundary supervision, while maintaining global semantic consistency, thereby significantly improving Precision and Recall.

[0075] Compared with SegFormer-ASPP (SegF-ASPP), its performance on the HYU dataset is slightly worse than that of the original segformer model, especially with a decrease of 2.23% and 2.49% in Recall and mIoU, respectively. This result verifies that after introducing the multi-scale context information of ASPP, the semantic understanding ability of the SegFormer model is enhanced, which is reflected in the improvement of Precision, but in small target segmentation and noise suppression, it is actually decreased, further proving that for identifying Enteromorpha with no fixed morphological features, optimizing shallow features and explicitly learning boundaries are more important.

[0076] After network improvement, the experimental detection precision is improved by an order of magnitude. Specifically, it is more accurate in edge detection of complex Enteromorpha, performs better in small target detection, and has stronger robustness in the face of complex marine noise, verifying the effectiveness of the improved algorithm part.

[0077] The model complexity and efficiency of HRnet, PSPNet, UNet, and the original SegFormer network and SegFormer-ASPP network are compared with the algorithm of the present application, and the results are shown in Table 6.

[0078] Table 6 Network complexity analysis

[0079] In the experiment, the input size is set to 512x512 when calculating GFLOPs, and the training time is the time required for one HYU iteration. As can be seen from Table 6, the mIoU of the algorithm of the present application has been greatly improved with a slight increase in GFLOPs, parameters and training time, and has a very high application value at a relatively low cost.

[0080] In some embodiments, as shown in Figure 10 the experimental prediction results of HRnet, PSPNet, U-Net, the original SegFormer network, SegFormer-ASPP network and the algorithm of the present application are shown. The performance of each algorithm is evaluated by visualizing the detection results. The pixels marked in green represent correct prediction, the pixels marked in blue represent missed detection, and the pixels marked in red represent false detection. As can be seen from Figure 9 when testing on the same image, the detection results of the improved network are better than those of other networks in the following aspects.

[0081] (1) The boundary detection is more fine and the edge transition is smooth: the edge integrity of the green algae coverage area is significantly improved. Figure 9 As can be seen from the first and third rows, HRNet, PSPNet and the original SegFormer have fuzzy or broken phenomenon in the boundary description of independent green algae clumps, especially at the complex water body and green algae junction. U-Net has improved in boundary positioning, but the edge transition is still not smooth enough. The improved network after introducing the boundary supervision branch and ECA can accurately capture the green algae contour, the edge continuity and transition are natural and reasonable, and effectively distinguish green algae from background water, so that the boundary detection effect is obviously better than other comparison methods.

[0082] (2) Enhance the detection ability of small or mixed green algae clumps: the improved network enhances the edge sensitive channel through the shallow ECA, and the boundary branch explicitly learns the edge information, effectively suppressing the missed detection and false detection caused by water reflection and similar green algae color. Figure 9The second row shows that when the small Enteromorpha clumps are similar to the surrounding background color and the Enteromorpha area is much smaller than other areas, UNet and the original SegFormer and SegFormer-ASPP are prone to miss these areas, while the improved network can completely detect the distribution range and edges of the small Enteromorpha clumps, and achieve more comprehensive Enteromorpha coverage recognition.

[0083] (3) The details are fully retained, and the boundaries are clearly distinguished from the background: Figure 9 The second and fourth rows show that in the multi-clump Enteromorpha aggregation area, other comparative networks have phenomena such as edge blurring, local fracture or excessive smoothing, while the improved network recovers more edge details by using the fusion of high-level semantic enhancement and boundary supervision, the boundary contour is more consistent with the artificial annotation, the distribution range difference is significantly reduced, and the overall detection accuracy is improved.

[0084] (4) There are still challenges under environmental interference: although the improved network has improved in boundary sensitivity and global semantic discrimination, in areas with significant light reflection, water ripples or surrounding plankton interference, some Enteromorpha edges may still have slight missed detection or false detection, indicating that the feature discrimination ability still needs to be further optimized in complex spectral environments.

[0085] In some embodiments of the present specification, a channel boundary attention-based Enteromorpha remote sensing image segmentation method is provided, which introduces a channel attention mechanism to enhance feature extraction capability, and combines a boundary feature enhancement module to improve target edge recognition accuracy. To verify the effectiveness of the method, multi-temporal remote sensing images of typical sea areas in the Yellow Sea were selected as experimental data, and were compared with traditional SegFormer models and other deep segmentation algorithms. The experimental results show that the method has achieved significant improvement in Enteromorpha detection accuracy, boundary detail retention and model robustness, providing a reliable technical means for the automated monitoring of large-scale Enteromorpha disasters. (1) By learning boundary information from shallow features, accurate detection of small Enteromorpha targets and improvement of boundary integrity are achieved. (2) By adaptively re-labeling key feature channels to enhance channel responses while suppressing background noise, Enteromorpha segmentation accuracy is improved. (3) The combination of boundary supervision branch and ECA module enables shallow features to strengthen edge details while maintaining global semantic consistency, achieving collaborative optimization of the decoder and the boundary branch. (4) A complete Enteromorpha remote sensing image segmentation framework is constructed, including multi-scale feature extraction, ECA channel re-labeling, boundary supervision and decoding fusion process, which can adapt to various satellite remote sensing data in the Yellow Sea. (5) Through lightweight design, the patent achieves comprehensive improvement in mIoU, F1, Precision and Recall indicators without significantly increasing computational complexity, effectively balancing accuracy and efficiency.

