Integrated Method and System for Broadband 3D Sample Generation and Identification of Dominant Species in Red Tide

Abstract

This invention belongs to the field of satellite remote sensing and red tide detection technology, and discloses an integrated method and system for generating and identifying broadband 3D samples of dominant red tide species. The method constructs an integrated model for generating broadband 3D red tide samples and identifying dominant species based on deep learning. It utilizes a red tide image generation module based on spectral feature transfer to obtain expanded samples of different dominant red tide species. A red tide dominant species identification module is constructed by integrating multi-scale feature extraction and cross-spatial attention mechanisms. The module is trained using a student-teacher semi-supervised learning model. Experimental results show that the model has good adaptability and can effectively identify three dominant red tide species: *Noctiluca scintillans*, *Noctiluca chloroticus*, and *Hacochloa erythrophora*, with an overall identification accuracy of 94.48%, which is 3.52%–8.66% higher than other comparative methods, providing a basis for red tide prevention and management.

Description

Technical Field

[0001] This invention belongs to the field of satellite remote sensing and red tide detection technology, and in particular relates to an integrated method and system for generating and identifying broadband three-dimensional samples of dominant red tide species. Background Technology

[0002] In recent years, small-scale red tides have frequently occurred nearshore. The spatial resolution of water color satellites (below 250m) is insufficient for monitoring small-scale, dispersed red tides, leading to the increasing use of medium-to-high spatial resolution satellites (higher than 50m) for red tide detection. However, medium-to-high spatial resolution satellites have low spectral resolution and poor distinguishability of ground features. Therefore, medium-to-high spatial resolution satellites are mostly used for detecting single algal species in single regions (such as *Noctiluca scintillans* in the coastal waters of China, *Margalefidinium polykrikoides* in the Korean Peninsula, and *Lingulodinium polydra* in the coastal waters of South Africa and California), with limited research on the identification of different algal species. Currently, some studies have used high-resolution broadband satellites such as Landsat 8 OLI, GF-1 WFV, and Sentinel 2 A / B to distinguish algal species when *Ulva prolifera* and *Sargassum fusiforme* coexist. However, these two are different macroalgae; the reflectance peak of *Ulva prolifera* is approximately 550nm, while that of *Sargassum fusiforme* is around 600nm, showing significant spectral differences. As microalgae, red tides exhibit small spectral differences among different species. For example, the fluorescence reflectance peaks of *Noctiluca scintillans*, *Prorocentrum dinoflagellates* from a certain region of the East China Sea, and *Aureobasidium globosum* are all located around 690 nm, with differences of less than 5 nm. However, due to limitations in sensor configuration, the bandwidth of high-resolution wide-band satellites is often greater than 50 nm, making it impossible to accurately identify different dominant red tide species based solely on spectral differences. Gernez et al. (2023) explored the application of Sentinel2 MSI in identifying dominant red tide species, determining the spectral shapes of six red tide algal species. However, this paper used seven bands of the MSI, especially two red-edge bands, which are not available on most wide-band satellites. Therefore, achieving accurate identification of dominant red tide species using only four bands (red, green, blue, and near-infrared) remains a critical problem that urgently needs to be solved.

[0003] Deep learning methods possess powerful big data mining and feature extraction capabilities, providing new insights for identifying dominant red tide species. Zhu et al. (2019) proposed a hyperspectral remote sensing identification model for phytoplankton based on deep neural networks, which successfully predicted the dominant red tide species and their spatial distribution in the Yangtze River Estuary when transferred to HICO data. Kim et al. (2019) proposed an automatic pixel-level detection method based on a deep CNN model (U-Net) to identify three red tide groups from GOCI images of the Korean Peninsula coast. Shin et al. (2022) constructed convolutional neural network models of different depths and used Sentinel-3 OLCI data to identify and detect toxic and non-toxic red tides. However, these models primarily model red tide algae species within the same region, and their generalization ability needs improvement. Furthermore, considering the differences between high-resolution broadband satellites and ocean color satellite sensors, these models cannot be applied to low-spectral-resolution high-resolution satellite data.

[0004] Furthermore, training deep learning models requires a large number of red tide samples of different dominant species. However, due to the frequency and distribution of red tide outbreaks, the distribution of samples of different dominant species is uneven, and sample interpretation relies on human-computer interaction, resulting in high production costs and a shortage of high-quality samples, making model training difficult. Therefore, it is necessary to develop a new deep learning-based method for identifying dominant red tide species to address the problems of uneven distribution of samples of different dominant red tide species and the identification of dominant species under low spectral resolution.

[0005] Based on the above analysis, the existing technologies have the following problems and defects: (1) High-resolution wide-band satellites have low spectral resolution and few bands. Currently, traditional red tide dominant species identification methods cannot accurately identify red tide dominant species under the condition that only four bands (red, green, blue, and near-infrared) are available. (2) Currently, most deep learning-based red tide dominant species identification models are designed for red tide algae species in the same region. The generalization ability of the models is weak, and they cannot be applied to high-resolution satellite data with low spectral resolution. (3) Currently, training of deep learning-based red tide dominant species identification models requires a large number of red tide samples of different dominant species. However, due to the influence of the frequency and distribution of red tide outbreaks, the distribution of red tide samples of different dominant species is uneven. Furthermore, sample interpretation relies on human-computer interaction, resulting in high production costs and a shortage of high-quality samples, making it difficult to train the model. Summary of the Invention

[0006] To overcome the problems existing in related technologies, the present invention discloses an integrated method and system for generating and identifying broadband three-dimensional samples of dominant red tide species. The technical solution is as follows: This invention is implemented as follows: an integrated method for generating and identifying broadband three-dimensional samples of dominant red tide species, comprising the following steps: S1, Construct an integrated model for generating three-dimensional samples of wideband red tides and identifying dominant species based on deep learning; S2, the red tide image generation module based on spectral feature transfer obtains red tide expansion samples of different dominant species, and a red tide dominant species identification module is constructed by integrating multi-scale feature extraction and cross-spatial attention mechanism; S3 trains the red tide dominant species identification module through a semi-supervised learning mode using student and teacher models. It then uses generated red tide samples and real red tide samples to achieve cross-temporal and spatial identification of red tide dominant species under wide-band conditions.

