An early warning method for marine ecological disasters based on the integration of multi-source heterogeneous big data

Abstract

The present invention discloses a marine ecological disaster early warning method based on multi-source heterogeneous big data fusion. The method comprises the following steps: separately acquiring remote sensing image data and measured environmental data of a target sea area; sequentially performing boundary segmentation and feature fusion on the remote sensing image data using an image segmentation network to obtain multispectral feature data; performing spatiotemporal alignment of the multispectral feature data with the environmental data to obtain a spatiotemporal aligned dataset; mapping each sample pair in the spatiotemporal aligned dataset into graph node data, analyzing the connection relationship between any two adjacent graph node data to construct an adjacency matrix; performing neighborhood aggregation processing on the adjacency matrix to generate a fused feature vector; inputting the fused feature vector into a pre-built marine disaster prediction model to obtain a multidimensional prediction result, and providing real-time ecological disaster early warning based on the multidimensional prediction result. The method of the present invention improves the real-time performance of marine ecological disaster early warning.

Description

Technical Field

[0001] The present invention relates to the technical field of marine environment monitoring, and in particular to a marine ecological disaster early warning method based on multi-source heterogeneous big data fusion. Background Art

[0002] Marine ecological disasters seriously threaten the safety of marine ecosystems and the sustainable development of mankind. Their occurrence is characterized by suddenness, regionality, and temporal and spatial heterogeneity, requiring accurate real-time early warning technology.

[0003] Existing technology uses the spectral characteristics of remote sensing images to identify algae accumulation areas. However, this method can only capture local features within a single scene. For example, in the case of red tide warnings, existing technology can only determine the approximate distribution of algae but cannot understand the dynamic changes in key environmental factors within the water body. This results in low accuracy in disaster predictions, which in turn affects the accurate assessment of key information such as the probability of disaster occurrence and the scope of impact, and fails to meet the needs of real-time prevention and control of marine ecological disasters. Summary of the Invention

[0004] The present invention provides a marine ecological disaster early warning method based on the fusion of multi-source heterogeneous big data to solve the technical problem of how to improve the existing marine ecological disaster early warning method and achieve the effect of improving the real-time performance of marine ecological disaster early warning.

[0005] To solve the above technical problems, an embodiment of the present invention provides a marine ecological disaster early warning method based on multi-source heterogeneous big data fusion, including:

[0006] Obtain remote sensing image data and measured environmental data of the target sea area respectively;

[0007] Constructing an image segmentation network including an attention mechanism and a multi-scale feature fusion module, and sequentially performing boundary segmentation and feature fusion on the remote sensing image data based on the image segmentation network to obtain optimized multispectral feature data;

[0008] Performing spatiotemporal alignment on the multispectral feature data and the environmental data according to a dynamic spatiotemporal calibration algorithm to obtain a spatiotemporal aligned data set;

[0009] Mapping each sample pair in the spatiotemporal alignment dataset into graph node data, analyzing the connection relationship between any two adjacent graph node data, and constructing an adjacency relationship matrix based on the analysis results, wherein the connection relationship reflects the feature correlation and spatiotemporal correlation between the graph node data;

[0010] Performing neighborhood aggregation processing on the adjacency relationship matrix based on a pre-built graph convolutional network to generate a fused feature vector, wherein the fused feature vector includes a spatiotemporal dependency relationship between the spectral feature data and the environmental data;

[0011] A marine disaster prediction model is constructed according to the cross-domain transfer learning method, and the fused feature vector is input into the marine disaster prediction model to obtain a multi-dimensional prediction result reflecting the marine ecological disaster warning. Based on the multi-dimensional prediction result, a real-time warning of ecological disasters in the target sea area is performed.

[0012] As one of the preferred solutions, the remote sensing image data is sequentially subjected to boundary segmentation and feature fusion based on the image segmentation network to obtain optimized multispectral feature data, including:

[0013] Extracting features from the multispectral band data in the remote sensing image data based on the image segmentation network to obtain multi-scale spectral features, wherein the multi-scale spectral features include shallow spectral features and deep spectral features;

[0014] The shallow spectral features and the deep spectral features are spliced using a channel splicing method to obtain a red tide probability map with the same resolution as the remote sensing image data;

[0015] Performing double-threshold segmentation on the red tide probability map to obtain red tide boundary coordinates, wherein the double-threshold segmentation is designed to extract a core red tide area based on a first segmentation threshold and to extract a fuzzy transition boundary between the red tide and normal seawater based on a second segmentation threshold;

[0016] The red tide boundary coordinates are spatially associated with the remote sensing image data to obtain multispectral feature data including multispectral data and the red tide boundary coordinates.

[0017] As one preferred solution, the channel splicing method is used to splice the shallow spectral features and the deep spectral features to obtain a red tide probability map with the same resolution as the remote sensing image data, including:

[0018] Adjusting the spatial dimensions of the shallow spectral feature and the deep spectral feature to obtain a first shallow spectral feature and a first deep spectral feature having the same resolution as the remote sensing image data;

[0019] Assigning corresponding adaptive weights to the first shallow spectral feature and the first deep spectral feature respectively to obtain a second shallow spectral feature and a second deep spectral feature, wherein the adaptive weights reflect the importance of the shallow spectral feature and the deep spectral feature in the red tide probability prediction;

[0020] splicing the second shallow spectral feature and the second deep spectral feature in the channel dimension to obtain a red tide probability characteristic map;

[0021] The probability that each pixel point in the red tide probability characteristic map is a red tide area is calculated to obtain the red tide probability map.

[0022] As one preferred solution, the step of performing spatiotemporal alignment on the multispectral feature data and the environmental data according to a dynamic spatiotemporal calibration algorithm to obtain a spatiotemporal aligned dataset includes:

[0023] Based on the passing time of the remote sensing satellite, the collection time of the environmental data is dynamically resampled by a linear interpolation algorithm to generate first environmental data that is synchronized with the multispectral feature data;

[0024] Based on the acquired geocoding parameters of the multispectral feature data, spatially registering the collection location of the environmental data in the first environmental data, and mapping the first environmental data to a spatial grid with the same resolution as the multispectral feature data to obtain second environmental data;

[0025] The second environmental data is subjected to rationality verification to obtain the spatiotemporal alignment dataset.

[0026] As one of the preferred solutions, the measured environmental data of the target sea area is obtained, including:

[0027] Acquiring initial environmental data of the target sea area based on a sensor, performing outlier detection on the initial environmental data using an outlier recognition algorithm based on density peak clustering, and obtaining first environmental data from which data deviating from a normal distribution is eliminated;

[0028] Performing filtering on the first environmental data using a Kalman filter algorithm to obtain smoothed second environmental data;

[0029] Feature extraction is performed on the second environmental data to construct a standardized data sample including a timestamp, a spatial position and characteristic parameters, wherein the characteristic parameters include a temperature change rate, a salinity gradient, a dissolved oxygen content fluctuation amplitude and a pH value anomaly coefficient.

