A deep learning-based marine environment monitoring and early warning method and system

Abstract

The application discloses a kind of based on deep learning's marine environment monitoring early warning method and system, it is related to marine environment monitoring early warning technical field;Including the following steps: comb multiple source historical monitoring data, execute prediction analysis, obtain the core feature standard set in preset period;Its technical key points are: the application realizes whole-process standardization, seamless link of each link, dynamic response, closed-loop management and control, and the effect that practicality greatly improves;Achieved to improve whole-process monitoring early warning efficiency, ensure each link accurate link, dynamically adapt pollution change, reduce pollution disposal cost, improve the overall ability of marine pollution prevention and control, while various technologies cooperate, significantly improve the efficiency of whole-process monitoring early warning, reduce each link link error, can dynamically respond to pollution change, greatly improve practicality and operability, comprehensively enhance the overall ability of marine pollution prevention and control, use effect is good, with good use prospect.

Description

Technical Field

[0001] This invention relates to the field of marine environmental monitoring and early warning technology, specifically to a marine environmental monitoring and early warning method and system based on deep learning. Background Technology

[0002] As a key component of the Earth's climate system, the evolution of the ocean's environmental elements directly impacts the development of multiple fields, including shipping safety, fisheries production, marine engineering, disaster early warning, and climate assessment. With the intensification of global climate change, extreme marine meteorological events such as typhoons, strong cold waves, and abnormal intrusions of the Kuroshio Current occur frequently, placing more stringent demands on the accuracy, timeliness, coverage, regional adaptability, and physical interpretability of marine environmental forecasts.

[0003] Marine environmental monitoring and early warning refers to the systematic and continuous monitoring and analysis of the marine environment, as well as the prediction and analysis of monitoring data, to promptly detect abnormal changes and potential risks in the marine environment, and to issue early warnings and alerts so that corresponding measures can be taken to protect the marine ecosystem and maintain marine security. The purpose of marine environmental monitoring and early warning is to protect the marine ecological environment, safeguard human interests, and promote sustainable marine development. This is achieved through monitoring and analyzing factors such as ocean water temperature, salinity, water quality, oxygen content, plankton, benthic organisms, marine pollutants, and climate change, enabling timely... To understand changes in the marine environment and assess whether there are potential environmental risks.

[0004] The invention patent with patent authorization announcement number CN116625327B describes a big data-driven marine environment monitoring and early warning method. This method includes the following steps: introducing Internet of Things (IoT) communication technology, combining remote sensing technology, UAV swarm cooperation, and biomarker technology to acquire various types of marine intelligence in real time. In this invention, by collecting and analyzing ocean current data, marine areas are divided to achieve targeted resource deployment for corresponding data collection equipment, providing a comprehensive and timely marine intelligence foundation. Distributed database technology is used to store marine environmental information flow, improving data reliability, availability, and scalability. Edge computing technology is used to perform preliminary analysis of large amounts of data at data collection points, reducing data transmission and processing latency. While improving the referenceability and storage completeness of big data, the method performs distributed analysis and modeling analysis in a two-stage processing of the data, reducing computational load while improving computational accuracy, achieving a more comprehensive early warning effect.

[0005] Existing methods for marine environmental monitoring and early warning have fundamental flaws, specifically: The anomaly detection system described in the aforementioned patent has issues with standardization, lacks a unified water quality assessment benchmark, and has not reached a unified standard for comparing monitoring parameters. At the same time, it does not have an effective interference removal mechanism and is easily affected by factors such as the natural environment and equipment errors, which may lead to misjudgment and missed detection of anomalies.

[0006] In summary, existing marine environmental monitoring and early warning methods do not meet market demands. Therefore, we propose a deep learning-based marine environmental monitoring and early warning method and system. Summary of the Invention