[0007] Another objective of this invention is to provide an integrated system for generating and identifying broadband three-dimensional samples of dominant red tide species. This system is used to regulate the aforementioned integrated method for generating and identifying broadband three-dimensional samples of dominant red tide species. The system includes: A red tide image generation module based on spectral feature transfer is used to generate three-dimensional red tide samples; Based on the dominant species identification module of cross-spatial feature fusion, the red tide image generation module based on spectral feature transfer obtains red tide expansion samples of different dominant species. The red tide dominant species identification module is constructed by integrating multi-scale feature extraction and cross-spatial attention mechanism. The red tide dominant species identification module is trained by semi-supervised learning mode of student model and teacher model. Using the generated red tide samples and real red tide samples, cross-temporal and spatial identification of red tide dominant species under wide band conditions is realized.

[0008] Combining all the above technical solutions, the beneficial effects of this invention are as follows: First, red tides, as a global marine ecological disaster, seriously threaten the marine ecological balance. Red tide outbreaks are characterized by diverse dominant species, with different dominant species having varying impacts. Accurate identification of toxic red tides is particularly crucial for red tide disaster prevention and control. Satellite remote sensing plays a vital role in red tide detection, especially with broadband satellites offering high spatial resolution in frequent small-scale nearshore red tides. However, broadband satellites suffer from low spectral resolution and limited bands, making the identification of dominant red tide species extremely challenging. Although deep learning methods possess powerful feature extraction capabilities, their application is limited by the uneven distribution of samples among various dominant red tide species. To address these issues, this invention utilizes the data-driven capabilities of deep learning to propose an integrated model for three-dimensional red tide sample generation and dominant species identification. This model comprises two modules: a red tide sample generation module based on spectral feature transfer, which generates high-quality three-dimensional red tide samples, resolving the problem of uneven distribution of dominant species samples. This invention also developed a dominant species identification module based on cross-spatial feature fusion. This module enhances interspecific separability by jointly extracting spectral and spatial structural features. Furthermore, a student-teacher semi-supervised learning model is employed to improve the generalization ability of the identification model. Experimental results show that the model can effectively identify three dominant red tide species: *Noctiluca scintillans*, *Noctiluca scintillans*, and *Hacochloa hemlock*, with an overall identification accuracy of 94.48%, which is 3.52%-8.66% higher than other comparative methods. The model exhibits good adaptability, applicable to the identification of dominant red tide species in different regions and to various broadband satellites with red, green, blue, and near-infrared bands. In addition, the red tide 3D sample generation module demonstrates good spatial and spectral fidelity, enabling dominant species identification under conditions of few or no samples. This proves the feasibility of using low-spectral-resolution broadband satellite imagery to identify red tide species, providing a basis for red tide prevention and management.

[0009] Secondly, this invention develops an integrated model (RT-SGINet) for generating broadband three-dimensional red tide samples and identifying dominant species, using HY-1C / D CZI data as an example. This model specifically includes two parts: a red tide image generation module based on spectral feature transfer, used to solve the problem of insufficient and imbalanced samples in red tide dominant species identification; and a red tide dominant species identification module based on cross-spatial feature fusion, combined with cross-domain attention and multi-scale feature extraction modules, to overcome the limitation of small spectral differences of broadband dominant species on the identification task. Finally, based on a small number of real red tide images and labels, expanded samples of red tides with different dominant species are obtained, and using the generated red tide samples and the real red tide samples produced, remote sensing identification of red tide dominant species in different complex marine scenarios is achieved through semi-supervised methods.

[0010] Third, after the technology is industrialized, it will enable 24-hour automated monitoring, drive the development of the industrial chain, and provide strong protection for marine ecological protection and public health and safety. This invention proposes for the first time an integrated model for generating broadband three-dimensional samples of red tides and identifying dominant species. Through the three-dimensional sample generation technology of spectral feature transfer and the deep fusion mechanism of cross-spatial feature fusion and semi-supervised learning, it realizes "intelligent interpretation of small sample marine remote sensing" and "wideband satellite identification of dominant species of red tides".

[0011] Fourth, this invention addresses the difficulties in identifying dominant species in red tides due to the low spectral resolution of broadband satellites, as well as the imbalance in sample distribution among different dominant species. It proposes an integrated model for generating three-dimensional red tide samples and identifying dominant species, achieving high-precision identification of three dominant species: *Noctiluca scintillans*, *Noctiluca scintillans*, and *Hacochloa erythrophora*, providing an effective solution to related technical problems. This invention overcomes the technical bias in the field of marine remote sensing regarding the difficulty of accurately identifying dominant species using broadband satellites, verifies the application prospects of broadband satellites in dominant species identification, and provides new ideas and references for the development of marine monitoring technology. Attached Figure Description

[0012] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure; Figure 1 This is a flowchart of the integrated method for generating and identifying three-dimensional samples of dominant red tide species based on high-resolution wideband satellites provided in this embodiment of the invention; Figure 2 This is a schematic diagram of red tide image samples of different dominant species of HY-1C / D CZI provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the HY-1C / D CZI spectral library provided in the embodiments of the present invention; (a) is a water body, (b) is *Noctiluca scintillans*, (c) is *Noctiluca scintillans*, and (d) is *Hacochloa erythrophora*. Figure 4 This is a diagram of the integrated model framework for generating broadband three-dimensional samples of red tides and identifying dominant species provided in this embodiment of the invention. Figure 5 This is a network architecture diagram of the red tide image generation model based on spectral feature transfer provided in this embodiment of the invention; Figure 6 This is a network architecture diagram of the red tide dominant species identification model based on cross-spatial feature fusion provided in an embodiment of the present invention; Figure 7This is a schematic diagram of the red tide sample generation results provided in the embodiments of the present invention; wherein, rows 1-2 are Noctiluca scintillans redissima, rows 3-4 are Noctiluca scintillans greenissima, and rows 5-6 are Hemlockella helioscopia; (a) is a label image, (b) is the pix2pix generation result, (c) is the RT-GAN generation result, (d) is the RTS-GAN generation result, and (e) is the HY-1C / D image; Figure 8 These are radiance spectral curves of different land features under different methods provided in the embodiments of the present invention; (a) is Noctiluca scintillans redissus, (b) is Noctiluca scintillans greenissus, (c) is Hemoglobina haca, (d) is water, (e) is cloud, and (f) is land. Figure 9 This is a schematic diagram of the identification results of the dominant red tide species provided in the embodiments of the present invention; wherein, (a) is a HY-1C / D image, (b) is a verification image, (c) is the identification result of RT-SGINet, (d) is the identification result of Deeplab V3+, and (e) is the identification result of ResU-Net; Figure 10 This is a schematic diagram of the identification results of dominant red tide species in different regions provided by the embodiments of the present invention; wherein, (a) and (b) are a certain province, (c) and (d) are a certain region of the Bohai Sea, (e) and (f) are a certain region of the Arabian Sea, (g) and (h) are a certain coast of Thailand, and (i) and (j) are a certain region of the East China Sea; Figure 11 This is a schematic diagram of the results of red tide dominant species identified by different satellites according to the embodiments of the present invention; wherein, (a) is a GF1 WFV image, (b) is a GF1 WFV result, (c) is a Sentinel-2B MSI image, and (d) is a Sentinel-2B MSI result. Detailed Implementation