[0030] Mapping each sample pair in the spatiotemporal alignment dataset into graph node data, analyzing the connection relationship between any two adjacent graph node data, and constructing an adjacency relationship matrix based on the analysis results, includes:

[0031] As one of the preferred solutions, each sample pair data in the spatiotemporal alignment dataset is mapped to graph node data in a graph structure, and each graph node data includes spatial attributes, temporal attributes and feature attributes;

[0032] Calculate the spatial distance of the spatial attributes of any two adjacent graph node data based on the Euclidean distance formula;

[0033] Calculate the feature similarity of the feature attributes of any two adjacent graph node data according to the Pearson correlation coefficient;

[0034] The connection relationship between any two adjacent graph node data is determined based on the spatial distance and the feature similarity, and an adjacency relationship matrix is constructed according to the connection relationship.

[0035] As one of the preferred solutions, the method of determining the connection relationship between any two adjacent graph node data based on the spatial distance and the feature similarity, and constructing an adjacency relationship matrix according to the connection relationship, includes:

[0036] Determine the spatial distance threshold and feature similarity threshold of the target sea area based on the geographical environment characteristics of the target sea area, the seasonal variation pattern of the marine ecosystem, and historical marine ecological disaster data;

[0037] Performing a preliminary judgment on the connection relationship between any two adjacent graph node data according to the spatial distance threshold and the feature similarity threshold;

[0038] If the preliminary judgment result is that there is a connection relationship between the two graph node data, the connection strength between the two graph node data is calculated based on the weighted summation method to generate an adjacency relationship matrix reflecting the spatiotemporal correlation and feature correlation between the graph node data.

[0039] As one of the preferred solutions, the construction of a marine disaster prediction model based on a cross-domain transfer learning method includes:

[0040] Extracting the network structure and training parameters of the land disaster prediction model as a basic marine disaster prediction model, and embedding a fully connected layer and a domain adversarial network module for marine ecological disaster prediction into the basic marine disaster prediction model to generate a first marine disaster prediction model;

[0041] Acquire historical data of marine ecological disasters as target domain data, and acquire historical data of land disasters as source domain data, and correct the parameters of the first marine disaster prediction model by minimizing the difference in feature distribution between the source domain data and the target domain data to obtain a second marine disaster prediction model;

[0042] Performing data enhancement processing on the marine ecological disaster historical data to obtain second marine ecological disaster historical data;

[0043] The second marine disaster prediction model is trained based on the second marine ecological disaster historical data, and the weight parameters of the second marine disaster prediction model are updated by optimizing the loss function to obtain a trained marine disaster prediction model.

[0044] As one preferred solution, performing data enhancement processing on the marine ecological disaster historical data to obtain the second marine ecological disaster historical data includes:

[0045] Adding Gaussian noise with a zero mean and adaptive standard deviation to the time series feature data in the marine ecological disaster historical data, and randomly shifting the spatial position coordinate data of the marine ecological disaster historical data to generate first enhanced data containing spatiotemporal disturbance;

[0046] Constructing a marine ecological data generation model based on a generative adversarial network, and generating second enhanced data that conforms to the characteristics of the target sea area based on the marine ecological data generation model, wherein the marine ecological data generation model is designed to learn the distribution pattern of marine ecological data by inputting feature vectors of real disaster samples;

[0047] The first enhanced data and the second enhanced data are validated, and the first enhanced data and the second enhanced data that have passed the validation are merged to obtain the second marine ecological disaster history data.

[0048] As one preferred solution, the real-time early warning of ecological disasters in the target sea area based on the multi-dimensional prediction results includes:

[0049] Obtain real-time hydrological and meteorological data of the target sea area;

[0050] Inputting the multi-dimensional prediction results, the real-time hydrological data, and the meteorological data into a pre-built disaster assessment model, and determining the risk level of ecological disasters based on the output results of the model;

[0051] Obtaining the ecologically sensitive areas of the target sea area, and determining the degree of threat to the ecologically sensitive areas based on the disaster impact range in the multi-dimensional prediction results;

[0052] Execute the early warning control instruction generated by the risk level and the threat degree.

[0053] Compared with the prior art, the embodiments of the present invention have the following advantages:

[0054] 1) The present invention obtains remote sensing images and measured environmental data separately, processes the remote sensing images using an image segmentation network with an attention mechanism and a multi-scale feature fusion module, and then achieves spatiotemporal alignment of the data through a dynamic spatiotemporal calibration algorithm. It constructs a graph node and adjacency relationship matrix, and generates a fusion feature vector through a graph convolutional network. This method integrates the advantages of multi-source data, takes into account the spatiotemporal dependency of spectral features and environmental data, breaks through the limitations of single data, and can more comprehensively and deeply capture the complex mechanisms of marine ecological disasters, thereby improving the accuracy and reliability of disaster prediction models.

[0055] 2) The present invention uses the cross-domain transfer learning method to construct a marine disaster prediction model, and inputs the fused feature vector into the model to obtain multi-dimensional prediction results, covering key information such as the probability of disaster occurrence and the scope of impact. It can provide real-time early warning of ecological disasters in the target sea area, effectively meeting the real-time prevention and control needs of sudden marine ecological disasters. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 It is a flow chart of a marine ecological disaster early warning method based on multi-source heterogeneous big data fusion in one embodiment of the present invention. DETAILED DESCRIPTION

[0057] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. The purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0058] In the description of the present invention, the terms "first," "second," "third," etc. are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Therefore, a feature specified as "first," "second," "third," etc. may explicitly or implicitly include one or more of the features. In the description of the present invention, unless otherwise specified, "plurality" means two or more.

[0059] In the description of the present invention, it should be noted that, unless otherwise expressly specified and limited, the terms "installed", "connected" and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or an indirect connection through an intermediate medium, or it can be a communication between the two components. The terms "vertical", "horizontal", "left", "right", "up", "down" and similar expressions used herein are for illustrative purposes only, and do not indicate or imply that the device or component referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention. The term "and / or" used herein includes any and all combinations of one or more related listed items. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0060] In describing the present invention, it should be noted that, unless otherwise defined, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art. The terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. Those skilled in the art will understand the specific meanings of the above terms in the present invention in specific circumstances.

[0061] One embodiment of the present invention provides a marine ecological disaster early warning method based on the fusion of multi-source heterogeneous big data. For details, please refer to Figure 1 , Figure 1 The figure shows a flow chart of a marine ecological disaster early warning method based on multi-source heterogeneous big data fusion in one embodiment of the present invention, which includes steps S1 to S6:

[0062] S1: Obtain remote sensing image data and measured environmental data of the target sea area respectively;

[0063] Remote sensing image data refers to spectral radiation information on the surface of the target sea area acquired using electromagnetic wave sensors via remote sensing platforms such as satellites and drones. It can include multispectral or hyperspectral data and can reflect parameters such as sea surface temperature, algae distribution, and water turbidity. Measured environmental data refers to in-situ environmental data collected in real time by sensors deployed in the target sea area. It can include water quality parameters, meteorological data, and ocean dynamic parameters.

