A red tide / hypoxia intelligent early warning method and device based on multi-source data fusion

Abstract

This invention discloses a red tide / hypoxia intelligent early warning method and device based on multi-source data fusion. The system constructs a three-dimensional monitoring architecture of "air-space-sea", integrating high-resolution satellite remote sensing, SAR data, and in-situ sensor information of jacket structures. Spatiotemporal data interpolation is achieved through the DINEOF algorithm, and transfer learning is combined to optimize the red tide identification model and LSTM network to predict hypoxia risk. A lightweight CNN is deployed at the edge computing layer to achieve real-time algal species classification. The cloud dynamically calculates the comprehensive risk index of red tide / hypoxia, triggering a graded response: yellow alerts automatically adjust feeding and oxygenation, and red alerts drive the jacket structures to sink for disaster avoidance. The system achieves low-power communication and device linkage through LoRaWAN and Modbus RTU protocols, forming a "monitoring-decision-execution" closed loop with a response latency of less than 10 minutes, significantly improving early warning accuracy and emergency automation level, and is suitable for intelligent aquaculture scenarios in deep sea areas.

Description

Technical Field

[0001] This invention relates to the field of marine environmental monitoring and disaster early warning technology, and in particular to a red tide / hypoxia intelligent early warning method and device based on multi-source data fusion. Background Technology

[0002] Red tide and hypoxia disaster monitoring, as core technologies for marine aquaculture environmental safety, are widely used in jacket aquaculture scenarios. Among related technologies, a multi-dimensional monitoring system has been constructed through the collaborative operation of satellite remote sensing, in-situ sensors, and edge computing. Specifically, this system covers the entire process from data acquisition to risk decision-making, including space-based observations from multispectral remote sensing (such as Sentinel-2MSI), synthetic aperture radar (SAR), and hyperspectral satellite (HY-1D), as well as a sea-based monitoring network consisting of dissolved oxygen sensors, ADCP vertical current profilers, and underwater cameras deployed on jackets. With the development of AI technology, existing systems are gradually incorporating CNN image classification and LSTM time-series prediction models, but a complete edge-cloud collaborative architecture has not yet been formed.

[0003] However, existing red tide / hypoxia early warning methods directly employ a single threshold method (such as chlorophyll concentration exceeding...). The system lacks a comprehensive analysis of multiple factors, including ocean currents, weather, and aquaculture activities, which can lead to a false alarm rate exceeding 30% or significant warning delays. Specifically, insufficient coordination between satellite remote sensing data (1 km spatial resolution) and fixed-point sensors (local static data) prevents the construction of a dynamic model covering the entire region. Traditional methods fail to distinguish between toxic dinoflagellates and harmless diatoms, resulting in warning failures. Furthermore, the lack of integration of key parameters such as feeding amount and current velocity gradient leads to a disconnect between the prediction model and the actual environment. In addition, the emergency response mechanism relies on manual decision-making (delay > 1 hour), making it difficult to establish automated closed-loop control. Summary of the Invention

[0004] The present invention aims to at least partially solve one of the technical problems in the related art.

[0005] Therefore, the first objective of this invention is to propose an intelligent early warning method for red tide / hypoxia based on multi-source data fusion.

[0006] Another objective of this invention is to propose a red tide / hypoxia intelligent early warning device based on multi-source data fusion.

[0007] The third objective of this invention is to provide a computer device.

[0008] A fourth objective of this invention is to provide a non-transitory computer-readable storage medium.

[0009] To achieve the above objectives, a first aspect of the present invention proposes a red tide / hypoxia intelligent early warning method based on multi-source data fusion, comprising:

[0010] S1, acquires satellite remote sensing data, jacket sensor data, and underwater image data; S2, based on the DINEOF algorithm, the satellite remote sensing data and the jacket sensor data are spatiotemporally interpolated to generate a dynamic risk heat map of the entire domain, and the cross-sea red tide characteristics are adapted and optimized by combining the transfer learning model. S3, a lightweight convolutional neural network deployed on edge computing nodes is used to perform real-time algae classification on the underwater image data. At the same time, a low oxygen prediction model including dissolved oxygen gradient, feeding amount and ocean current stagnation index is built in the cloud through a long short-term memory network, and the prediction threshold is dynamically updated. S4, based on the quantitative results of the red tide risk index and the hypoxia risk index, uses the ModbusRTU protocol to link the aerator, feeding system and jacket hydraulic device, and performs feeding amount adjustment, aerator start-up or jacket lowering operation according to the preset graded response strategy.

[0011] In one embodiment of the present invention, S1 includes: S11 acquires chlorophyll distribution data at a resolution of 10 meters through multispectral satellites, sea surface wind speed and wave height data through synthetic aperture radar, and algal spectral characteristic data through hyperspectral satellites. S12 uses an underwater camera deployed on the jacket to collect algae density images at a frequency of 3 frames per minute, and transmits the image data to the edge computing node after compression via the LoRaWAN protocol; S13, the satellite remote sensing data includes multispectral, synthetic aperture radar and hyperspectral data, and the jacket sensor data includes dissolved oxygen vertical distribution data and ocean current velocity data.

[0012] In one embodiment of the present invention, S2 includes: S21. The DINEOF algorithm is used to interpolate the chlorophyll concentration data in the cloud-covered area to generate a dynamic risk heat map of the entire region. S22, the ResNet-50 pre-trained network was fine-tuned using transfer learning techniques to extract the red tide spectral features from the field data and optimize the model parameters.

[0013] In one embodiment of the present invention, S3 includes: S31 uses a lightweight MobileNetV3 convolutional neural network to classify underwater images in real time, ensuring that the inference latency of a single frame image is less than 200ms. S32 uses a cloud-based LSTM neural network model to update the predicted weights of dissolved oxygen gradient, feeding amount, and ocean current stagnation index hourly to adapt to environmental changes.

[0014] In one embodiment of the present invention, S4 includes: S41, when the red tide risk index reaches 0.4-0.6, reduce the feeding amount by 50% and start the aerator through the ModbusRTU protocol. The aerator adopts a pure oxygen to liquid ratio of 1:5. S42, when the low oxygen risk index exceeds 0.6, controls the hydraulic system of the jacket to sink to a disaster avoidance depth of 20 meters via the ModbusRTU protocol, and at the same time shuts down the feeding system.

[0015] The present invention discloses a red tide / hypoxia intelligent early warning method and device based on multi-source data fusion, which can realize high-precision dynamic early warning and automated emergency response for red tide and hypoxia disasters, significantly improve the accuracy of early warning and shorten the response time, and effectively form a closed-loop control of "monitoring-decision-execution".