[0013] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0014] The innovations of this invention are as follows: (1) This invention innovatively constructs an integrated model for generating three-dimensional red tide samples and identifying dominant species, achieving high-precision identification of dominant red tide species under low spectral resolution conditions of wide-band satellites. (2) This invention achieves high-precision identification (94.48%) of Noctiluca scintillans, Noctiluca scintillans, and Hemangiospermae under low spectral resolution and few band conditions for the first time, which is 3.52%-8.66% higher than existing methods, and has good cross-regional and cross-platform adaptability. (3) This invention constructs a three-dimensional red tide sample generation module based on spectral feature transfer, breaking through the limitations of traditional two-dimensional image generation, and effectively alleviating the problem of uneven distribution of dominant species samples through high-fidelity synthetic data. (4) This invention proposes an integrated model for generating three-dimensional red tide samples and identifying dominant species to address the difficulty in identifying dominant red tide species caused by low spectral resolution of wide-band satellites and the problem of unbalanced sample distribution among different dominant species. The model includes two modules: a red tide sample generation module based on spectral feature transfer is constructed to generate high-quality three-dimensional red tide samples and solve the problem of uneven distribution of dominant species samples. A dominant species identification module based on cross-spatial feature fusion was developed. This module enhances interspecific separability by jointly extracting spectral and spatial structural features. Furthermore, a student-teacher semi-supervised learning model is employed to improve the generalization ability of the identification model. Experimental results show that the model can effectively identify three dominant red tide species: *Noctiluca scintillans*, *Noctiluca scintillans*, and *Hacochloa hemlock*, with an overall identification accuracy of 94.48%, which is 3.52%-8.66% higher than other comparative methods. The model exhibits good adaptability, applicable to the identification of dominant red tide species in different regions and to various broadband satellites with red, green, blue, and near-infrared bands. In addition, the red tide 3D sample generation module demonstrates good spatial and spectral fidelity, enabling dominant species identification under conditions of few or no samples.

[0015] Example 1, such as Figure 1 As shown in the figure, the integrated method for generating and identifying broadband three-dimensional samples of dominant red tide species provided in this embodiment of the invention includes the following steps: S1, Construct an integrated model for generating three-dimensional samples of wideband red tides and identifying dominant species based on deep learning; S2, the red tide image generation module based on spectral feature transfer obtains red tide expansion samples of different dominant species, and a red tide dominant species identification module is constructed by integrating multi-scale feature extraction and cross-spatial attention mechanism; S3 trains the red tide dominant species identification module through a semi-supervised learning mode using student and teacher models. It then uses generated red tide samples and real red tide samples to achieve cross-temporal and spatial identification of red tide dominant species under wide-band conditions.

[0016] In step S2, a red tide image generation module based on spectral feature transfer is constructed using a generative adversarial network. This module consists of a generator and a discriminator. The generator takes real labeled images as input and generates red tide and non-red tide features of the image by constraining the label data. The generator consists of a shallow feature extraction layer, a global deep feature extraction layer, and a spectral constraint layer. The shallow feature extraction layer includes 3×3 convolution, normalization, and ReLU activation function layers, while the global deep feature extraction layer includes several residual global correlation attention modules. After global feature extraction, red tide spectral features are embedded using a linear mixing method, followed by image restoration through deconvolution. The shallow features are combined with the spectral information of different dominant red tide species in the spectral library to achieve red tide sample generation under red tide spectral constraints.

[0017] The discriminator adopts a multi-scale discriminative structure. After outputting multi-scale discriminative features, it calculates the discriminative loss at different scales based on the conditional discriminative loss and takes the average value as the final discriminative result.

[0018] Design a spectral domain-aware adversarial loss function for the generator. The expression is: ; ; In the formula, For the first A real image, For the first One generated image, For the expected value, For the discriminator in a known labeled image Under the premise that the generated image is considered It is the confidence level of the real image. The cosine angle of the spectrum. The number of bands for the HY-1C / D CZI is 4; For real images, For label images, To generate an image; For multi-scale discriminators, the loss function It is in the calculation The loss at each scale is used as the average of the losses to adjust the discriminant parameters as the final loss. The expression is: ; In the formula, To determine the scale quantity, For the expected value, They were respectively in the second At each scale, the discriminator can detect known label images. Under the premise of assuming a real image and generating images It is the confidence level of the real image; For real images Data distribution.

[0019] In step S2, the construction of red tide samples of different dominant species includes: Sample image pairs are obtained by randomly cropping remote sensing images and their corresponding labels, and 10% of the samples are randomly selected for model testing. Based on the construction of the sample set, spectra are randomly selected from the images, and red tide spectra are screened through K-Means clustering to construct a real red tide spectral library, which is used to constrain the generation of red tide samples.

[0020] In step S2, the red tide dominant species identification module based on cross-spatial feature fusion includes two branches. The first branch adopts an encoding and decoding structure to extract the spectral spatial features of red tides and non-red tides. The second branch extracts the spatial features between different red tide algal species by combining multi-scale feature extraction and cross-spatial attention module. The features of the two branches are fused using dense residual module to obtain the identification result of the red tide dominant species.