[0064] Understandably, remote sensing data alone only captures spectral information about the sea surface, such as the distribution of surface algae during red tides, but fails to reveal the dynamics of key environmental factors within the water column. Measured data alone, however, only reflects localized conditions, such as salinity at a specific buoy, making it difficult to construct spatial correlation models for disaster occurrence. Combining these two approaches creates a complete data chain linking "sea surface spectral characteristics, water environmental parameters, and disaster driving mechanisms," supporting disaster analysis.

[0065] Specifically, in this embodiment, the acquired remote sensing image data is preprocessed, including:

[0066] 1) Radiometric correction: removes the effects of atmospheric scattering and sensor noise to restore the true surface radiation brightness;

[0067] 2) Geometric correction: Correcting image geometric distortion based on ground control points or satellite attitude data;

[0068] 3) Band screening: Select sensitive bands based on the type of disaster. For example, red tide warnings focus on the reflectivity differences between the red light (650nm) and near-infrared (850nm) bands.

[0069] Preferably, in one embodiment of the present invention, obtaining measured environmental data of the target sea area includes:

[0070] Initial environmental data of the target sea area is obtained based on sensors, and an outlier recognition algorithm based on density peak clustering is used to detect outliers in the initial environmental data to obtain first environmental data that excludes data that deviates from normal distribution;

[0071] Performing filtering on the first environmental data using a Kalman filter algorithm to obtain smoothed second environmental data;

[0072] Feature extraction is performed on the second environmental data to construct a standardized data sample including a timestamp, spatial location and characteristic parameters, wherein the characteristic parameters include temperature change rate, salinity gradient, dissolved oxygen content fluctuation amplitude and pH value anomaly coefficient.

[0073] Density peak clustering is an unsupervised outlier detection algorithm based on local density and distance. It calculates the local density of each data point and its minimum distance to high-density points to identify outliers with significantly lower density than their neighbors. The Kalman filter is a recursive filtering method for time series data. By establishing a state-space model and using the state estimate at the previous moment and the current observation, it recursively calculates the optimal state estimate at the current moment, effectively suppressing Gaussian noise.

[0074] The temperature change rate is the change in water temperature per unit time, reflecting the thermal dynamic stability of the marine environment; the salinity gradient is the difference in salinity within a unit spatial distance, characterizing the degree of water mixing and the impact of runoff; the dissolved oxygen content fluctuation amplitude is the difference between the maximum and minimum dissolved oxygen concentrations within a certain time window, reflecting the stability of the water body's redox environment; the pH value anomaly coefficient is the degree of deviation of the current pH value from the historical average value for the same period, and is used to identify acid-base balance anomalies.

[0075] It should be noted that the initial environmental data collected by sensors are often affected by equipment failure, biological attachment or environmental disturbances, and contain outliers and high-frequency noise. Direct use will lead to model misjudgment.

[0076] Specifically, in this embodiment, parameters such as water temperature, salinity, dissolved oxygen, and pH value of the target sea area are collected through sensors such as buoys, submersibles, and AUVs to form an initial data set containing timestamps and spatial locations.

[0077] The density peak clustering algorithm is used to set the distance threshold, where the distance threshold can be dynamically determined according to the standard deviation of historical data. The local density and distance parameters of each data point are calculated, and outliers are identified and eliminated to obtain the first environmental data.

[0078] The Kalman filter algorithm is applied to the first environmental data to construct a state space model. Through the prediction-update recursive process, high-frequency noise is filtered out, long-term trends are retained, and smoothed second environmental data are generated.

[0079] The temperature change rate and dissolved oxygen fluctuation amplitude are calculated based on the time window. The salinity gradient is calculated using the inverse distance weighted method based on the spatial location of the sensor. The pH anomaly coefficient is calculated using the Z-score. The timestamp, spatial location, and the above characteristic parameters are integrated and normalized to form a feature vector with a unified dimension.

[0080] Through three-level processing, the measured data are transformed from original discrete points into high-quality samples containing spatiotemporal dynamic characteristics, providing a reliable data basis for multi-source data fusion and disaster prediction model construction, and significantly improving the early warning method's ability to capture complex changes in the marine environment.

[0081] S2: Construct an image segmentation network that includes an attention mechanism and a multi-scale feature fusion module. Based on the image segmentation network, perform boundary segmentation and feature fusion on the remote sensing image data in sequence to obtain optimized multispectral feature data.

[0082] Preferably, in one embodiment of the present invention, the remote sensing image data is sequentially subjected to boundary segmentation and feature fusion based on an image segmentation network to obtain optimized multispectral feature data, including:

[0083] Based on the image segmentation network, the multi-spectral band data in the remote sensing image data is extracted to obtain multi-scale spectral features, which include shallow spectral features and deep spectral features.

[0084] The shallow spectral features and deep spectral features are spliced using the channel splicing method to obtain a red tide probability map with the same resolution as the remote sensing image data;

[0085] Double-threshold segmentation is performed on the red tide probability map to obtain the red tide boundary coordinates, where the double-threshold segmentation is designed to extract the core red tide area based on the first segmentation threshold and the fuzzy transition boundary between the red tide and normal seawater based on the second segmentation threshold;

[0086] The red tide boundary coordinates are spatially associated with remote sensing image data to obtain multispectral feature data including multispectral data and red tide boundary coordinates.

[0087] The attention mechanism is a neural network module that simulates human visual attention. It uses weight distribution to focus the network on key areas, suppress irrelevant information, and improve the targeted nature of feature extraction. Multi-scale feature fusion utilizes features from different layers of a convolutional neural network, integrating multi-scale information through splicing or weighting, addressing the inaccurate segmentation of complex boundaries caused by single-scale features.

[0088] It should be noted that while multispectral data from remote sensing images can reveal spectral anomalies on the sea surface, they present two major challenges. The first is blurred boundaries: the edge of a red tide is affected by water mixing and illumination angle, resulting in a gradual transition in spectral characteristics. Traditional single-threshold segmentation methods are prone to missing edges or misidentifying noise. The second is an imbalance in feature scales: shallow features can pinpoint red tide details, while deep features can identify the entire red tide. However, when used independently, these two methods cannot achieve both accuracy and integrity.

[0089] By focusing on the spectral features related to red tides through the attention mechanism, combining multi-scale feature fusion to retain details and semantic information, and then distinguishing the core area from the transition boundary through double threshold segmentation, the coordinates of the red tide boundary can be accurately extracted and associated with the original spectral data, providing high-precision spatial features for subsequent spatiotemporal alignment (matching with measured environmental data) and disaster range deduction, solving the problems of rough boundary segmentation and insufficient feature utilization in traditional methods.

[0090] Specifically, in this embodiment, an image segmentation network based on U-Net or DeepLab is constructed, embedding a spatial attention module and multi-scale convolution, wherein,

[0091] Shallow feature extraction: The shallow features of the multispectral image are extracted through a 3×3 convolutional layer, with the same resolution as the original image;

[0092] Deep feature extraction: Semantic features are extracted through downsampling and deep convolution, with reduced resolution but enhanced semantic information.