[0016] To achieve the above objectives, a third aspect of this application provides a computer device, including a processor and a memory; wherein the processor runs a program corresponding to the executable program code stored in the memory, in order to implement a red tide / hypoxia intelligent early warning method based on multi-source data fusion as described in the first aspect embodiment.

[0017] To achieve the above objectives, a fourth aspect of this application provides a non-transitory computer-readable storage medium storing a computer program that, when executed by a processor, implements a red tide / hypoxia intelligent early warning method based on multi-source data fusion as described in the first aspect embodiment.

[0018] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0019] Figure 1 This is a flowchart of a red tide / hypoxia intelligent early warning method based on multi-source data fusion according to an embodiment of the present invention; Figure 2 This is a diagram of a red tide / hypoxia intelligent early warning structure based on multi-source data fusion according to an embodiment of the present invention; Figure 3 It is a computer device according to an embodiment of the present invention. Detailed Implementation

[0020] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0021] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0022] The following description, with reference to the accompanying drawings, describes an intelligent early warning method and apparatus for red tide / hypoxia based on multi-source data fusion according to an embodiment of the present invention.

[0023] Example 1 Figure 1 This is a flowchart of a red tide / hypoxia intelligent early warning method based on multi-source data fusion according to an embodiment of the present invention, such as... Figure 1 As shown, it includes: like Figure 1 As shown, a red tide / hypoxia intelligent early warning method based on multi-source data fusion includes the following steps: S1 acquires satellite remote sensing data, jacket sensor data, and underwater image data.

[0024] Specifically, this step involves the acquisition and integration of multi-source data. In some implementations, the system uses satellite remote sensing platforms, jacket sensor networks, and underwater imaging equipment to simultaneously acquire multispectral, synthetic aperture radar (SAR), and hyperspectral data, while also acquiring dissolved oxygen vertical distribution data and ocean current velocity data, in order to achieve multi-dimensional perception of red tides and hypoxia disasters.

[0025] Furthermore, satellite remote sensing data primarily originates from the Sentinel-2 MSI sensor, the hyperspectral satellite HY-1D, and SAR satellites (such as Sentinel-1). Multispectral data provides chlorophyll concentration retrieval capability at 10-meter resolution, while hyperspectral data is used to extract spectral characteristics of algae, improving species resolution accuracy in red tide identification. SAR data is used to penetrate cloud layers, acquiring sea surface roughness and wind field information to assist in predicting red tide diffusion paths. A jacketed structure sensor network is deployed in aquaculture areas, including multi-point dissolved oxygen sensors (such as YSIProODO), which collect vertical dissolved oxygen distribution data at 0.5-meter intervals with an accuracy of ±0.2 mg / L and a sampling frequency of once every 10 minutes. ADCP equipment (such as Nortek Vector) measures ocean current velocity at a 1Hz sampling rate, with a vertical resolution of up to 1 meter and a velocity measurement range of 0.01–2.0 m / s, meeting the dynamic modeling requirements for low-oxygen diffusion trends. Underwater image data is captured by high-definition cameras (such as GoPro Hero10), with an image resolution of 4K and a frame rate of 30fps, and is processed in real time through edge computing nodes.

[0026] Furthermore, satellite remote sensing data must meet a spatial resolution of 10 meters and a revisit cycle of at least once a day to ensure data timeliness and spatial continuity. Jacket structure sensor data must have Modbus RTU communication capabilities, support RS-485 physical layer transmission, and the data packet format must conform to the ISO / IEC 11073 standard. Underwater image data must meet preprocessing requirements such as illumination compensation and motion blur correction to improve the recognition accuracy of subsequent CNN models.

[0027] Furthermore, this step is deployed in practical applications in nearshore aquaculture areas, especially suitable for sea areas prone to typhoons and with severe water stratification. Through the synchronous acquisition of multi-source data, the system can construct a high spatiotemporal resolution environmental state model, providing high-quality input for subsequent DINEOF interpolation and LSTM prediction.

[0028] Furthermore, this step enables early and multi-dimensional perception of red tides and hypoxia hazards, significantly improving the spatiotemporal coverage and data reliability of the early warning system. Through joint monitoring of vertical dissolved oxygen and ocean current velocity, the system can identify the formation and diffusion mechanisms of hypoxia zones, providing crucial decision-making basis for subsequent closed-loop control.

[0029] Furthermore, S1 includes: S11 acquires chlorophyll distribution data at a resolution of 10 meters via multispectral satellites, sea surface wind speed and wave height data via synthetic aperture radar, and algal spectral characteristic data via hyperspectral satellites.

[0030] Specifically, this step provides high-precision, multi-dimensional environmental parameter inputs for intelligent early warning of red tide and hypoxia disasters through the collaborative acquisition of multi-source remote sensing data.

[0031] Furthermore, the system employs multispectral satellites (such as Sentinel-2MSI) to acquire chlorophyll distribution data with a resolution of 10 meters. Its operating band covers the visible to near-infrared range (400–2500 nm). Chlorophyll a concentration is extracted using inversion algorithms (such as the maximum chlorophyll index method), achieving a resolution of 10 meters, thus meeting the refined monitoring needs of nearshore aquaculture areas. Simultaneously, synthetic aperture radar (SAR) is used to acquire sea surface wind speed and wave height data. By measuring changes in the sea surface backscattering coefficient and combining this with empirical models of wind speed and wave height (such as CMOD5 or WINDSAT models), it enables all-weather, all-time monitoring of the sea surface dynamic environment, particularly suitable for cloud cover or nighttime scenarios.

[0032] Furthermore, hyperspectral satellites (such as HY-1D) are used to collect spectral characteristic data of algae, with a spectral resolution of 5–10 nm, covering the 300–1000 nm wavelength range. This allows for the identification of reflectance characteristic curves of different algae, thus enabling preliminary classification of red tide algae species. This data complements multispectral chlorophyll data, providing key input features for subsequent red tide identification models based on transfer learning.

[0033] Furthermore, multispectral satellite data must meet a spatial resolution of 10 meters and a daily revisit frequency, SAR data must have a resolution of 20 meters and a revisit frequency of 2–3 days, and hyperspectral data must have a spectral resolution of 5–10 nm and a spatial resolution of 10–30 meters. All data must undergo atmospheric correction, geometric correction, and timestamp alignment to ensure consistency of multi-source data in the spatiotemporal dimensions.