[0021] Feature extraction employs a multi-resolution parallel extraction architecture, using different Rate parameters (1, 6, 8, 12) to simultaneously extract features from multiple receptive fields. Specifically, the input features first pass through four parallel branches, each using a different dilated convolution Rate value for feature extraction. Rate=1 corresponds to standard convolution, capturing local detail features; Rate=6 provides a medium receptive field, suitable for capturing mid-scale features of red tide patches; Rate=8 and Rate=12 correspond to larger receptive fields, used to capture large-scale red tide distribution patterns and contextual information. The output features of the four parallel branches have the same spatial resolution but contain semantic information at different scales. Finally, they are concatenated along the channel dimension using a concatenation operation to form multi-scale fused features.

[0022] In terms of specific application scenarios, red tide distribution exhibits spatial heterogeneity. The purpose of employing cross-spatial attention is to capture long-distance spatial semantic relationships by establishing bidirectional spatial associations between local and global levels and between global and local levels, thereby achieving spatial feature associations of red tide algae species in different regions. However, traditional attention mechanisms are difficult to establish long-distance spatial associations from a mechanistic perspective. This invention directly compared the overall computational complexity of the comparative methods. Compared to DeepLab_v3 and Res_UNet, the FLOPs of this patented method are reduced by 2.25%-33.55%.

[0023] In step S2, the identification loss of dominant red tide species includes supervision loss. and unsupervised losses ; Monitoring losses The multi-task loss function is used, and its expression is: ; In the formula, For cross-entropy loss, For edge-weighted loss, These are the weighting coefficients.

[0024] Multi-task loss functions include cross-entropy loss. and edge-weighted loss Among them, edge-weighted loss focuses on red tide boundary extraction, while cross-entropy loss dominates global classification; ; In the formula, These represent the total number of samples and the total number of categories, respectively. For the first The sample is kind, For the first The sample belongs to the first The probability of a class These are the pixel position coordinates. For the image spatial domain, For position Category The probability, For position Edge strength at the location; Unsupervised loss Cross-entropy loss is calculated using pseudo-labels with high confidence. ; In the formula, These are the horizontal and vertical coordinates of the image, respectively. In unlabeled images Weighting In unlabeled images Predicted probability, In unlabeled images Pseudo-labels at the location.

[0025] In step S3, the model architectures of the student model and the teacher model are completely identical. The teacher model is responsible for obtaining the pseudo-labels of the generated red tide images, while the student model simultaneously identifies the generated red tide images and satellite red tide images. During the training process, parameters are passed to the teacher model using an exponential moving average method.

[0026] The method also includes an evaluation of the accuracy of red tide dominant species identification and an evaluation of the red tide sample generation effect; The accuracy of red tide dominant species identification was evaluated using six accuracy evaluation methods: overall accuracy, average accuracy, precision, recall, F1-score, and Kappa coefficient. ; ; ; In the formula, For overall accuracy, For average accuracy, For accuracy, For recall rate, The F1 score coefficient. Kappa coefficient This represents the total number of samples. For the number of dominant species, For the first The number of correctly identified samples in each class. For the first The number of samples in each class The number of correct and incorrect entries classified into this category, respectively. This represents the number of items that belong to this category but have been misclassified. To correctly identify the number of items that do not belong to this category; Four evaluation methods were used to assess the effectiveness of feature migration-based red tide sample generation: visual assessment, SSIM, PSNR, and spectral angle. Among them, the spectral angle was calculated only using the red tide water area. ; ; In the formula, Real footage of red tide And generate red tide images Structural similarity between them Peak signal-to-noise ratio, It is the spectral angle. To generate an image The maximum value, For the first A real image of a red tide. For the first A generated red tide image It is the stability constant. The images are real red tide footage and generated red tide footage, respectively. These represent the mean and standard deviation of the image pixel values, respectively. The dynamic range of pixel values. , , The first The training data and the spectral values ​​predicted by the network, The number of training samples. It is an inverse cosine function.

[0027] Example 2: The integrated system for generating broadband three-dimensional red tide samples and identifying dominant species provided in this embodiment of the invention includes: A red tide image generation module based on spectral feature transfer is used to generate three-dimensional red tide samples; Based on the dominant species identification module of cross-spatial feature fusion, the red tide image generation module based on spectral feature transfer obtains red tide expansion samples of different dominant species. The red tide dominant species identification module is constructed by integrating multi-scale feature extraction and cross-spatial attention mechanism. The red tide dominant species identification module is trained by semi-supervised learning mode of student model and teacher model. Using the generated red tide samples and real red tide samples, cross-temporal and spatial identification of red tide dominant species under wide band conditions is realized.

[0028] To further demonstrate the positive effects of the above embodiments, the present invention conducts the following experiments based on the above technical solutions.

[0029] 1. Data and Methods; 1.1 Data; The Haiyang-1C and Haiyang-1D satellites (HY-1C / D), as a new generation of domestically produced medium- and high-resolution ocean optical satellites, were launched in September 2018 and June 2020, respectively. They carry an Ocean Color and Temperature Scanner (COCTS), a Coastal Zone Imager (CZI), an Ultraviolet Imager (UVI), an onboard calibration spectrometer, and an Automatic Identification System (AIS). The CZI sensor has a spatial resolution of 50m and a swath width of 950km. Through a network of the C and D satellites (morning and afternoon), a revisit cycle of twice every three days can be achieved. Compared to other common medium- and high-resolution satellites, the HY-1C / DCZI satellites, due to their wide swath width, high resolution, and short revisit cycle, have become the main data source for operational red tide monitoring. They can effectively compensate for the insufficient spatial resolution of ocean color satellites and are helpful for conducting detailed detection of small-scale red tides (see Table 1).

[0030] Table 1 Parameters of HY-1C / D CZI Sensor

[0031] Considering the main dominant species categories, toxicity / harmfulness, and available data of red tide outbreaks, this invention takes three dominant algal species—Noctiluca scintillans, Noctiluca scintillans, and Hemlockella helioscopia—as examples to conduct a red tide dominant species identification experiment. Specific information is shown in Table 2.