[0093] Preferably, in one embodiment of the present invention, a channel splicing method is used to splice shallow spectral features and deep spectral features to obtain a red tide probability map with the same resolution as the remote sensing image data, including:

[0094] Adjusting the spatial dimensions of the shallow spectral features and the deep spectral features to obtain the first shallow spectral features and the first deep spectral features with the same resolution as the remote sensing image data;

[0095] Corresponding adaptive weights are assigned to the first shallow spectral feature and the first deep spectral feature respectively, to obtain the second shallow spectral feature and the second deep spectral feature, wherein the adaptive weights reflect the importance of the shallow spectral feature and the deep spectral feature in the prediction of red tide probability;

[0096] The second shallow spectral feature and the second deep spectral feature are spliced in the channel dimension to obtain the red tide probability feature map;

[0097] The probability of each pixel in the red tide probability characteristic map being a red tide area is calculated to obtain a red tide probability map.

[0098] It should be noted that the shallow and deep spectral features of remote sensing images are complementary. Directly concatenating unweighted features can lead to an imbalance between detail and semantic information. Spatial dimension adjustment ensures feature spatial alignment, and adaptive weight assignment highlights the key role of different features in red tide prediction. Ultimately, channel concatenation fuses multi-scale information to generate high-precision red tide probability maps, addressing the lack of adaptability of traditional fixed-weight fusion to complex spectral features.

[0099] Specifically, in this embodiment, after shallow and deep feature extraction, a channel attention module is introduced to dynamically adjust the weight of each band feature according to the red tide spectral characteristics to suppress the interference of irrelevant targets such as clouds and ships.

[0100] Through upsampling, the deep features are restored to the original image resolution, consistent with the dimension of the shallow features; the shallow and deep features are spliced along the channel dimension to generate a red tide probability feature map containing details and semantic information; through the convolution layer and Sigmoid activation function, the probability of each pixel in the probability feature map belonging to the red tide area is calculated to obtain a red tide probability map of the same resolution. For example, a pixel value of 0.7 indicates a 70% probability of being a red tide pixel.

[0101] Next, the thresholds are set, wherein the first threshold, i.e., the high threshold, is used to extract the core red tide area; the second threshold, i.e., the low threshold, is used to extract the transition boundary outside the core area.

[0102] For the binary image after double threshold segmentation, the OpenCV contour detection algorithm is used to extract the polygon vector coordinates, and the minimum circumscribed rectangle of the core area and the boundary vertex coordinates of the transition area are recorded.

[0103] The red tide boundary coordinates are aligned with the geocoding of the original remote sensing image to ensure that each boundary point corresponds to a multispectral pixel in the image; and boundary labels are added to the original multispectral data to form multispectral feature data containing spectral features and boundary attributes.

[0104] S3: The multispectral feature data and the environmental data are spatiotemporally aligned according to the dynamic spatiotemporal calibration algorithm to obtain a spatiotemporal aligned dataset;

[0105] Preferably, in one embodiment of the present invention, the multispectral feature data and the environmental data are spatiotemporally aligned according to a dynamic spatiotemporal alignment algorithm to obtain a spatiotemporally aligned dataset, including:

[0106] Based on the passing time of the remote sensing satellite, the collection time of the environmental data is dynamically resampled through a linear interpolation algorithm to generate the first environmental data that is synchronized with the multispectral feature data.

[0107] Based on the geocoding parameters of the acquired multispectral feature data, spatially register the collection location of the environmental data in the first environmental data, and map the environmental data to a spatial grid with the same resolution as the multispectral feature data to obtain the second environmental data;

[0108] The second environment data was rationally verified to obtain a spatiotemporally aligned dataset.

[0109] The dynamic spatiotemporal calibration algorithm aims to adjust data from different sources and with different spatiotemporal resolutions to achieve consistency in both temporal and spatial dimensions, thereby achieving effective data fusion. The linear interpolation algorithm is a data processing method that estimates data values at unknown time points based on the linear relationship between known data points, thereby achieving data synchronization in the temporal dimension. Geocoding parameters, such as latitude and longitude, projected coordinates, are used to identify geographic spatial locations and can determine the specific location of multispectral feature data in geographic space.

[0110] It's important to note that multispectral signature data and environmental data are derived from remote sensing imagery and sensor measurements, respectively, and therefore differ in time and space. Multispectral signature data are typically acquired during satellite overflights and have specific temporal and spatial resolutions; environmental data, on the other hand, are continuously collected by sensors and have a more discrete temporal and spatial distribution. Without spatiotemporal alignment, the two data types cannot be effectively combined, hindering subsequent analysis and prediction of marine ecological disasters. Spatiotemporal alignment can integrate data from different sources into a common spatiotemporal framework, providing an accurate and consistent data foundation for subsequent model development and disaster warning.

[0111] The spatiotemporal calibration method and accuracy indicators of the present invention are shown in Table 1.

[0112] Table 1

[0113]

[0114] Specifically, in this embodiment, the target time point for time synchronization is determined based on the remote sensing satellite's transit time. For each parameter in the environmental data, a linear interpolation algorithm is used to estimate the data value at the target time point based on its acquisition time and corresponding value. After linear interpolation, first environmental data is generated that is time-synchronized with the multispectral feature data, ensuring temporal consistency between the two data types.

[0115] Obtain the geocoding parameters for the multispectral signature data and determine its spatial reference and resolution. For each acquisition location in the first environmental data, map it to a spatial grid with the same resolution as the multispectral signature data based on the geocoding parameters. Spatial interpolation methods such as nearest neighbor interpolation and bilinear interpolation can be used to convert discrete environmental data points into continuous spatially distributed data. Spatial registration is then used to obtain the second environmental data, spatially aligning the environmental data with the multispectral signature data for subsequent data fusion and analysis.

[0116] The secondary environment data is checked for plausibility, such as whether the data's value range is within a reasonable range and whether the data's changing trends conform to physical laws. Thresholds and rules can be set to determine data plausibility, and any non-compliant data can be corrected or eliminated. For example, if temperature data shows abnormally high or low temperatures, this could be due to a sensor failure or data transmission error, requiring further processing. After this plausibility verification, the final spatiotemporally aligned dataset is obtained, which is consistent with the multispectral feature data in both time and space, ensuring data quality.

[0117] S4: Map each sample pair in the spatiotemporal alignment dataset into graph node data, analyze the connection relationship between any two adjacent graph node data, and construct an adjacency relationship matrix based on the analysis results, where the connection relationship reflects the feature correlation and spatiotemporal correlation between the graph node data;

[0118] Preferably, in one embodiment of the present invention, each sample pair data in the spatiotemporal alignment dataset is mapped to graph node data, and the connection relationship between any two adjacent graph node data is analyzed, and an adjacency relationship matrix is constructed based on the analysis results, including:

[0119] Map each sample pair in the spatiotemporal alignment dataset into graph node data in the graph structure. Each graph node data includes spatial attributes, temporal attributes, and feature attributes.

[0120] Calculate the spatial distance of the spatial attributes of any two adjacent graph node data based on the Euclidean distance formula;

[0121] Calculate the feature similarity of the feature attributes of any two adjacent graph node data based on the Pearson correlation coefficient;

[0122] The connection relationship between any two adjacent graph node data is determined based on spatial distance and feature similarity, and the adjacency relationship matrix is constructed according to the connection relationship.