[0034] Furthermore, this step is primarily used for environmental status sensing in nearshore jacket aquaculture areas, especially during peak red tide seasons or in areas with low oxygen risk. Through the rapid acquisition and fusion of remote sensing data, macroscopic environmental background information is provided to the system. Combined with in-situ sensor data from the jacket, early identification and risk assessment of red tide and low oxygen events can be achieved.

[0035] Furthermore, the technical effect of this step is that, through the collaborative acquisition of multi-source remote sensing data, the spatial resolution and temporal continuity of environmental parameters are significantly improved, providing high-quality input for subsequent DINEOF interpolation and transfer learning models, thereby enhancing the accuracy of red tide identification and the timeliness of hypoxia prediction. This is a key foundation for realizing the "air-space-sea" three-dimensional monitoring architecture.

[0036] S12 uses an underwater camera deployed on a jacket to collect algae density images at a frequency of 3 frames per minute, and transmits the image data to the edge computing node after compression via the LoRaWAN protocol.

[0037] Specifically, in some implementations, the steps of this invention involve acquiring algae density images at a frequency of 3 frames per minute using an underwater camera deployed on a jacket structure, and then transmitting the compressed image data to an edge computing node via the LoRaWAN protocol. This achieves real-time sensing and efficient transmission of dynamic changes in algae in nearshore aquaculture areas. The technical implementation of this step is based on the characteristics of the low-power, low-bandwidth underwater environment, employing customized image acquisition and transmission strategies to ensure the timeliness and integrity of the data.

[0038] Furthermore, underwater cameras typically employ industrial-grade equipment with an IP68 waterproof rating, equipped with low-light enhancement and automatic white balance adjustment to adapt to low-light and murky underwater environments. The cameras operate at a fixed frequency. Images are acquired at a rate of (frames per minute), which effectively controls the amount of data while ensuring data continuity and reducing the processing and transmission burden at the edge. The acquired image data is first compressed losslessly or near-losslessly using lightweight image compression algorithms (such as JPEG-LS or WebP) to reduce transmission bandwidth requirements while preserving key algal morphology and density information.

[0039] Furthermore, the resolution of image acquisition is typically set to... Pixels, Frame Rate Compression ratio controlled at to Between, to ensure the typical transmission rate of LoRaWAN ( Under these conditions, the transmission delay of a single frame image does not exceed [a certain value]. The LoRaWAN protocol supports a maximum packet length of [missing information]. Therefore, the image data needs to be segmented before transmission and reassembled and decoded at the edge computing nodes.

[0040] Furthermore, this step is applicable to intelligent aquaculture systems integrated into offshore wind turbine jacket platforms, especially in areas prone to red tides (such as nearshore waters and estuaries). Through high-frequency image acquisition and real-time edge processing, abnormal changes in algae density can be quickly identified, providing crucial input for subsequent red tide early warning. Cameras are typically deployed below the aquaculture cages. To capture the vertical distribution characteristics of algae in the water.

[0041] Furthermore, this step enables continuous, low-power monitoring of algal density, providing high spatiotemporal resolution in-situ image data for red tide early warning models. Through real-time processing by edge computing nodes, the system can... The system can complete image decoding and preliminary classification within the system, significantly improving the speed of early warning response, reducing data transmission energy consumption, and enhancing the robustness and practicality of the system in complex marine environments.

[0042] S2, based on the DINEOF algorithm, the satellite remote sensing data and the jacket sensor data are spatiotemporally interpolated to generate a dynamic risk heat map of the entire domain, and the cross-sea red tide characteristics are adapted and optimized by combining the transfer learning model.

[0043] Specifically, in some implementations, this invention employs the DINEOF algorithm to perform spatiotemporal interpolation processing on satellite remote sensing data and jacket sensor data to generate a global dynamic risk heat map. DINEOF is a data interpolation method based on empirical orthogonal functions (EOF). Its principle is to extract the main spatial modes of the data field through principal component analysis (PCA) and use these modes to reconstruct missing data. Specifically, the system first performs time alignment and spatial gridding processing on multi-source data (such as chlorophyll concentration data from Sentinel-2MSI, vertical velocity profile data from ADCP, and dissolved oxygen data from jacket sensors) to construct a unified spatiotemporal data matrix. Subsequently, the main spatial feature vectors are extracted through EOF decomposition and interpolated using time coefficients to fill in data gaps caused by cloud cover or equipment blind spots.

[0044] Furthermore, the DINEOF algorithm in this invention employs an iterative optimization strategy. In each iteration, the interpolation result is updated by minimizing the reconstruction error until a preset convergence condition (such as the error being less than 10%) is met. (Or the number of iterations exceeds 100). The interpolated data achieves a spatial resolution of 100 meters and a temporal resolution of 1 hour, significantly improving the spatial continuity and temporal dynamics of red tide and hypoxia risk.

[0045] Furthermore, regarding the adaptation and optimization of the transfer learning model, the system extracts spectral features of cross-oceanic red tides based on a ResNet-50 pre-trained network and performs localized optimization of the model through a fine-tuning strategy. Specifically, the model input consists of the chlorophyll concentration field after DINEOF interpolation and the spectral features of the underwater image, with the output being the probability distribution of red tide occurrence. During fine-tuning, the learning rate is set to... The batch size is 32, and the number of training rounds does not exceed 10, in order to avoid overfitting and improve the model's generalization ability.

[0046] Furthermore, by combining spatiotemporal interpolation with transfer learning, the model transfer problem caused by the differences in red tide characteristics across sea areas was effectively solved, improving the accuracy and applicability of the global risk heat map and providing high-quality data support for subsequent intelligent early warning and automated control.

[0047] Furthermore, S2 includes: S21. The DINEOF algorithm is used to interpolate the chlorophyll concentration data in the cloud-covered area to generate a dynamic risk heat map of the entire region.

[0048] Specifically, in some implementations, this invention employs the DINEOF algorithm to interpolate chlorophyll concentration data in cloud-covered areas to generate a dynamic risk heatmap covering the entire region. DINEOF is an interpolation method based on empirical orthogonal functions (EOF). Its principle involves performing singular value decomposition (SVD) on remote sensing data in the spatiotemporal domain to extract the main spatiotemporal feature modes, and then using these modes to reconstruct and interpolate missing areas. This method is particularly suitable for processing marine remote sensing data with spatiotemporal correlations, effectively reducing the impact of data loss due to cloud cover on the overall analysis accuracy.