[0032] Noctiluca scintillans, a typical dominant species in red tides, belongs to the dinoflagellates and has been reported in all temperate, subtropical, and tropical coastal regions. Noctiluca scintillans includes two types: red Noctiluca scintillans (red... Noctiluca scintillans ) and green noctiluca Noctiluca scintillans Among them, *Noctiluca scintillans* is widely distributed in coastal areas worldwide, and its outbreaks easily lead to fish deaths and water quality deterioration. *Noctiluca scintillans* often appears as red or golden stripes and produces blue fluorescence at night, hence its nickname "blue tears." *Noctiluca scintillans* differs from *Noctiluca scintillans* in its nutrient acquisition methods, growth areas, and spectrum. *Noctiluca scintillans* cells contain a large number of symbiotic green algae, so the water turns green during outbreaks. It is mainly distributed in tropical sea areas such as Southeast Asia and a certain region of the Arabian Sea. Every spring, the area affected by *Noctiluca scintillans* red tides in a certain region of the Arabian Sea can reach millions of square kilometers. *Akashiwo sanguinea*, originally named *Gymnodinium sanguineum Hirasaka*, is a toxic red tide algae species commonly found in coastal and estuarine areas. During outbreaks, the water turns brownish-red, causing deaths of fish and seabirds. Analysis shows that *Hemanthocarpus erythrocarpus* is a eurythermal and euryhaline algae, which has a strong adaptability to changes in water temperature and salinity. It grows rapidly and has been recorded to cause serious red tide events in waters around the world, including Europe, North America, South America, and Asia (Japan and China).

[0033] Table 2 List of Experimental Images

[0034] 1.2 Sample Set Construction; Considering the distribution characteristics of different dominant red tide species, remote sensing images and their corresponding labels were randomly cropped into 128×128 pixel sample image pairs, generating a total of 1285 samples, such as... Figure 2 As shown, 10% of the samples were randomly selected for model testing, and the statistical results of the samples of different dominant red tide species are shown in Table 3. Due to differences in the effective satellite data, the sample sizes of different dominant species were uneven. For example, although *Noctiluca scintillans* occurs globally, its strip-like and discrete distribution resulted in fewer effective samples, while *Hacochloa hemangiosus* was mostly distributed in a planar pattern, resulting in more effective samples. Ultimately, the number of *Hacochloa hemangiosus* samples was approximately twice that of *Noctiluca rubiginosa* and *Noctiluca scintillans*.

[0035] Table 3. Sample Size Statistics

[0036] Based on the constructed sample set, this invention also randomly selected 10,000 spectra from images and used K-Means clustering to screen out typical red tide spectra to construct a real red tide spectral library, such as... Figure 3 As shown, this is used to constrain the generation of red tide samples in order to improve the spectral fidelity of the generated samples.

[0037] 1.3 Accuracy Evaluation Indicators; The accuracy evaluation indicators of this invention include two parts: the accuracy evaluation indicators for identifying dominant red tide species and the evaluation indicators for the effectiveness of red tide sample generation.

[0038] The performance of the method described in this invention was evaluated using six accuracy assessment methods: overall accuracy (OA), average accuracy (AA), precision, recall, F1-score, and Kappa coefficient. The detailed formulas are as follows: ; ; ; To evaluate the effectiveness of the feature transfer-based red tide sample generation model, four evaluation methods were employed: visual assessment, SSIM, PSNR, and spectral angle. The spectral angle was calculated using only the red tide water region.

[0039] ; ; 2. An Integrated Method for Broadband 3D Sample Generation and Identification of Dominant Red Tide Species: Addressing the challenge of identifying dominant red tide species using high-resolution broadband satellites, this invention constructs an integrated model (RT-SGINet) based on deep learning for generating 3D red tide samples and identifying dominant species, achieving cross-temporal and spatial identification of dominant red tide species under broadband conditions. The network structure of the semi-supervised deep learning red tide dominant species identification model with a joint generative adversarial network is as follows: Figure 4 As shown, the model consists of two parts: RTS_GAN and CSF-RSI. The input is red tide samples and their corresponding labels, and the output is red tide generating samples and red tide dominant species identification results.

[0040] 2.1 Red Tide Image Generation Module Based on Spectral Feature Transfer; Red tide outbreaks exhibit significant seasonality and regionality. The limited number and uneven distribution of remote sensing samples of different dominant species make it difficult to meet the demands of deep learning models for large-scale, high-quality training data. Furthermore, the creation of red tide samples relies on expert experience, is time-consuming, labor-intensive, and difficult. To alleviate the sample shortage problem, image augmentation and other techniques are commonly used to expand the training data. However, existing augmentation methods struggle to simultaneously preserve the spectral structure and spatial edge features of red tide images, easily introducing noise or blurring the target area, thus affecting the subsequent identification of dominant species. Therefore, this invention constructs a red tide image generation model based on spectral feature transfer using generative adversarial networks (see...). Figure 5 Specifically, it includes two parts: a generator and a discriminator.

[0041] Significant differences exist in the spectral responses and spatial distributions of different dominant red tide species. To enhance the ability of generated samples to express the spatial-spectral characteristics of red tides, the generator uses real labeled images as input, constrains the red tide and non-red tide features of the generated images through label data, and introduces a real wide-band spectral library to enhance the spectral fidelity of the generated images. Specifically, the generator consists of a shallow feature extraction layer, a global deep feature extraction layer, and a spectral constraint layer. The shallow feature extraction layer includes 3×3 convolution, normalization, and ReLU activation function layers, while the global deep feature extraction layer includes several residual global correlation attention modules. After global feature extraction, red tide spectral features are embedded using a linear mixture method, followed by image restoration through deconvolution. By combining the shallow features and fusing the spectral information of different dominant red tide species in the spectral library, red tide samples under red tide spectral constraints are generated.

[0042] Considering the small-scale targets and edge information in red tide images, the discriminator adopts a multi-scale discrimination structure to fully utilize the spatial information of the red tide at different scales, enhancing the perception of local details and target boundaries. After outputting multi-scale discrimination features, discrimination losses at different scales are calculated based on the conditional discrimination loss, and the average value is taken as the final discrimination result. Furthermore, a spectral domain-aware adversarial loss function is designed for the generator: ; Based on conditional generative adversarial loss, the spectral similarity between the generated image and the real image is further constrained by the following formula to ensure that the generated image has a reliable red tide spectrum.