[0123] In the graph structure, each graph node represents a sample pair in the spatiotemporal alignment dataset, including spatial, temporal, and feature attributes. The adjacency matrix describes the connections between nodes in the graph. The elements in the matrix indicate whether there is a connection between two nodes and the strength of the connection.

[0124] The Pearson correlation coefficient is a statistic that measures the linear correlation between two variables. In this step, it is used to calculate the feature similarity between the feature attributes of adjacent graph node data. The Euclidean distance formula is used to calculate the distance between two points in space. In this step, it is used to calculate the spatial distance between the spatial attributes of adjacent graph node data.

[0125] It's important to note that while spatiotemporally aligned datasets achieve temporal and spatial uniformity, the relationships between the data are unclear. Mapping the data into graph nodes and constructing an adjacency matrix allows for the spatial and temporal correlations and feature correlations between the data to be represented in a graph structure, facilitating the subsequent use of graph analysis methods such as graph convolutional networks to uncover underlying patterns and regularities within the data. This graph structure can better capture both local and global information between the data, helping to improve the accuracy and reliability of marine ecological disaster predictions.

[0126] Traverse each sample pair in the spatiotemporal alignment dataset and map it to graph node data in the graph structure. Assign spatial attributes, temporal attributes, and feature attributes to each graph node data.

[0127] The spatial distance between the spatial attributes of any two adjacent graph node data is calculated using the Euclidean distance formula, and the feature similarity between the feature attributes of any two adjacent graph node data is calculated based on the Pearson correlation coefficient.

[0128] Preferably, in one embodiment of the present invention, the connection relationship between any two adjacent graph node data is determined based on spatial distance and feature similarity, and an adjacency relationship matrix is constructed according to the connection relationship, including:

[0129] Based on the geographical environment characteristics of the target sea area, the seasonal changes in the marine ecosystem and historical marine ecological disaster data, the spatial distance threshold and feature similarity threshold of the target sea area are determined;

[0130] Make a preliminary judgment on the connection relationship between any two adjacent graph node data based on the spatial distance threshold and feature similarity threshold;

[0131] If the preliminary judgment result is that there is a connection relationship between the two graph node data, the connection strength between the two graph node data is calculated based on the weighted summation method to generate an adjacency relationship matrix that reflects the spatiotemporal correlation and feature correlation between the graph node data.

[0132] Specifically, in this embodiment, the spatial distance threshold and feature similarity threshold of the target sea area are determined based on the geographical environment characteristics of the target sea area, the seasonal variation pattern of the marine ecosystem and the historical marine ecological disaster data.

[0133] For any two adjacent graph nodes, compare their spatial distance with the spatial distance threshold, as well as their feature similarity with the feature similarity threshold. If the spatial distance is less than or equal to the spatial distance threshold and the feature similarity is greater than or equal to the feature similarity threshold, it is preliminarily determined that the two graph nodes are connected.

[0134] If the preliminary judgment result shows that there is a connection relationship between the two graph node data, the connection strength between the two graph node data is calculated based on the weighted summation method. For example, the weights of spatial distance and feature similarity can be set as and ,in . Then the connection strength is expressed as:

[0135]

[0136] in, is the connection strength, is the spatial distance, is the spatial distance threshold, is the feature similarity.

[0137] Constructing the adjacency matrix , the elements of the matrix Representing graph nodes and If there is no connection between two graph nodes, then ;otherwise, .

[0138] S5: Based on the pre-built graph convolutional network, the adjacency matrix is subjected to neighborhood aggregation processing to generate a fused feature vector. The fused feature vector includes the spatiotemporal dependency relationship between the spectral feature data and the environmental data.

[0139] Graph convolutional networks (GCNs) are neural networks specifically designed for processing graph-structured data. They extract feature information from graph data by aggregating and transforming the features of nodes and their neighbors. Neighborhood aggregation is a core operation of GCNs, specifically weighted summarization of the features of a node and its neighbors to capture the relationships and feature information between nodes.

[0140] It's important to note that after the multi-source spectral and environmental data are spatiotemporally aligned and constructed into a graph structure, the spatiotemporal and feature correlations between the data are represented through an adjacency matrix. However, the features of these data at different nodes are scattered and their mutual influences are not fully reflected. By leveraging the neighborhood aggregation processing of a graph convolutional network, the features of a node and its neighbors can be fused, revealing the potential spatiotemporal dependencies between the spectral and environmental data. The resulting fused feature vector can more comprehensively and accurately reflect the state of the marine ecosystem, providing more valuable information for the subsequent prediction of marine ecological disasters.

[0141] Specifically, in this embodiment, the pre-built graph convolutional network typically includes multiple graph convolutional layers, each with a specific number of neurons. Common graph convolutional layers use specific convolution operations, such as Chebyshev convolution or first-order approximate convolution, to randomly initialize the weight parameters in the graph convolutional network. These parameters are continuously adjusted and optimized during subsequent training.

[0142] The previously constructed adjacency relationship matrix is input into the graph convolutional network as the topological structure information of the graph. The data of each graph node constitutes the node feature matrix, which also serves as the input of the graph convolutional network.

[0143] In each layer of the graph convolutional network, for each node, its neighbor nodes are determined according to the adjacency matrix. Then the features of the neighbor nodes and the features of the node itself are weighted and summed according to certain rules. Taking the first-order approximate convolution as an example, Layer Node Features The update formula is:

[0144]

[0145] in, For nodes The set of neighbor nodes of and Node and The degree, It is The learnable weight matrix of the layer, is the activation function.

[0146] By stacking multiple graph convolutional layers and continuously performing neighborhood aggregation operations, information can be propagated more widely within the graph. The output of each layer serves as the input to the next layer. After multiple layers of processing, node features can be integrated with information from more distant nodes, thereby capturing more complex spatiotemporal dependencies.

[0147] After multi-layer neighborhood aggregation processing in the graph convolutional network, a feature representation for each node is finally obtained. By combining the feature representations of all nodes in a specific manner, a fused feature vector is generated. This fused feature vector contains the spatiotemporal dependencies between the spectral feature data and the environmental data, and serves as input for subsequent marine disaster prediction models.

[0148] S6: Construct a marine disaster prediction model based on the cross-domain transfer learning method, input the fused feature vector into the marine disaster prediction model, obtain a multi-dimensional prediction result that reflects the marine ecological disaster warning, and conduct real-time warning of ecological disasters in the target sea area based on the multi-dimensional prediction result.

[0149] Preferably, in one embodiment of the present invention, a marine disaster prediction model is constructed according to a cross-domain transfer learning method, including:

[0150] The network structure and training parameters of the land disaster prediction model are extracted as the basic marine disaster prediction model. The fully connected layer and domain adversarial network module for marine ecological disaster prediction are embedded in the basic marine disaster prediction model to generate the first marine disaster prediction model.