[0049] Furthermore, the DINEOF algorithm first preprocesses chlorophyll concentration data from multispectral satellites (such as Sentinel-2MSI), including outlier removal, timestamp alignment, and spatial coordinate unification. Then, the data is organized into a spatiotemporal matrix, where rows represent time series and columns represent spatial grid points. By introducing an iterative optimization mechanism, DINEOF interpolates missing regions using known data points in each iteration and updates the EOF modes until the interpolation error converges to a preset threshold.

[0050] Furthermore, key parameters of the DINEOF algorithm include the number of EOF modes (typically the first 10-20 modes are selected to retain key information), the number of iterations (generally set to 50-100 times to ensure convergence), and the interpolation window size (5×5 or 7×7 pixels are recommended, corresponding to a spatial range of approximately 500-700 meters). In addition, the interpolation results must meet a spatial resolution of 100 meters to ensure consistency with the fusion accuracy of the jacket sensor data.

[0051] Furthermore, this interpolation step primarily addresses the missing data from satellites such as MODIS or Sentinel-2 in cloud-obscured areas, thereby providing a continuous and complete chlorophyll concentration distribution map for subsequent red tide risk index calculations. Combined with in-situ data collected from edge computing nodes, the DINEOF interpolation results can serve as input to the red tide early warning model, improving the accuracy and real-time performance of comprehensive risk assessment.

[0052] Furthermore, the introduction of the DINEOF algorithm significantly improves the spatiotemporal continuity of chlorophyll concentration data, making the generation of dynamic risk heatmaps across the entire region more reliable. Compared to traditional linear interpolation or Kriging methods, DINEOF effectively reduces interpolation errors while preserving the physical characteristics of the data, thus providing a high-quality data foundation for intelligent early warning of red tides / hypoxia.

[0053] S22, the ResNet-50 pre-trained network was fine-tuned using transfer learning techniques to extract the red tide spectral features from the field data and optimize the model parameters.

[0054] Specifically, in some implementations, this invention uses transfer learning techniques to fine-tune a ResNet-50 pre-trained network to extract red tide spectral features from field data and optimize model parameters, thereby improving the accuracy and adaptability of red tide identification. ResNet-50, as a deep convolutional neural network, uses large-scale general image datasets (such as ImageNet) for its original training and possesses powerful image feature extraction capabilities. In this invention, the network is redeployed for red tide spectral image recognition tasks, and through transfer learning strategies, the general features of the pre-trained model are transferred to specific red tide identification scenarios.

[0055] First, high-resolution (10-meter) sea surface spectral images were acquired from multispectral satellites (such as Sentinel-2MSI). Combined with an underwater hyperspectral imaging device deployed on a jacket platform, a training dataset containing spectral features such as chlorophyll a, algal pigments, and suspended particles was constructed. This dataset underwent preprocessing, including radiometric correction, atmospheric correction, band selection (such as blue, green, red, and near-infrared bands), and labeling (e.g., whether a red tide occurred, algal species, etc.).

[0056] Furthermore, during transfer learning, the first 48 convolutional layers of ResNet-50 were frozen to retain its general image feature extraction capabilities learned on ImageNet. Subsequently, for the red tide recognition task, the final fully connected layer was reconstructed, and a new classification head was introduced to adapt to the specific red tide category classification in this invention. The model was fine-tuned using the Adam optimizer, with a learning rate set to... The batch size is 32, and the training epochs are 20. The loss function used is cross-entropy loss, which measures the difference between the model output and the true label.

[0057] Furthermore, the model is iteratively trained using red tide spectral images collected on-site, gradually adjusting the weight parameters of the final layer to improve its ability to identify the spectral characteristics of local algae. For example, to address the differences in algal species distribution between the Bohai Bay and the South China Sea, a regional feature enhancement module is introduced during the fine-tuning process. By weighted fusion of spectral response curves from different sea areas, the model's ability to generalize to cross-regional red tide characteristics is enhanced.

[0058] Furthermore, its technical advantages are reflected in two aspects: firstly, transfer learning significantly reduces the amount of data and computational resources required for model training, thereby improving the model's convergence speed; secondly, the fine-tuned ResNet-50 model can more accurately extract red tide spectral features, providing high-quality input data for subsequent risk index calculation and early warning decisions. In practical applications, the model is deployed on a cloud server, with model weights updated hourly to adapt to environmental changes and ensure the real-time performance and accuracy of red tide identification.

[0059] S3 utilizes a lightweight convolutional neural network deployed on edge computing nodes to perform real-time algae classification on the underwater image data. Simultaneously, a low-oxygen prediction model including dissolved oxygen gradient, feeding amount, and ocean current stagnation index is constructed in the cloud through a long short-term memory network, and the prediction threshold is dynamically updated.

[0060] Specifically, in some implementations, the present invention uses a lightweight convolutional neural network (CNN) deployed on edge computing nodes to perform real-time algae classification on underwater image data, while using a long short-term memory network (LSTM) to build a hypoxia prediction model in the cloud and dynamically update the prediction threshold to achieve intelligent early warning and response to red tides and hypoxia disasters.

[0061] Furthermore, the edge computing nodes employ embedded devices to deploy lightweight CNN models, enabling low-latency classification of image data captured by underwater cameras. This model, pre-trained and fine-tuned on a field dataset, possesses the ability to identify common red tide algae (such as dinoflagellates and diatoms). The image input size is [size missing]. Pixels, using RGB three-channel format, model inference latency controlled within This meets the needs of real-time monitoring. The classification results include algal species and their relative densities, which are used to assist in the calculation of the red tide risk index.

[0062] Furthermore, in the cloud, the LSTM network receives multi-source time-series data from edge nodes and sensors, including dissolved oxygen vertical gradient, feed quantity, and ocean current stagnation index (defined as current velocity below a certain threshold). (duration). Model input dimension is ,in The time step is typically 24 hours, and the output is the probability of low oxygen levels for the next 72 hours. The LSTM model dynamically updates its weights hourly to adapt to nonlinear changes and seasonal fluctuations in environmental parameters, thereby optimizing the prediction threshold and improving the accuracy of early warnings.

[0063] Furthermore, the dissolved oxygen gradient is adopted The difference, in units of Feeding amount is defined as the mass of feed fed per unit time (unit: The current stagnation index is calculated using ADCP data. The cloud model is updated at a frequency of [missing information]. The prediction window is The output is the probability of hypoxia. .