Claims

1. A method for segmenting *Ulva prolifera* remote sensing images based on channel boundary attention, characterized in that, include: S1: The remote sensing image of Ulva prolifera is processed using an improved adaptive histogram equalization spectral enhancement algorithm to obtain a spectrally enhanced and optimized image; S2: Based on channel boundary attention, an image segmentation model is constructed using a feature extraction module, a boundary enhancement module, and a multi-scale feature fusion and decoding module; S3: Use spectral enhancement to optimize the image and train the image segmentation model to obtain a trained image segmentation model; S4: Using the trained image segmentation model, inference is performed on the Ulva prolifera remote sensing image to obtain the Ulva prolifera remote sensing image segmentation result, thus completing the segmentation of the Ulva prolifera remote sensing image.

2. The method for segmenting *Ulva prolifera* remote sensing images based on channel boundary attention according to claim 1, characterized in that, The expression for the spectrally enhanced optimized image is: ； ； ； ； ； ； ； in, This indicates a spectral enhancement optimized image. Represents the x-coordinate of the image. Represents the vertical coordinate of the image. This indicates the final enhanced image obtained after converting the enhanced HSV back to RGB. This represents the minimum possible value of an image pixel (e.g., 0). This indicates the maximum possible value of an image pixel. This represents a reconstructed HSV image after brightness enhancement, containing... aisle, Indicates the hue channel. This represents the saturation channel. This represents the enhanced luminance channel (Value). This represents a function that converts an HSV image to the RGB color space. Represents the local enhancement coefficient. Represents the local enhancement coefficient The baseline gain is used to ensure that basic luminance retention is maintained even when there are no significant changes in texture or contrast (i.e., local variance is close to zero). This indicates the sensitivity adjustment coefficient. Let represent the local standard deviation centered at (x,y) with a window size of w×w. Let represent the local mean centered at (x, y) with a window size of w×w. This represents a small constant used to prevent the denominator from being zero. Indicates the brightness compensation item. Indicates the brightness balance factor. This represents the global average brightness. Represents the luminance component. This indicates the image after converting the original RGB image to the HSV color space. This indicates the original multispectral remote sensing image. Convert to HSV color space This represents the original multispectral remote sensing image.

3. The method for segmenting *Ulva prolifera* remote sensing images based on channel boundary attention according to claim 1, characterized in that, The image segmentation model includes: The feature extraction module is used to extract features from the spectral enhancement and optimization image using the channel boundary attention mechanism to obtain the original feature map; The channel boundary attention module is used to perform local interaction relationship analysis on the original feature map to obtain the input feature map; The boundary enhancement module is used to perform gradient mapping calculations on the input feature map to obtain the boundary response map; The multi-scale feature fusion and decoding module is used to perform cascaded and weighted fusion of boundary response maps based on feature fusion paths to obtain the segmentation results of the Ulva prolifera remote sensing image.

4. The method for segmenting *Ulva prolifera* remote sensing images based on channel boundary attention according to claim 3, characterized in that, The expression for the original feature map is: ； ； ； in, Indicates the recalibrated first Output characteristics of each channel Indicates the first Attention weights for each channel, This represents the c-th channel feature of the input feature map. Indicates the first The global descriptor value obtained by global average pooling of the channel. Indicates the feature map height. Indicates the width of the feature map. Indicates the x-axis index of spatial location. Indicates the y-axis index of the spatial location. This represents the Sigmoid function. Represents ReLU. and This represents the weight matrix.

5. The method for segmenting *Ulva prolifera* remote sensing images based on channel boundary attention according to claim 3, characterized in that, The expression for the input feature map is: ； ； ； ； ； in, Indicates the input feature map, Let H and W represent a 3D feature tensor from a shallow layer of the encoder, where H and W represent the spatial resolution and C represents the number of channels. Represents the channel description vector. Indicates the x-axis index of spatial location. Indicates the y-axis index of the spatial location. This indicates the number of channels in the tensor. Represents the channel weight vector. This represents the Sigmoid activation function. Indicates length is One-dimensional convolution, This indicates the characteristics after channel recalibration. The omitted H and W represent spatial resolution. Indicates the first The weighting coefficients of each channel Indicates the length of the core. This represents the operation of rounding down to the nearest odd integer. and This represents hyperparameters.

6. The method for segmenting *Ulva prolifera* remote sensing images based on channel boundary attention according to claim 1, characterized in that, The expression for the loss function of the image segmentation model is: ； ； ； ； ； in, This represents the loss function of the image segmentation model. The weights representing the cross-entropy loss are... The weighting coefficients representing the boundary loss. The weighting coefficients representing the loss of morphological consistency. Represents cross-entropy loss, Indicates boundary loss. This indicates a loss of morphological consistency. This represents the total number of pixels (all pixels in the training batch). This represents the true label of the i-th pixel (0: background, 1: seaweed). This represents the predicted probability of the i-th pixel (the probability of it belonging to the seaweed category). This represents the boundary weight coefficient of the i-th pixel. This represents the predicted boundary value (0–1) for the i-th pixel. This represents the true boundary value (0 or 1) of the i-th pixel. Indicates the enhancement coefficient. This represents the distance from a pixel to the actual boundary. This represents the Sigmoid activation function. Indicates the amount of change.

Images