[0151] The historical data of marine ecological disasters are obtained as target domain data, and the historical data of land disasters are obtained as source domain data. By minimizing the difference in feature distribution between the source domain data and the target domain data, the parameters of the first marine disaster prediction model are corrected to obtain the second marine disaster prediction model.

[0152] Perform data enhancement processing on the historical data of marine ecological disasters to obtain the second historical data of marine ecological disasters;

[0153] The second marine disaster prediction model is trained based on the historical data of the second marine ecological disaster, and the weight parameters of the second marine disaster prediction model are updated by optimizing the loss function to obtain a trained marine disaster prediction model.

[0154] Cross-domain transfer learning involves transferring knowledge and models learned in one domain (the source domain) to a different but related domain (the target domain) to improve the performance of the target domain model. This is particularly useful when data in the target domain is scarce. A land disaster prediction model is a machine learning or deep learning model designed to predict land disasters. It has a specific network structure and trained parameters.

[0155] A fully connected layer is a layer in a neural network. Each neuron in a fully connected layer is connected to all neurons in the previous layer. It is responsible for nonlinearly transforming and combining features and is commonly used in classification or regression tasks. The domain adversarial network module uses adversarial training to enable the model to learn common features between source and target domain data while minimizing the difference in feature distribution between them, thereby improving the model's generalization ability in the target domain. Data augmentation processing involves various transformations and expansions of the original data to increase the diversity and quantity of the data, thereby improving the model's generalization ability and robustness.

[0156] Marine ecological disaster data is generally scarce. Directly building and training marine disaster prediction models may face overfitting due to insufficient data, resulting in poor generalization. However, terrestrial disaster data is relatively abundant, and terrestrial and marine ecosystems share certain similarities, such as the fact that disasters are both influenced by environmental factors. Therefore, cross-domain transfer learning methods, leveraging the knowledge and experience of terrestrial disaster prediction models, can provide a good initial state for marine disaster prediction models. Furthermore, data augmentation can further expand historical marine ecological disaster data, improving the model's generalization ability and ultimately building a more accurate and reliable marine disaster prediction model.

[0157] Specifically, in this embodiment, the network structure and training parameters of a pre-trained land disaster prediction model are extracted to serve as a basic marine disaster prediction model. A fully connected layer and a domain adversarial network module for marine ecological disaster prediction are embedded within this basic marine disaster prediction model. The fully connected layer performs further nonlinear transformations and combinations on the input features to adapt them to the task of marine ecological disaster prediction. The domain adversarial network module is used to address differences in feature distribution between the source and target domains.

[0158] Historical data on marine ecological disasters is collected as target domain data, while historical data on land disasters is obtained as source domain data. A domain adversarial network module is used to minimize the differences in feature distributions between the source and target domain data through adversarial training. Specifically, the domain adversarial network module consists of a feature extractor and a discriminator. The feature extractor attempts to extract common features between the source and target domain data, while the discriminator attempts to distinguish whether the data comes from the source or target domain. Through continuous adversarial training, the feature extractor gradually learns feature representations that can eliminate domain differences, thereby correcting the parameters of the first marine disaster prediction model and obtaining a second marine disaster prediction model.

[0159] Preferably, in one embodiment of the present invention, data enhancement processing is performed on the marine ecological disaster historical data to obtain the second marine ecological disaster historical data, including:

[0160] Adding Gaussian noise with a mean of zero and an adaptive standard deviation to the time series feature data in the historical data of marine ecological disasters, and randomly shifting the spatial position coordinate data of the historical data of marine ecological disasters to generate first enhanced data containing spatiotemporal disturbances;

[0161] A marine ecological data generation model based on a generative adversarial network is constructed to generate second enhanced data that conforms to the characteristics of the target sea area. The marine ecological data generation model is designed to learn the distribution pattern of marine ecological data by inputting feature vectors of real disaster samples.

[0162] The first enhanced data and the second enhanced data are validated, and the first enhanced data and the second enhanced data that have passed the validation are merged to obtain the second marine ecological disaster historical data.

[0163] Gaussian noise is random noise with a Gaussian distribution, whose mean is zero and whose standard deviation determines the strength of the noise. Adding Gaussian noise can simulate measurement errors and uncertainties in real data.

[0164] It's important to note that historical data on marine ecological disasters is typically limited, which can lead to overfitting and poor generalization in models trained on this data. Data augmentation can increase the diversity and quantity of data, enabling the model to learn more diverse feature patterns, thereby improving its generalization and robustness. Furthermore, augmented data generated through various methods can simulate the actual conditions of marine ecological disasters from multiple perspectives, further enhancing the model's adaptability to complex situations.

[0165] Specifically, in this embodiment, Gaussian noise is added to each data point in the time series feature data of historical marine ecological disaster data, such as the water temperature series at different times in a particular sea area. The mean of the Gaussian noise is zero, and the standard deviation is adaptively adjusted based on the characteristics of the data. The spatial coordinate data of the historical marine ecological disaster data, such as longitude and latitude, is randomly offset. Through these operations, first enhanced data containing spatiotemporal disturbances is generated.

[0166] A marine ecological data generation model is constructed based on a generative adversarial network. The generator receives a random noise vector as input and attempts to generate marine ecological disaster data that matches the characteristics of the target sea area. The discriminator receives the feature vectors of real disaster samples and the generated marine ecological disaster data as input and attempts to distinguish between real and generated data.

[0167] The model uses feature vectors from real-world disaster samples as input, and through adversarial training, the generator learns the distribution patterns of marine ecological data. Specifically, during training, the parameters of the generator and discriminator are alternately updated, continuously improving the discriminator's discriminative power while simultaneously generating data that increasingly resembles the distribution of real data. After the generative model is trained, a random noise vector is fed into the generator to generate secondary augmented data that matches the characteristics of the target ocean area.

[0168] The validity of the first and second augmented data is verified. For example, the validity of the data value range is checked, such as whether the water temperature data is within the reasonable ocean temperature range; and the physical relationship of the data is checked to see if it is consistent with reality, such as whether the relationship between salinity and density is reasonable.

[0169] Some thresholds and rules can be set to judge the validity of the data, and data that does not meet the requirements can be eliminated or corrected. The first enhanced data and the second enhanced data that have passed the validity verification are merged with the original marine ecological disaster historical data to obtain the second marine ecological disaster historical data.

[0170] Input historical data on the second marine ecological disaster into the second marine disaster prediction model for training. Define an appropriate loss function, such as the cross-entropy loss function or the mean squared error loss function, to measure the difference between the model's predictions and the true labels. Use an optimization algorithm to optimize the loss function and continuously update the weight parameters of the second marine disaster prediction model until the model performance reaches a satisfactory level, resulting in a trained marine disaster prediction model.

[0171] The fused feature vector is input into the marine disaster prediction model to obtain multi-dimensional prediction results reflecting marine ecological disaster warnings, where the multi-dimensional prediction results include:

[0172] Disaster probability: predict the possibility of marine ecological disasters occurring in specific sea areas and time periods;

[0173] Disaster type: output the probability distribution of disaster type;

[0174] Impact scope and intensity: Output the spatial scope and severity of the disaster impact.