[0064] Furthermore, this step is applicable to the intelligent early warning system of the guide frame aquaculture platform, especially under extreme weather conditions such as typhoons and cold waves. It can quickly identify algae species and predict low-oxygen diffusion trends, providing data support for emergency measures such as automatically controlling aerators and adjusting cage depth. Through edge and cloud collaborative computing, the system improves prediction accuracy while ensuring real-time performance, and controls response latency within a certain range. This method is significantly superior to the traditional fixed threshold method.

[0065] Furthermore, a lightweight CNN is used to achieve low-power, low-latency algae identification, and an LSTM model is combined to model multi-factor time-series data, effectively improving the prediction capabilities for red tides and hypoxia disasters. A dynamic prediction threshold update mechanism can adapt to the environmental characteristics of different sea areas, reducing false alarm rates and improving the system's intelligence and practicality.

[0066] Furthermore, S3 includes: S31 uses a lightweight MobileNetV3 convolutional neural network to classify underwater images in real time, ensuring that the inference latency of a single frame image is less than 200ms.

[0067] Specifically, in some implementations, this invention employs a lightweight MobileNetV3 convolutional neural network for real-time classification of underwater images to achieve rapid identification of algae species, thereby providing crucial input for red tide early warning. This step is technically implemented using an embedded AI inference framework deployed on edge computing nodes. By optimizing the model structure and computational efficiency, the inference latency for a single frame image is ensured to be less than 200ms.

[0068] Furthermore, MobileNetV3 employs depthwise separable convolutions and a linear bottleneck structure, significantly reducing the model's computational complexity while maintaining high classification accuracy. Its network structure includes multiple inverse residual modules, achieving efficient feature extraction through an expansion-convolution-compression process. In this invention, MobileNetV3 is pre-trained on the ImageNet dataset and fine-tuned for underwater algae image datasets to adapt to uneven lighting and color distortion issues in underwater environments. The image input size is 224×224×3, in RGB format, and is fed into the model after normalization.

[0069] Furthermore, model inference latency is controlled within 200ms, primarily relying on hardware acceleration (such as the NVIDIA Volta GPU in Jetson AGXXavier) and model quantization techniques (such as INT8 quantization). The model has 3 input channels, and the number of output nodes is set according to the number of target algae species (e.g., 5 common red tide algae species). During inference, the model's computational cost (FLOPs) is controlled below 1.2G, and the number of parameters (Params) is less than 5M to meet the real-time and power consumption requirements of edge devices. The classification confidence threshold is set to 0.7; prediction results exceeding this threshold will be used for subsequent risk index calculations.

[0070] Furthermore, this step is deployed on the edge computing nodes of the guide tube aquaculture platform, working in real-time with underwater cameras. The cameras acquire underwater images at 30fps. After preprocessing by the edge node's module (including white balance correction and histogram equalization), the image data is classified by the MobileNetV3 model. This model runs on an embedded Linux system and uses TensorRT for inference acceleration, ensuring high throughput and low latency even in resource-constrained environments.

[0071] Furthermore, the technical advantage of this step lies in achieving rapid identification of underwater algae through a lightweight CNN model, providing real-time and accurate species information for red tide early warning, thereby enhancing the system's ability to assess red tide risks. Combined with subsequent multi-source data fusion and risk index calculation, this step constitutes a key link in the closed loop of "image recognition-risk assessment-emergency response" in the entire early warning system, demonstrating significant practical value and innovation.

[0072] S32 uses a cloud-based LSTM neural network model to update the predicted weights of dissolved oxygen gradient, feeding amount, and ocean current stagnation index hourly to adapt to environmental changes.

[0073] Specifically, in some implementations, the present invention uses an LSTM neural network model deployed in the cloud to dynamically update the prediction weights of dissolved oxygen gradient, feeding amount and ocean current stagnation index every hour, so as to achieve real-time adaptation and accurate prediction of hypoxia risk.

[0074] Furthermore, the LSTM model receives multi-source time-series data from the edge computing layer, including dissolved oxygen vertical gradient, feeding amount, and current stagnation index, where SI is defined as current velocity below a certain threshold. The duration of the data is recorded. This data is uploaded to a cloud server via the LoRaWAN protocol and, after timestamp alignment, fed into the LSTM model. The model structure typically contains 2-3 layers of LSTM units, with each layer containing between 64 and 128 units to balance computational efficiency and prediction accuracy. The model is trained using the Adam optimizer, with a learning rate set to... The mean squared error (MSE) is used as the loss function.

[0075] Furthermore, the input dimension of the dissolved oxygen gradient is... ,in The time step is typically 60 minutes. The number of sensors deployed on the jacket is typically 5-8, distributed across different water depths. Feeding data is input using the cumulative hourly feed rate (kg / h), while the ocean current stagnation index is expressed as a time series in minutes. The model is fine-tuned online hourly based on the latest data, updating the prediction weights of each factor to adapt to dynamic environmental changes caused by weather, tides, or aquaculture activities.

[0076] Furthermore, this step is primarily used for real-time prediction and early warning decision support for hypoxia risk. For example, when dissolved oxygen concentration in aquaculture areas decreases, feeding increases, or ocean current stagnation intensifies, the model can quickly adjust weights to improve its sensitivity to potential hypoxia events. This mechanism is particularly suitable for scenarios prone to hypoxia, such as after typhoons, low-velocity currents at night, or high-density feeding.

[0077] Furthermore, this step significantly improves the accuracy and response speed of hypoxia early warning. Through dynamic weight adjustment, the model can more accurately capture environmental change trends, thereby reducing the false alarm rate of hypoxia prediction to below 10% and ensuring a response delay of less than 10 minutes.

[0078] S4, based on the quantitative results of the red tide risk index and the hypoxia risk index, uses the ModbusRTU protocol to link the aerator, feeding system and jacket hydraulic device, and performs feeding amount adjustment, aerator start-up or jacket lowering operation according to the preset graded response strategy.

[0079] Specifically, in the execution control layer of this invention, based on the quantification results of the red tide risk index and the hypoxia risk index, the system uses the Modbus RTU protocol to achieve coordinated control of the aerator, feeding system, and jacket hydraulic device, thereby executing a graded response strategy. This step is based on the deep integration of industrial automation control standards and intelligent decision-making logic.

[0080] Furthermore, the Modbus RTU protocol, as the core communication mechanism of this system, operates at the RS-485 physical layer and adopts a master-slave communication architecture, supporting point-to-point or one-to-many control command issuance. The system master control unit (such as an edge computing node or cloud server) communicates with the execution device through a serial port or industrial Ethernet gateway, sending Modbus frames containing device address, function code, register address, and data value. For example, when the red tide risk index reaches the yellow warning threshold (0.4-0.6), the system will send a control command to the feeding system to adjust the feeding amount to 50% of the current set value, and simultaneously send a start signal to the aerator, setting its oxygen-to-liquid ratio to 1:5 to ensure that the oxygen supply per unit time meets the oxygenation needs of the aquaculture water.