[0043] ; For multi-scale discriminators, the loss function It is in the calculation The loss at each scale is used as the average of the losses to adjust the discriminant parameters as the final loss. The expression is: ; 2.2 Red Tide Dominant Species Identification Module Based on Cross-Spatial Feature Fusion; The spectral differences among dominant red tide species are small, and broadband satellite spectral resolution is low with limited bands, making it impossible to directly use traditional index methods for dominant species identification. Furthermore, the distribution of dominant red tide species exhibits significant regionality, requiring high model generalization ability. Therefore, this invention proposes a red tide dominant species identification model based on cross-spatial feature fusion (see...). Figure 6 The model was trained using a student-teacher semi-supervised learning model to improve its generalization ability. The model consists of two branches. The first branch employs an encoding / decoding structure to extract rich spectral spatial features of both red tides and non-red tides. The second branch further extracts spatial features among different red tide algal species by combining multi-scale feature extraction and cross-spatial attention module calculations, as shown in the formula below. Finally, a dense residual module is used to fuse the features from both branches to obtain the identification results of the dominant red tide species.

[0044] ; ; In the formula, These are the channel-weighted feature maps, the cross-spatial attention fusion feature maps, and the final output feature maps, respectively. For global average pooling, For normalization function, For convolution operations, As input features, Vertical average pooling, For matrix multiplication, This is horizontal average pooling.

[0045] Furthermore, the model architectures of the student model and the teacher model are completely identical. The teacher model is responsible for obtaining pseudo-labels for the generated red tide images, while the student model simultaneously identifies the generated red tide images and satellite red tide images. During training, parameters are passed to the teacher model using the Exponential Moving Average (EMA) method.

[0046] The implementation process of the student model and the teacher model is as follows: (1) Training phase: The teacher network is pre-trained using red tide images (labeled data) to enable it to have basic recognition capabilities; (2) Co-training phase: The teacher network is responsible for generating pseudo-labels for images (unlabeled data), while the student network trains on both red tide images (labeled data) and generated images (unlabeled data); (3) The student network is trained using the designed loss function, while the teacher model updates parameters using an exponential shift method.

[0047] Table 8 shows the experimental results for different proportions of data generated in semi-supervised training. The results indicate that the required labeled data can be reduced by approximately 75% (1:3). Compared to manually added Gaussian noise, the generalization ability test for different regions, times, and sensors is more suitable for the application scenario of this patent. The generalization ability test for different regions, times, and sensors has already been conducted. Figure 10 , Figure 11 Displayed in China; Figure 10 , Figure 11 These are actual deployment examples, showing the results of red tide dominant species identification from different regions and different satellites using the same satellite. The loss in identifying dominant species in red tides includes two parts: surveillance loss. and unsupervised losses . Employing a multi-task loss function: Specifically, this includes cross-entropy loss: And edge-weighted loss: Among them, the edge-weighted loss focuses on red tide boundary extraction, while the cross-entropy loss dominates the global classification. Cross-entropy loss is calculated using pseudo-labels with high confidence levels. This avoids recognition errors caused by low-confidence pseudo-labels.

[0048] 3. Results and Analysis; The model was implemented using the PyTorch framework and executed on an NVIDIA RTX A6000. The model training learning rate was initialized to 1×10⁻⁶. -4 An adaptive learning rate reduction method was used to adjust the learning rate, with a batch size of 32 and an epoch of 300. Training was stopped if the loss function increased more than 20 times or the number of iterations exceeded 300. Of these, 1156 samples were used for network training, and 183 samples were used to generate red tide image quality checks.

[0049] 3.1 Red Tide Sample Generation Results; To verify the effectiveness of red tide sample generation, qualitative and quantitative evaluations were performed on different generated samples, and the results were compared with those of models without spectral information constraints (RT_GAN) and pix2pix image generation models. The sample generation results for different algal species are shown below. Figure 7 As shown in the figure, all three methods yielded spatial distributions approximately similar to the original image. However, the image generated by the method of this invention exhibits a clear red tide geometry, sharp edges, and detail information closely resembling the real image. Quantitative evaluation results show that the method of this invention achieves average SSIM and PNSR values ​​of 0.98 and 50.28, respectively, outperforming the other two methods (see Table 4).

[0050] Table 4 Comparison of the accuracy of red tide generation samples using different methods

[0051] The method of this invention improves the spectral fidelity of the generated samples by introducing a red tide spectral library to constrain the spectral information of the generated images. As shown in Table 4, the average spectral angle between the generated images and the real images of the three dominant species is 1.33°, and they are consistent with the reference images in terms of spectral shape and spectral values ​​(see Table 4). Figure 8 The other two methods produce images with significant spectral differences from the original images; the spectral angle between the image generated by *Noctiluca scintillans* and the real image is as high as 14°. Apart from the three types of red tide algae, the generated images obtained by this invention also exhibit good spectral shape representation of seawater, clouds, and land, demonstrating far superior spectral preservation compared to the other methods.

[0052] 3.2 Results of Dominant Species Identification in Red Tide: The performance of the algorithm was qualitatively and quantitatively evaluated using test samples. Quantitative results (see Table 5) show that the model proposed in this invention has high accuracy in identifying dominant species in red tides, with an overall accuracy of 94.38% and an average accuracy of 92.21%. The F1-scores of each dominant species are above 0.91. In comparison, *Noctiluca scintillans* exhibits a slight increase in its near-infrared spectrum, differing from the spectra of *Noctiluca rubiginosa* and *Hacochloa rubiginosa*, and thus has the highest identification accuracy. Figure 9 The identification results show that the present invention can effectively identify three dominant species with fewer misidentifications, and the identification effect of edges and strip-shaped regions in a single category is better.

[0053] Table 5. Accuracy of Red Tide Dominant Species Identification

[0054] Furthermore, to verify the effectiveness of this invention, commonly used image semantic segmentation models (Deeplab V3+, ResU-Net) were selected as comparison models and compared with the model of this invention. The comparison methods used the same red tide sample generation and training strategies as the method of this invention. In terms of computational efficiency, the computational complexity of the model of this invention is 19.56 GFLOPs, a reduction of 2.25% compared to Deeplab V3+'s 20.01 GFLOPs, and a reduction of 33.55% compared to ResU-Net's 29.43 GFLOPs, significantly improving computational efficiency while maintaining high accuracy. The results show (see...) Figure 9The model of this invention exhibits better recognition performance than the comparative methods, especially in areas with low biomass and at the edge of red tides. Quantitative evaluation results, as shown in Table 6, indicate that the deep learning-based red tide dominant species identification model developed in this invention has the highest recognition accuracy, outperforming other methods in all three evaluation metrics, with an overall accuracy (OA) of 94.48%. Compared to the ResU-Net model, the model of this invention further extracts spatial features of different dominant red tide species through multi-scale feature extraction and a cross-spatial attention module, significantly improving the recognition performance of different dominant species. Deeplab V3+ also employs multi-scale feature extraction in its encoder module, but this invention, by combining it with a cross-spatial attention module, enhances the cross-spatial feature extraction capability of red tides, improving the OA by 3.52%.