[0175] Preferably, in one embodiment of the present invention, a real-time early warning of ecological disasters in a target sea area is performed based on multi-dimensional prediction results, including:

[0176] Obtain real-time hydrological and meteorological data of the target sea area;

[0177] Input multi-dimensional prediction results, real-time hydrological data, and meteorological data into a pre-built disaster assessment model, and determine the risk level of ecological disasters based on the model output;

[0178] Obtain ecologically sensitive areas in the target sea area and determine the degree of threat to the ecologically sensitive areas based on the disaster impact range in the multi-dimensional prediction results;

[0179] Execute early warning control instructions generated by risk level and threat degree.

[0180] Real-time hydrological data includes dynamic water-related data such as current flow velocity, water level, temperature, and salinity in the target sea area, reflecting the current water conditions. Real-time meteorological data includes current weather conditions in the target sea area, such as wind speed, direction, temperature, and precipitation. Meteorological factors can have a significant impact on marine ecology. The disaster assessment model is a pre-built model that comprehensively analyzes multi-dimensional forecast results, real-time hydrological data, and meteorological data to assess the level of ecological disaster risk.

[0181] Risk levels categorize the potential harm caused by ecological disasters, such as low, medium, and high risk, to facilitate the implementation of different response measures. Ecologically sensitive areas are areas within the target sea area where ecosystems are more fragile and susceptible to disasters, such as marine protected areas and fishery aquaculture areas. The threat level assesses the potential impact of a disaster on an ecologically sensitive area based on the disaster's impact area as determined by multi-dimensional predictions. Early warning and control directives are a series of response instructions tailored to the risk level and threat level, such as issuing warnings of varying levels and initiating emergency responses.

[0182] It's important to note that while multi-dimensional forecasts provide preliminary information on marine ecological disasters, the marine environment is dynamic, and real-time hydrological and meteorological data can reflect the actual conditions in the current sea area. By incorporating this data into disaster assessment models, the risk level of ecological disasters can be more accurately determined. Furthermore, considering the unique characteristics of ecologically sensitive areas, assessing their threat levels can help inform targeted protective measures. Finally, executing early warning and control commands can ensure timely response to disasters, mitigating their impact on marine ecosystems and human activities.

[0183] Specifically, in this embodiment, hydrological monitoring equipment, such as buoys and underwater sensors, deployed in the target sea area collects data such as water velocity, water level, water temperature, and salinity in real time. These devices can regularly transmit data to a data center to ensure the timeliness and accuracy of the data. Real-time meteorological data for the target sea area, including wind speed, wind direction, temperature, precipitation, and other information, is obtained from the meteorological department. Alternatively, meteorological monitoring stations can be set up around the sea area to conduct independent data collection.

[0184] Multi-dimensional forecast results, real-time hydrological data, and meteorological data are fed into a pre-built disaster assessment model. This model can be based on a machine learning algorithm. The model conducts a comprehensive analysis of the input data, taking into account the interactions between various factors, and outputs an ecological disaster risk level. For example, high water temperature, low wind speed, and high nutrient content may increase the risk of red tide.

[0185] Using technologies such as Geographic Information Systems (GIS), we can obtain information on the location, extent, and ecological characteristics of ecologically sensitive areas in the target sea area. This information can include the boundaries of marine protected areas and the distribution of fishery aquaculture zones. Based on the multi-dimensional prediction results, the disaster impact area is overlaid with the location of the ecologically sensitive areas for analysis. This determines whether the ecologically sensitive areas are within the disaster impact area and the extent of the impact. The threat level of the ecologically sensitive areas is assessed based on factors such as the degree of overlap between the disaster impact area and the ecologically sensitive areas and the vulnerability of the ecologically sensitive areas. The threat level can be categorized as mild, moderate, or severe.

[0186] By comprehensively considering multi-dimensional forecast results, real-time hydrological data, and meteorological data, we can more accurately assess the risk level of ecological disasters and reduce misjudgments and omissions. By assessing the threat level of ecologically sensitive areas, we can implement targeted protective measures to reduce the impact of disasters on these areas and protect the marine ecosystem. The implementation of early warning and control instructions can promptly activate emergency response mechanisms, implement effective countermeasures, and minimize the damage to the marine ecosystem and human activities caused by disasters.

[0187] Compared with the prior art, the embodiments of the present invention have the following advantages:

[0188] 1) The present invention obtains remote sensing images and measured environmental data separately, processes the remote sensing images using an image segmentation network with an attention mechanism and a multi-scale feature fusion module, and then achieves spatiotemporal alignment of the data through a dynamic spatiotemporal calibration algorithm. It constructs a graph node and adjacency relationship matrix, and generates a fusion feature vector through a graph convolutional network. This method integrates the advantages of multi-source data, takes into account the spatiotemporal dependency of spectral features and environmental data, breaks through the limitations of single data, and can more comprehensively and deeply capture the complex mechanisms of marine ecological disasters, thereby improving the accuracy and reliability of disaster prediction models.

[0189] 2) The present invention uses the cross-domain transfer learning method to construct a marine disaster prediction model, and inputs the fused feature vector into the model to obtain multi-dimensional prediction results, covering key information such as the probability of disaster occurrence and the scope of impact. It can provide real-time early warning of ecological disasters in the target sea area, effectively meeting the real-time prevention and control needs of sudden marine ecological disasters.

[0190] The above embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous modifications and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

Claims

1. A marine ecological disaster early warning method based on multi-source heterogeneous big data fusion, characterized by: include: Obtain remote sensing image data and measured environmental data of the target sea area respectively; Constructing an image segmentation network including an attention mechanism and a multi-scale feature fusion module, and sequentially performing boundary segmentation and feature fusion on the remote sensing image data based on the image segmentation network to obtain optimized multispectral feature data; Performing spatiotemporal alignment on the multispectral feature data and the environmental data according to a dynamic spatiotemporal alignment algorithm to obtain a spatiotemporal aligned dataset; wherein performing spatiotemporal alignment on the multispectral feature data and the environmental data according to the dynamic spatiotemporal alignment algorithm to obtain a spatiotemporal aligned dataset comprises: Based on the passing time of the remote sensing satellite, the collection time of the environmental data is dynamically resampled by a linear interpolation algorithm to generate first environmental data that is synchronized with the multispectral feature data; Based on the acquired geocoding parameters of the multispectral feature data, spatially registering the collection location of the environmental data in the first environmental data, and mapping the first environmental data to a spatial grid with the same resolution as the multispectral feature data to obtain second environmental data; Performing rationality verification on the second environmental data to obtain the spatiotemporal alignment dataset; Mapping each sample pair data in the spatiotemporal alignment dataset into graph node data, analyzing the connection relationship between any two adjacent graph node data, and constructing an adjacency relationship matrix based on the analysis results, wherein the connection relationship reflects the feature correlation and spatiotemporal correlation between the graph node data; wherein mapping each sample pair data in the spatiotemporal alignment dataset into graph node data, analyzing the connection relationship between any two adjacent graph node data, and constructing the adjacency relationship matrix based on the analysis results, includes: Mapping each sample pair in the spatiotemporal alignment dataset into graph node data in a graph structure, wherein each graph node data includes spatial attributes, temporal attributes, and feature attributes; Calculate the spatial distance of the spatial attributes of any two adjacent graph node data based on the Euclidean distance formula; Calculate the feature similarity of the feature attributes of any two adjacent graph node data according to the Pearson correlation coefficient; Determine the connection relationship between any two adjacent graph node data based on the spatial distance and the feature similarity, and construct an adjacency relationship matrix according to the connection relationship; Performing neighborhood aggregation processing on the adjacency relationship matrix based on a pre-built graph convolutional network to generate a fused feature vector, wherein the fused feature vector includes a spatiotemporal dependency relationship between the spectral feature data and the environmental data; A marine disaster prediction model is constructed according to the cross-domain transfer learning method, and the fused feature vector is input into the marine disaster prediction model to obtain a multi-dimensional prediction result reflecting the marine ecological disaster warning. Based on the multi-dimensional prediction result, a real-time warning of ecological disasters in the target sea area is performed.