[0081] Furthermore, when the hypoxia risk index exceeds the red warning threshold (>0.6), the system will trigger the sinking operation of the jacket hydraulic system, sending a target depth command (e.g., 20 meters) to the hydraulic controller via the Modbus RTU protocol, and simultaneously shutting down the feeding system to prevent organic load from further exacerbating water hypoxia. This sinking depth can be calculated using the formula... Where DO_{def} represents the deviation of dissolved oxygen from the threshold. The ocean current stagnation index, This is an empirical coefficient for the sea area. This formula is used to dynamically adjust the disaster avoidance depth of the jacket structure to adapt to the hydrodynamic conditions of different sea areas.

[0082] Furthermore, Modbus RTU communication baud rate is typically set to 9600 bps, with 8 data bits, 1 stop bit, and even parity to ensure communication stability and real-time performance. Response latency is controlled within 10 minutes to meet the rapid response needs of aquaculture environments to sudden disasters.

[0083] Furthermore, by working collaboratively with edge computing nodes and cloud servers, automated interventions for red tides and hypoxia disasters can be achieved. Its technological value lies in quantifying and mapping environmental risk indices to physical actions, significantly improving the accuracy and timeliness of disaster response, reducing the cost of manual intervention, and protecting the living environment and economic benefits of aquaculture organisms.

[0084] Furthermore, S4 includes: S41, when the red tide risk index reaches 0.4-0.6, the feeding amount is reduced by 50% and the aerator is started through the ModbusRTU protocol. The aerator adopts a pure oxygen to liquid ratio of 1:5.

[0085] Specifically, when the red tide risk index reaches the range of 0.4 to 0.6, the system will enter a yellow alert phase and trigger corresponding emergency control strategies via the Modbus RTU protocol. This step achieves initial intervention in the aquaculture environment to reduce the likelihood of red tide outbreaks and alleviate the risk of low oxygen levels.

[0086] Furthermore, the Modbus RTU protocol, a widely adopted serial communication protocol in industrial automation, is based on the RS-485 physical layer and boasts advantages such as strong anti-interference capabilities, long transmission distance (up to 1200 meters), and stable communication. The system monitors the red tide risk index in real time through edge computing nodes (such as NVIDIA Jetson AGXXavier). When the index reaches a set threshold, the edge node generates control commands, which are sent to the on-site feeding control system and aerator controller via the Modbus RTU protocol. Specifically, the feeding amount will be reduced by 50% to decrease the input of organic load into the water, thereby slowing down the excessive proliferation of algae. Simultaneously, the aerator will be activated, employing a pure oxygen-to-liquid ratio of 1:5, meaning that 1 unit volume of pure oxygen is mixed with 5 units of seawater to improve the diffusion efficiency and solubility of oxygen in the water.

[0087] Furthermore, the red tide risk index is calculated by combining chlorophyll concentration (weight 40%), algal toxicity coefficient (30%), meteorological stagnation index (20%), and historical disaster weight (10%), with a value range of 0 to 1, and 0.4 to 0.6 indicating moderate risk. The aerator's air supply mode parameters (air-liquid ratio 1:5) are the optimal configuration derived from a water dissolved oxygen kinetic model and actual aquaculture environment testing, ensuring maximum oxygenation effect with limited energy consumption.

[0088] Furthermore, this procedure is applicable to jacketed seawater aquaculture systems, particularly in areas with high stocking densities and weak water exchange capacity. Through the Modbus RTU protocol and linkage with field equipment, the system can complete closed-loop control from risk identification to equipment response within 10 minutes, significantly improving emergency response efficiency.

[0089] Furthermore, the technical effect of this step lies in the fact that by reducing the amount of feed and activating efficient aeration equipment, the system can effectively suppress the development of red tide risk at its initial stage, while improving dissolved oxygen levels in the water and reducing the risk of hypoxia. Its innovation lies in directly linking the risk index with Modbus RTU control commands and optimizing equipment operation strategies through quantitative parameters (such as gas-liquid ratio), thereby achieving intelligent and automated aquaculture environment control.

[0090] S42, when the low oxygen risk index exceeds 0.6, controls the hydraulic system of the jacket to sink to a disaster avoidance depth of 20 meters via the ModbusRTU protocol, and at the same time shuts down the feeding system.

[0091] Specifically, when the hypoxia risk index exceeds 0.6, the system will control the hydraulic system of the guide frame to descend to a disaster avoidance depth of 20 meters via the Modbus RTU protocol, and simultaneously shut down the feeding system. This step reduces the risk of death of farmed organisms due to hypoxia through automation, achieving a rapid closed-loop disaster emergency response.

[0092] Furthermore, the Modbus RTU protocol, a widely adopted serial communication protocol in industrial control, is based on the RS-485 physical layer and boasts advantages such as strong anti-interference capabilities, stable transmission, and suitability for long-distance communication. The system calculates a hypoxia risk index in real time through edge computing nodes. This index is dynamically weighted by dissolved oxygen at the bottom layer (50% weight), organic load (30% weight), and ocean current stratification intensity (20% weight). When the index value exceeds a threshold of 0.6, the system issues a red alert and triggers the Modbus RTU command transmission mechanism. The control signal sends a sinking command to the hydraulic control system through a preset register address (e.g., 0x0A01), setting the target depth to 20 meters. This depth is verified based on the hypoxia tolerance physiological limits of cultured organisms and historical disaster data, ensuring the survival of organisms even under extreme hypoxia conditions.

[0093] Furthermore, the Modbus RTU communication baud rate is set to 9600 bps, with 8 data bits and 1 stop bit, and even parity (E) to ensure communication reliability even in complex marine electromagnetic environments. The sinking process requires an execution accuracy of ±0.5 meters, and the response delay must be controlled within 10 minutes. The feeding system's shutdown operation is achieved through a Modbus RTU-controlled relay module, cutting off the power to the feeding motor to prevent further increase in the water's organic load, thereby mitigating the deterioration of hypoxia.

[0094] Furthermore, this procedure is applicable to jacketed deep-sea aquaculture systems, especially in areas prone to red tides or low oxygen levels (such as nearshore eutrophic waters and aquaculture areas with poor water exchange capacity). Upon receiving a red alert signal, the system immediately performs sinking and shutdown operations to avoid the risk of organism asphyxiation due to a sudden drop in dissolved oxygen. Simultaneously, this response mechanism is linked to the BeiDou short message system to send emergency instructions to management personnel, ensuring consistency between remote monitoring and on-site control.