[0055] Table 6 Comparison of the accuracy of red tide dominant species identification using different methods

[0056] 4. Discussion; 4.1 Impact of Red Tide Generated Samples on Dominant Species Identification; To investigate the impact of red tide generated samples on dominant species identification, this invention introduced different proportions of generated samples for dominant species identification experiments while maintaining a fixed total training sample set (see Table 7). The results show that the identification accuracy is comparable to that of the real sample set when different proportions of generated samples are introduced, with the best accuracy achieved when 50% generated samples are introduced. This is mainly because the spatial-spectral constraints and spectral feature transfer in SFT-Net further enhance the red tide edge features in the generated samples. The experimental results further demonstrate that the spatial-spectral information of the red tide generated images in this invention is close to that of the real images, enabling the identification of dominant red tide species under conditions of few or no samples.

[0057] Table 7. Dominant species identification results with different generated sample ratios

[0058] Furthermore, this invention also conducted dominant species identification experiments using training sample sets augmented with samples generated at different ratios (see Table 8). The results show that as the number of samples increases, the identification accuracy begins to increase significantly, but the rate of increase slows down when the number of augmented samples exceeds three times. When the number of augmented samples exceeds five times, the dominant species identification accuracy decreases. Based on this, the ratio of real to generated images was set to 1:3 during the training of the dominant species identification model.

[0059] Table 8. Results of dominant species identification in expanded sample sets generated at different ratios.

[0060] 4.2 Regional Adaptability Analysis; To test the regional applicability of the model of this invention, HY-1C / D CZI data from a certain province on March 14, 2022, a certain region in the Bohai Sea on September 17, 2023, a certain region in the Arabian Sea on February 20, 2020, a certain coastal area of ​​Thailand on August 10, 2023, and a certain region in the East China Sea on October 1, 2021 were used to conduct experiments on the identification of dominant red tide species under different red tide events in different regions.

[0061] Quantitative results show that this invention achieves high accuracy, with average F1-scores of 0.86, 0.92, and 0.89 for the three dominant species (see Table 9). The identification results of dominant species in different regions are as follows: Figure 10 As shown, thanks to the feature extraction of cross-spatial attention, the method of the present invention is less sensitive to regional changes and can be widely applied to the identification of dominant red tide species in different sea areas and at different times, with good generalization ability.

[0062] Table 9. Accuracy of Identification of Dominant Red Tide Species in Different Regions

[0063] 4.3 Applicability to Different Satellites; To evaluate the applicability of the algorithm of this invention to other medium-to-high spatial resolution wide-band satellite imagery, this invention uses *Nostoc commune* as an example, directly employing the trained model to conduct recognition experiments on GF1 WFV and Sentinel-2B MSI satellite images respectively. For Sentinel-2B MSI, bands similar to HY-1C / D CZI (blue, green, red, and near-infrared) were selected for testing. Figure 11 The image shows the identification results. Experimental results indicate that this method successfully identified the outbreak range of the dominant species in two regions. This demonstrates that the method of this invention has minimal impact on the model due to the satellite sensor's bandwidth and center wavelength, and can be widely applied to wideband satellites with similar band settings, exhibiting good applicability.

[0064] 5. Summary; Addressing the challenges of small spectral differences and uneven sample numbers among different dominant red tide species in medium-to-high resolution satellite red tide identification, this invention, using HY-1C / D CZI as an example, constructs an integrated model for three-dimensional red tide sample generation and dominant species identification based on deep learning. This model first utilizes a red tide generation module based on spectral feature transfer to obtain expanded samples of different dominant red tide species, balancing the number of different dominant species and increasing the diversity of red tide samples. Then, a red tide dominant species identification module is constructed by integrating multi-scale feature extraction and cross-spatial attention mechanisms to improve the identification capability of dominant red tide species in different sea areas and seasons. Finally, through a student-teacher semi-supervised training mode, the generated red tide samples and real red tide samples are used to identify *Noctiluca scintillans*, *Noctiluca greenis*, and *Hacochloa erythrophora*.

[0065] The RT-SGINet model achieved good results in red tide dominant species identification, effectively identifying different dominant species with OA and AA scores of 94.48% and 92.21%, respectively, and a Kappa score of 0.93, all of which are improvements in accuracy compared to other typical deep learning semantic segmentation models. Furthermore, the model exhibits good generalization ability and can be applied to the identification of red tide algae species in different regions and using satellites with red, green, blue, and near-infrared bands. In addition, the red tide generation module based on spectral feature transfer proposed in this invention can generate red tide sample images with spatial and spectral information close to the real image, effectively enabling dominant species identification under conditions of few or no sample annotations.