[0007] To achieve the above objectives, the present invention provides the following technical solution: A deep learning-based method for marine environmental monitoring and early warning includes the following steps: By analyzing historical monitoring data from multiple sources and performing predictive analysis, a set of core feature standards for a preset period is obtained. The region is divided, and satellite remote sensing monitoring data and buoy matrix monitoring data in each region are collected and analyzed in real time to obtain satellite scores and matrix scores. The edge depth model constructed by combining the input of the core feature standard set is then processed to obtain preliminary judgment results. Analyze the preliminary judgment results. When the preliminary judgment results are abnormal, start the hybrid depth model to analyze the core feature standard set, satellite remote sensing monitoring data and buoy matrix monitoring data to obtain the pollution type and confidence level. The concentration data was summarized using the buoy-based contour line generation method to generate pollution concentration contour lines and upload the coordinates. A preliminary pollution range map was drawn using the Kriging interpolation algorithm combined with satellite imagery. Based on the preliminary pollution range map, the preliminary coordinates of the pollution source were calculated using a reverse extrapolation algorithm to obtain the pollution range and the preliminary location coordinates of the pollution source. Send the pollution range and preliminary location coordinates of the pollution source to the verification platform, execute the on-site verification strategy, and receive on-site verification data; By analyzing on-site verification data, pollution range and preliminary location coordinates of pollution sources, real-time satellite remote sensing monitoring data, real-time buoy matrix monitoring data and real-time meteorological data, the pollution diffusion trend in the future period can be obtained. Based on a pre-set comprehensive assessment system, the scores of each indicator are evaluated according to four categories: abnormal pollution concentration, diffusion rate, ecological harm, and social impact. The pollution score is obtained by weighted summation, and corresponding early warning strategies and dynamic adjustment strategies are implemented until the pollution is eliminated.

[0008] Preferably, the multi-source historical monitoring data covers historical satellite remote sensing monitoring data, historical meteorological data, and historical buoy matrix monitoring data. Among them, the satellite remote sensing monitoring data includes hyperspectral satellite imagery, thermal infrared satellite imagery, synthetic aperture radar satellite scattering coefficient data, and spatial distribution data of water quality parameters, as well as chlorophyll fluorescence intensity, colored soluble organic matter absorption coefficient, oil film reflectance, and sea surface temperature obtained through further analysis; the meteorological data includes wind speed, precipitation, and sea surface temperature; and the buoy matrix monitoring data includes long-term monitoring sequence data of phycobiliprotein concentration, microbial metabolic activity, suspended particulate matter particle size distribution, and volatile organic pollutant concentration.

[0009] Preferably, the predictive analysis includes: Data standardization and structured organization techniques are used to process multi-source historical monitoring data to obtain a standardized data sample set; Spatiotemporal convolutional neural network technology is used to process spatially correlated data in the data sample set to obtain spatial dimension features; We use bidirectional long short-term memory neural network technology to process time-series data in the data sample set to obtain time-dimensional features; Dynamic weight allocation technology is used to process spatial and temporal features to obtain a weighted feature set adapted to different periods; The weighted feature set is processed using feature mapping and range calibration techniques to obtain the dynamic reference range of each type for each period and each spatial grid point, forming the core feature standard set within the preset period, and divided into satellite reference range set and buoy matrix reference range set.

[0010] Preferably, real-time acquisition and analysis of satellite remote sensing monitoring data includes first performing data cleaning to remove noise and eliminate cloud cover interference, then comparing it with the satellite reference range set to calculate the deviation ratio of each satellite monitoring element, then scoring it according to the single-element scoring standard, and finally weighting and summing according to the importance of the elements to obtain the satellite anomaly score for each region. The real-time acquisition and analysis of buoy matrix monitoring data includes first performing data cleaning to remove noise, then comparing it with the buoy matrix baseline range set to calculate the deviation ratio of each monitoring element in each buoy, then scoring according to the single-element scoring standard, and finally weighting and summing according to the importance of the elements to obtain the buoy anomaly score of each buoy, and calculating the average of the buoy anomaly scores of all buoys in the region to obtain the matrix score.

[0011] Preferably, the edge depth model processing, constructed by combining the input of the core feature standard set, includes: Data normalization and integration techniques are used to process satellite remote sensing monitoring data, buoy matrix monitoring data, satellite scores, and matrix scores to obtain a standardized dataset; By using the cross-attention mechanism to allocate weights, mining and integrating cross-correlation features, and then sequentially carrying out anomaly feature extraction, inference operations, and balance deviation verification, the anomaly confidence and balance coefficient are obtained. Anomaly threshold determination is performed based on anomaly confidence level and balance coefficient. If the anomaly confidence level is lower than the preset confidence threshold and the balance coefficient is lower than the set standard threshold, it is judged as normal; otherwise, it is judged as abnormal.

[0012] The preferred method for analyzing data using a hybrid depth model is as follows: By employing data pruning and time-series processing techniques, we analyze the core feature standard set, satellite remote sensing monitoring data, and buoy matrix monitoring data to obtain a standard feature dataset. By using convolutional neural network technology to analyze satellite remote sensing image data and satellite baseline range set in standard feature dataset, a satellite pollution spatial feature set is obtained; The converter neural network technology was used to analyze the buoy matrix monitoring data and the buoy matrix baseline range set in the standard feature dataset to obtain the buoy pollution time series feature set; Cross-attention mechanism technology was used to analyze the spatial feature set of satellite pollution and the temporal feature set of buoy pollution, and the core identification feature set was obtained after feature fusion. Classify and infer the pollution type and pollution confidence level by classifying the core identification feature set.