[0095] Furthermore, the technical advantage of this step lies in achieving automation and precision in disaster response by quantifying the mapping relationship between risk indices and actions. Descending to a 20-meter refuge depth effectively avoids surface hypoxic zones, while shutting down the feeding system controls organic matter input at the source, forming a dual protection mechanism.

[0096] The red tide / hypoxia intelligent early warning method based on multi-source data fusion in this invention can achieve high-precision real-time early warning and automated emergency response for red tide and hypoxia disasters, effectively reducing the false alarm rate and improving the timeliness and pertinence of disaster prevention and control.

[0097] Example 2 The red tide / hypoxia early warning system of this invention adopts a three-dimensional "air-space-sea" monitoring architecture, achieving precise early warning and automated emergency response through multi-source data fusion and intelligent decision-making. The core of the system consists of four layers: The data acquisition layer integrates multispectral satellites (such as Sentinel-2MSI), synthetic aperture radar (SAR), and hyperspectral satellites (HY-1D) to acquire chlorophyll distribution and algal spectral characteristics at a resolution of 10 meters. Simultaneously, it constructs an in-situ monitoring network through dissolved oxygen sensors, ADCP vertical current profilers, and underwater cameras deployed on the jacket structure to collect water quality parameters and algal density images in real time. The edge computing layer performs data preprocessing on embedded nodes on the jacket structure (such as NVIDIA Jetson AGXXavier), including outlier filtering, timestamp alignment, and real-time algal species classification of underwater images using a lightweight CNN model (MobileNetV3). After compression, the data is transmitted to the cloud via LoRaWAN with low power consumption.

[0098] The cloud-based fusion analysis layer uses the DINEOF algorithm to perform spatiotemporal interpolation on satellite and sensor data, filling in data gaps in cloud-obscured areas. It also incorporates transfer learning techniques to optimize the red tide prediction model: extracting cross-ocean red tide spectral features based on a ResNet-50 pre-trained network, and fine-tuning the model's adaptability using on-site data. The hypoxia early warning system relies on an LSTM neural network, inputting dissolved oxygen vertical gradient, feeding amount, and ocean current stagnation index (duration of current velocity below 0.1 m / s) to predict the probability of hypoxia in the next 72 hours. In the risk grading rules, the red tide risk index integrates chlorophyll concentration (weight 40%), algal toxicity coefficient (30%), meteorological stagnation index (20%), and historical disaster weight (10%), while the hypoxia risk index is dynamically calculated from bottom dissolved oxygen (50%), organic load (30%), and ocean current stratification intensity (20%).

[0099] The execution control layer implements a tiered response strategy: During a yellow alert (index 0.4-0.6), the feeding amount is automatically reduced by 50% and the aerator is activated (pure oxygen to liquid ratio 1:5); during a red alert (index > 0.6), the jacket structure is hydraulically lowered to a disaster avoidance depth of 20 meters, the feeding system is simultaneously shut down, and emergency instructions are pushed via BeiDou short message service. The entire system uses the Modbus RTU protocol to link hardware devices, forming a closed loop of "monitoring-decision-execution," ensuring a response delay of less than 10 minutes.

[0100] The core technological breakthrough of this invention lies in the multi-source data collaboration mechanism and the edge-cloud collaborative computing architecture. The former utilizes closed-loop calibration of satellite remote sensing and jacket sensors, employing the DINEOF algorithm to generate a risk heat map with 100-meter accuracy, and combines a transfer learning model to address the issue of cross-oceanic red tide characteristic differences (such as the difference in algal species distribution between the Bohai Bay and the South China Sea). The latter deploys a lightweight CNN at the edge to achieve real-time algal classification (inference latency <200ms), while the cloud uses LSTM to dynamically optimize prediction thresholds and updates model weights hourly to adapt to environmental changes. The key features include: 1) a data fusion method, including red tide heat map generation technology based on DINEOF and transfer learning, and an LSTM hypoxia prediction model integrating feeding factors; 2) a hardware deployment scheme for an air-space-sea integrated monitoring system, particularly emphasizing the collaborative design of vertical ADCP and edge computing nodes; and 3) closed-loop control logic, clearly defining the quantitative mapping relationship between the risk index and the jacket sinking depth and aerator power. For example, when dissolved oxygen is below 3 mg / L and the flow rate remains stagnant, the system automatically calculates the sinking depth (…). (where K is the sea area empirical coefficient).

[0101] Example 3 To achieve the above embodiments, such as Figure 2 As shown, this embodiment also provides a red tide / hypoxia intelligent early warning device 10 based on multi-source data fusion, including: Data acquisition module 100 is used to acquire satellite remote sensing data, jacket sensor data and underwater image data; The spatiotemporal interpolation and transfer learning module 200 is used to perform spatiotemporal interpolation processing on the satellite remote sensing data and jacket sensor data based on the DINEOF algorithm, generate a dynamic risk heat map of the entire domain, and combine the transfer learning model to adapt and optimize the characteristics of red tides across the sea area. The edge computing and cloud prediction module 300 is used to perform real-time algae classification on the underwater image data using a lightweight convolutional neural network deployed on edge computing nodes. At the same time, it constructs a hypoxia prediction model in the cloud through a long short-term memory network, which includes dissolved oxygen gradient, feeding amount and ocean current stagnation index, and dynamically updates the prediction threshold. The linkage control module 400 is used to link the aerator, feeding system and jacket hydraulic device through ModbusRTU protocol according to the quantitative results of red tide risk index and hypoxia risk index, and to perform feeding amount adjustment, aerator start-up or jacket lowering operation according to preset graded response strategy.

[0102] Furthermore, the data acquisition module 100 is also used for: Chlorophyll distribution data with a resolution of 10 meters was obtained through multispectral satellites, sea surface wind speed and wave height data were obtained through synthetic aperture radar, and algal spectral characteristic data were obtained through hyperspectral satellites. The underwater camera deployed on the jacket structure acquires algae density images at a frequency of 3 frames per minute, and transmits the compressed image data to the edge computing node via the LoRaWAN protocol; The satellite remote sensing data includes multispectral, synthetic aperture radar, and hyperspectral data, while the jacket sensor data includes dissolved oxygen vertical distribution data and ocean current velocity data.