2. The marine ecological disaster early warning method based on multi-source heterogeneous big data fusion according to claim 1 is characterized in that: The step of sequentially performing boundary segmentation and feature fusion on the remote sensing image data based on the image segmentation network to obtain optimized multispectral feature data includes: Extracting features from the multispectral band data in the remote sensing image data based on the image segmentation network to obtain multi-scale spectral features, wherein the multi-scale spectral features include shallow spectral features and deep spectral features; The shallow spectral features and the deep spectral features are spliced using a channel splicing method to obtain a red tide probability map with the same resolution as the remote sensing image data; Performing double-threshold segmentation on the red tide probability map to obtain red tide boundary coordinates, wherein the double-threshold segmentation is designed to extract a core red tide area based on a first segmentation threshold and to extract a fuzzy transition boundary between the red tide and normal seawater based on a second segmentation threshold; The red tide boundary coordinates are spatially associated with the remote sensing image data to obtain multispectral feature data including multispectral data and the red tide boundary coordinates.

3. The marine ecological disaster early warning method based on multi-source heterogeneous big data fusion according to claim 2 is characterized in that: The channel splicing method is used to splice the shallow spectral features and the deep spectral features to obtain a red tide probability map with the same resolution as the remote sensing image data, including: Adjusting the spatial dimensions of the shallow spectral feature and the deep spectral feature to obtain a first shallow spectral feature and a first deep spectral feature having the same resolution as the remote sensing image data; Assigning corresponding adaptive weights to the first shallow spectral feature and the first deep spectral feature respectively to obtain a second shallow spectral feature and a second deep spectral feature, wherein the adaptive weights reflect the importance of the shallow spectral feature and the deep spectral feature in the red tide probability prediction; splicing the second shallow spectral feature and the second deep spectral feature in the channel dimension to obtain a red tide probability characteristic map; The probability that each pixel point in the red tide probability characteristic map is a red tide area is calculated to obtain the red tide probability map.

4. The marine ecological disaster early warning method based on multi-source heterogeneous big data fusion according to claim 1 is characterized in that: Obtain measured environmental data of the target sea area, including: Acquiring initial environmental data of the target sea area based on a sensor, performing outlier detection on the initial environmental data using an outlier recognition algorithm based on density peak clustering, and obtaining first environmental data from which data deviating from a normal distribution is eliminated; Performing filtering on the first environmental data using a Kalman filter algorithm to obtain smoothed second environmental data; Feature extraction is performed on the second environmental data to construct a standardized data sample including a timestamp, a spatial position and characteristic parameters, wherein the characteristic parameters include a temperature change rate, a salinity gradient, a dissolved oxygen content fluctuation amplitude and a pH value anomaly coefficient.

5. The marine ecological disaster early warning method based on multi-source heterogeneous big data fusion according to claim 1 is characterized in that: The determining of the connection relationship between any two adjacent graph node data based on the spatial distance and the feature similarity, and constructing an adjacency relationship matrix according to the connection relationship, includes: Determine the spatial distance threshold and feature similarity threshold of the target sea area based on the geographical environment characteristics of the target sea area, the seasonal variation pattern of the marine ecosystem, and historical marine ecological disaster data; Performing a preliminary judgment on the connection relationship between any two adjacent graph node data according to the spatial distance threshold and the feature similarity threshold; If the preliminary judgment result is that there is a connection relationship between the two graph node data, the connection strength between the two graph node data is calculated based on the weighted summation method to generate an adjacency relationship matrix reflecting the spatiotemporal correlation and feature correlation between the graph node data.

6. The marine ecological disaster early warning method based on multi-source heterogeneous big data fusion according to claim 1 is characterized in that: The marine disaster prediction model is constructed according to the cross-domain transfer learning method, including: Extracting the network structure and training parameters of the land disaster prediction model as a basic marine disaster prediction model, and embedding a fully connected layer and a domain adversarial network module for marine ecological disaster prediction into the basic marine disaster prediction model to generate a first marine disaster prediction model; Acquire historical data of marine ecological disasters as target domain data, and acquire historical data of land disasters as source domain data, and correct the parameters of the first marine disaster prediction model by minimizing the difference in feature distribution between the source domain data and the target domain data to obtain a second marine disaster prediction model; Performing data enhancement processing on the marine ecological disaster historical data to obtain second marine ecological disaster historical data; The second marine disaster prediction model is trained based on the second marine ecological disaster historical data, and the weight parameters of the second marine disaster prediction model are updated by optimizing the loss function to obtain a trained marine disaster prediction model.

7. The marine ecological disaster early warning method based on multi-source heterogeneous big data fusion according to claim 6 is characterized in that: The performing of data enhancement processing on the marine ecological disaster historical data to obtain second marine ecological disaster historical data includes: Adding Gaussian noise with a zero mean and adaptive standard deviation to the time series feature data in the marine ecological disaster historical data, and randomly shifting the spatial position coordinate data of the marine ecological disaster historical data to generate first enhanced data containing spatiotemporal disturbance; Constructing a marine ecological data generation model based on a generative adversarial network, and generating second enhanced data that conforms to the characteristics of the target sea area based on the marine ecological data generation model, wherein the marine ecological data generation model is designed to learn the distribution pattern of marine ecological data by inputting feature vectors of real disaster samples; The first enhanced data and the second enhanced data are validated, and the first enhanced data and the second enhanced data that have passed the validation are merged to obtain the second marine ecological disaster history data.

8. The marine ecological disaster early warning method based on multi-source heterogeneous big data fusion according to claim 1 is characterized in that: The real-time early warning of ecological disasters in the target sea area based on the multi-dimensional prediction results includes: Obtain real-time hydrological and meteorological data of the target sea area; Inputting the multi-dimensional prediction results, the real-time hydrological data, and the meteorological data into a pre-built disaster assessment model, and determining the risk level of ecological disasters based on the output results of the model; Obtaining the ecologically sensitive areas of the target sea area, and determining the degree of threat to the ecologically sensitive areas based on the disaster impact range in the multi-dimensional prediction results; Execute the early warning control instruction generated by the risk level and the threat degree.

Images