[0066] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention and within the spirit and principles of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for generating and identifying broadband three-dimensional samples of dominant red tide species, characterized in that, The method includes the following steps: S1, Construct an integrated model for generating three-dimensional samples of wideband red tides and identifying dominant species based on deep learning; S2, the red tide image generation module based on spectral feature transfer obtains red tide expansion samples of different dominant species, and a red tide dominant species identification module is constructed by integrating multi-scale feature extraction and cross-spatial attention mechanism; S3 trains the red tide dominant species identification module through a semi-supervised learning mode using student and teacher models, and uses generated red tide samples and real red tide samples to achieve cross-temporal and spatial identification of red tide dominant species under wide-band conditions. In step S2, the red tide image generation module based on spectral feature transfer is constructed using a generative adversarial network. The red tide image generation module based on spectral feature transfer includes two parts: a generator and a discriminator. The generator takes real labeled images as input and generates red tide and non-red tide features of the image by constraining the label data. The discriminator adopts a multi-scale discriminant structure. After outputting multi-scale discriminant features, it calculates the discriminant loss at different scales based on the conditional discriminant loss and takes the average value as the final discriminant result. The generator consists of a shallow feature extraction layer, a global deep feature extraction layer, and a spectral constraint layer. The shallow feature extraction layer includes a 3×3 convolution, normalization, and ReLU activation function layer, while the global deep feature extraction layer includes several residual global correlation attention modules. After global feature extraction, red tide spectral features are embedded using a linear mixing method, followed by deconvolution for image restoration. The generator combines the shallow features with spectral information of different dominant red tide species from the spectral library to generate red tide samples under spectral constraints. Design a spectral domain-aware adversarial loss function for the generator. The expression is: ； ； In the formula, For the first A real image, For the first One generated image, For the expected value, For the discriminator in a known labeled image Under the premise that the generated image is considered It is the confidence level of the real image. The cosine angle of the spectrum. The number of bands for the HY-1C / D CZI is 4; For real images, For label images, To generate an image; For multi-scale discriminators, the loss function It is in the calculation The loss at each scale is used as the average of the losses to adjust the discriminant parameters as the final loss. The expression is: ； In the formula, To determine the scale quantity, For the expected value, They were respectively in the second At each scale, the discriminator can detect known label images. Under the premise of assuming a real image and generating images It is the confidence level of the real image; In step S2, the red tide dominant species identification module includes two branches. The first branch adopts an encoding and decoding structure to extract the spectral spatial features of red tides and non-red tides. The second branch extracts the spatial features between different red tide algal species by combining multi-scale feature extraction and cross-spatial attention module. The features of the two branches are fused using the dense residual module to obtain the identification result of the red tide dominant species. Feature extraction employs a multi-resolution parallel extraction architecture, simultaneously extracting features across multiple receptive fields using different Rate parameters. Specifically, the input features first pass through four parallel branches, each employing a different dilated convolution Rate value for feature extraction. Rate=1 corresponds to standard convolution, capturing local detail features; Rate=6 provides a medium receptive field, suitable for capturing mid-scale features of red tide patches; Rate=8 and Rate=12 correspond to larger receptive fields, used to capture large-scale red tide distribution patterns and contextual information. The output features of the four parallel branches have the same spatial resolution but contain semantic information at different scales. Finally, they are concatenated along the channel dimension using a concatenation operation to form multi-scale fused features. The loss in identifying dominant red tide species includes monitoring loss. and unsupervised losses ; Monitoring losses The multi-task loss function is used, and its expression is: ； In the formula, For cross-entropy loss, For edge-weighted loss, These are the weighting coefficients; Multi-task loss functions include cross-entropy loss. and edge-weighted loss Among them, edge-weighted loss focuses on red tide boundary extraction, while cross-entropy loss dominates global classification; ； ； In the formula, These represent the total number of samples and the total number of categories, respectively. For the first The sample is kind, For the first The sample belongs to the first The probability of a class These are the pixel position coordinates. For the image spatial domain, For position Category The probability, For position Edge strength at the location; Unsupervised loss Cross-entropy loss is calculated using pseudo-labels with high confidence. ； In the formula, These are the horizontal and vertical coordinates of the image, respectively. In unlabeled images Weighting In unlabeled images Predicted probability, In unlabeled images Pseudo-labels at the location.

2. The integrated method for generating and identifying broadband three-dimensional samples of dominant red tide species according to claim 1, characterized in that, In step S2, the construction of red tide samples of different dominant species includes: Sample image pairs are obtained by randomly cropping remote sensing images and their corresponding labels, and 10% of the samples are randomly selected for model testing. Based on the construction of the sample set, spectra are randomly selected from the images, and red tide spectra are screened through K-Means clustering to construct a real red tide spectral library, which is used to constrain the generation of red tide samples.

3. The integrated method for generating and identifying broadband three-dimensional samples of dominant red tide species according to claim 1, characterized in that, In step S3, the teacher model is responsible for obtaining pseudo-labels for the generated red tide images, while the student model simultaneously identifies both the generated red tide images and satellite red tide images. During the training process, parameters are passed to the teacher model using an exponential moving average method. The implementation process of the student model and the teacher model is as follows: (1) Training phase: The teacher network is pre-trained using red tide images to enable it to recognize the red tide; (2) Collaborative training phase: The teacher network is responsible for generating pseudo-labels for the images, while the student network is trained on both the red tide images and the generated images. (3) The student network is trained using the designed loss function, while the teacher model updates the parameters using an exponential shift method.

4. The integrated method for generating and identifying broadband three-dimensional samples of dominant red tide species according to claim 1, characterized in that, The method also includes an evaluation of the accuracy of red tide dominant species identification and an evaluation of the red tide sample generation effect; The accuracy of red tide dominant species identification was evaluated using six accuracy evaluation methods: overall accuracy, average accuracy, precision, recall, F1-score, and Kappa coefficient. ； ； ； ； ； ； In the formula, For overall accuracy, For average accuracy, For accuracy, For recall rate, The F1 score coefficient. Kappa coefficient This represents the total number of samples. For the number of dominant species, For the first The number of correctly identified samples in each class. For the first The number of samples in each class The number of correct and incorrect entries classified into this category, respectively. This represents the number of items that belong to this category but have been misclassified. To correctly identify the number of items that do not belong to this category; Four evaluation methods were used to assess the effectiveness of feature migration-based red tide sample generation: visual assessment, SSIM, PSNR, and spectral angle. Among them, the spectral angle was calculated only using the red tide water area. ； ； ； ； In the formula, Real footage of red tide And generate red tide images Structural similarity between them Peak signal-to-noise ratio, It is the spectral angle. To generate an image The maximum value, For the first A real image of a red tide. For the first A generated red tide image It is the stability constant. The images are real red tide footage and generated red tide footage, respectively. These represent the mean and standard deviation of the image pixel values, respectively. The dynamic range of pixel values. , , The first The training data and the spectral values ​​predicted by the network, The number of training samples. It is an inverse cosine function.

5. An integrated system for generating and identifying broadband three-dimensional samples of dominant red tide species, characterized in that, This system is used to control the integrated method for generating and identifying broadband three-dimensional samples of dominant red tide species as described in any one of claims 1-4. The system includes: A red tide image generation module based on spectral feature transfer is used to generate three-dimensional red tide samples; Based on the dominant species identification module of cross-spatial feature fusion, the red tide image generation module based on spectral feature transfer obtains red tide expansion samples of different dominant species. The red tide dominant species identification module is constructed by integrating multi-scale feature extraction and cross-spatial attention mechanism. The red tide dominant species identification module is trained by semi-supervised learning mode of student model and teacher model. Using the generated red tide samples and real red tide samples, cross-temporal and spatial identification of red tide dominant species under wide band conditions is realized.

Images