[0103] Furthermore, the spatiotemporal interpolation and transfer learning module 200 is also used for: The DINEOF algorithm was used to interpolate chlorophyll concentration data in cloud-covered areas to generate a dynamic risk heat map of the entire region. The ResNet-50 pre-trained network was fine-tuned using transfer learning techniques to extract red tide spectral features from field data and optimize model parameters.

[0104] The present invention discloses a red tide / hypoxia intelligent early warning method and device based on multi-source data fusion, which can realize high-precision dynamic early warning and automated emergency response for red tide and hypoxia disasters, significantly improve the accuracy of early warning and shorten the response time, and effectively form a closed-loop control of "monitoring-decision-execution".

[0105] Example 4 To implement the methods of the above embodiments, the present invention also provides a computer device, such as... Figure 3 As shown, the computer device 600 includes a memory 601 and a processor 602; wherein, the processor 602 reads the executable program code stored in the memory 601 to run a program corresponding to the executable program code, so as to implement the various steps of the red tide / hypoxia intelligent early warning method based on multi-source data fusion described above.

[0106] Example 5 To implement the above embodiments, this application also proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a red tide / hypoxia intelligent early warning method based on multi-source data fusion as described in the foregoing embodiments.

[0107] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0108] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A red tide / hypoxia intelligent early warning method based on multi-source data fusion, characterized in that, include: S1, acquires satellite remote sensing data, jacket sensor data, and underwater image data; S2, based on the DINEOF algorithm, the satellite remote sensing data and the jacket sensor data are spatiotemporally interpolated to generate a dynamic risk heat map of the entire domain, and the cross-sea red tide characteristics are adapted and optimized by combining the transfer learning model. S3, a lightweight convolutional neural network deployed on edge computing nodes is used to perform real-time algae classification on the underwater image data. At the same time, a low oxygen prediction model including dissolved oxygen gradient, feeding amount and ocean current stagnation index is built in the cloud through a long short-term memory network, and the prediction threshold is dynamically updated. S4, based on the quantitative results of the red tide risk index and the hypoxia risk index, uses the ModbusRTU protocol to link the aerator, feeding system and jacket hydraulic device, and performs feeding amount adjustment, aerator start-up or jacket lowering operation according to the preset graded response strategy.

2. The method as described in claim 1, characterized in that, S1 includes: S11 acquires chlorophyll distribution data at a resolution of 10 meters through multispectral satellites, sea surface wind speed and wave height data through synthetic aperture radar, and algal spectral characteristic data through hyperspectral satellites. S12 uses an underwater camera deployed on the jacket to collect algae density images at a frequency of 3 frames per minute, and transmits the image data to the edge computing node after compression via the LoRaWAN protocol; S13, the satellite remote sensing data includes multispectral, synthetic aperture radar and hyperspectral data, and the jacket sensor data includes dissolved oxygen vertical distribution data and ocean current velocity data.

3. The method as described in claim 1, characterized in that, The S2 includes: S21. The DINEOF algorithm is used to interpolate the chlorophyll concentration data in the cloud-covered area to generate a dynamic risk heat map of the entire region. S22, the ResNet-50 pre-trained network was fine-tuned using transfer learning techniques to extract the red tide spectral features from the field data and optimize the model parameters.

4. The method as described in claim 1, characterized in that, The S3 includes: S31 uses a lightweight MobileNetV3 convolutional neural network to classify underwater images in real time, ensuring that the inference latency of a single frame image is less than 200ms. S32 uses a cloud-based LSTM neural network model to update the predicted weights of dissolved oxygen gradient, feeding amount, and ocean current stagnation index hourly to adapt to environmental changes.

5. The method as described in claim 1, characterized in that, The S4 includes: S41, when the red tide risk index reaches 0.4-0.6, reduce the feeding amount by 50% and start the aerator through the ModbusRTU protocol. The aerator adopts a pure oxygen to liquid ratio of 1:

5. S42, when the low oxygen risk index exceeds 0.6, controls the hydraulic system of the jacket to sink to a disaster avoidance depth of 20 meters via the ModbusRTU protocol, and at the same time shuts down the feeding system.

6. A red tide / hypoxia intelligent early warning device based on multi-source data fusion, characterized in that, include: The data acquisition module is used to acquire satellite remote sensing data, jacket sensor data, and underwater image data; The spatiotemporal interpolation and transfer learning module is used to perform spatiotemporal interpolation processing on the satellite remote sensing data and jacket sensor data based on the DINEOF algorithm, generate a dynamic risk heat map of the entire domain, and combine the transfer learning model to adapt and optimize the characteristics of red tides across the sea area. The edge computing and cloud prediction module is used to perform real-time algae classification on the underwater image data using a lightweight convolutional neural network deployed on edge computing nodes. At the same time, it constructs a hypoxia prediction model in the cloud through a long short-term memory network, which includes dissolved oxygen gradient, feeding amount and ocean current stagnation index, and dynamically updates the prediction threshold. The linkage control module is used to link the aerator, feeding system and jacket hydraulic device through ModbusRTU protocol according to the quantitative results of red tide risk index and hypoxia risk index, and to perform feeding amount adjustment, aerator start-up or jacket lowering operation according to preset graded response strategy.

7. The apparatus as claimed in claim 6, characterized in that, The data acquisition module is also used for: Chlorophyll distribution data with a resolution of 10 meters was obtained through multispectral satellites, sea surface wind speed and wave height data were obtained through synthetic aperture radar, and algal spectral characteristic data were obtained through hyperspectral satellites. The underwater camera deployed on the jacket structure acquires algae density images at a frequency of 3 frames per minute, and transmits the compressed image data to the edge computing node via the LoRaWAN protocol; The satellite remote sensing data includes multispectral, synthetic aperture radar, and hyperspectral data, while the jacket sensor data includes dissolved oxygen vertical distribution data and ocean current velocity data.

8. The apparatus as claimed in claim 6, characterized in that, The spatiotemporal interpolation and transfer learning module is also used for: The DINEOF algorithm was used to interpolate chlorophyll concentration data in cloud-covered areas to generate a dynamic risk heat map of the entire region. The ResNet-50 pre-trained network was fine-tuned using transfer learning techniques to extract red tide spectral features from field data and optimize model parameters.

9. A computer device, characterized in that, Including processor and memory; The processor reads executable program code stored in the memory to run a program corresponding to the executable program code, so as to implement a red tide / hypoxia intelligent early warning method based on multi-source data fusion as described in any one of claims 1-5.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements a red tide / hypoxia intelligent early warning method based on multi-source data fusion as described in any one of claims 1-5.

Images