[0013] Preferably, the steps for generating the pollution range and preliminary location coordinates of the pollution source are as follows: The real-time collected buoy matrix monitoring data is verified, and invalid concentration data caused by wind, waves, and equipment errors are removed. Then, the data collected by all buoys that are initially judged to be abnormal are summarized to construct an abnormal dataset. The contamination gradient is obtained by calculating the extent by which all data in the anomalous dataset exceeds the baseline range of the buoy matrix. The pollution gradient contour lines were generated using linear interpolation, key coordinate points were selected, and the data was transmitted to the satellite data processing center. The key coordinate points were calibrated, and the Kriging interpolation algorithm was used to fill the monitoring blind spots by combining satellite image features to draw a preliminary pollution range map. Based on the interpolation results, the pollution boundary is delineated using the set pollution gradient as the critical value, and the coordinates of the core pollution area are determined. By summarizing satellite pollution plume trajectories, anomalous datasets, real-time meteorological data, and prior baseline data, spatiotemporal alignment and preprocessing are completed, and data affected by natural factors are removed to obtain the analysis dataset. The dataset was processed and analyzed using a time-space converter neural network, and the pollution diffusion starting point was calculated using a reverse inference algorithm to obtain the preliminary location coordinates of the pollution source. By combining real-time satellite remote sensing monitoring data with anomaly datasets, the specific type of pollution source can be determined.

[0014] Preferably, when analyzing on-site verification data, pollution range and preliminary location coordinates of pollution sources, real-time satellite remote sensing monitoring data, real-time buoy matrix monitoring data and real-time meteorological data, fluid dynamics simulation technology is used to process the data to obtain preliminary prediction results of pollution diffusion trends; space-time converter neural network technology is used to process satellite remote sensing monitoring data and buoy matrix monitoring data to obtain deviation correction parameters; and result fusion and error calibration technology is used to process the preliminary prediction results and deviation correction parameters to obtain the pollution diffusion trend in the future period.

[0015] Preferably, the corresponding early warning strategies are as follows: When the pollution score is higher than 80 points, it is judged as Level 1 pollution, an emergency response is initiated, operations in the polluted area are stopped, cleaning equipment is deployed, high-frequency monitoring is carried out and early warning information is issued, and a support request signal is sent; when the pollution score is between 60 and 80 points, it is judged as Level 2 pollution, a special response is initiated, activities in the polluted area are restricted, the responsible party is notified to rectify the situation, medium- and high-frequency monitoring is implemented and early warning information is issued; when the pollution score is lower than 60 points, it is judged as Level 3 pollution, a routine response is initiated, monitoring work is strengthened, the pollution spread trend is tracked, the responsible party is notified to conduct inspections and a warning message is issued.

[0016] A deep learning-based marine environmental monitoring and early warning system includes a historical analysis module, an anomaly detection module, a pollution detection module, an analysis and location module, an on-site verification module, a pollution prediction module, and a detection and early warning module. Historical analysis module: Organizes multi-source historical monitoring data, performs predictive analysis, and obtains a set of core feature standards within a preset period; Anomaly detection module: Divides regions, collects and analyzes satellite remote sensing monitoring data and buoy matrix monitoring data in each region in real time, obtains satellite scores and matrix scores, and processes the edge depth model constructed by combining the input of the core feature standard set to obtain preliminary detection results; Pollution determination module: Analyzes preliminary determination results. When the preliminary determination result is abnormal, the hybrid depth model is activated to analyze the core feature standard set, satellite remote sensing monitoring data and buoy matrix monitoring data to obtain the pollution type and confidence level. Analysis and positioning module: The buoy-based contour line generation method is used to summarize concentration data, generate pollution concentration contour lines and upload coordinates. The Kriging interpolation algorithm is used in combination with satellite imagery to draw a preliminary pollution range map. Based on the preliminary pollution range map, the reverse deduction algorithm is used to calculate the preliminary coordinates of the pollution source, and the pollution range and the preliminary positioning coordinates of the pollution source are obtained. Field verification module: Sends the pollution range and preliminary location coordinates of the pollution source to the verification platform, executes the field verification strategy, and receives the field verification data; Pollution prediction module: Analyzes on-site verification data, pollution range and preliminary location coordinates of pollution sources, real-time satellite remote sensing monitoring data, real-time buoy matrix monitoring data and real-time meteorological data to obtain the pollution diffusion trend in the future period; Detection and early warning module: Based on a preset comprehensive evaluation system, it evaluates the scores of four categories of indicators: abnormal pollution concentration, diffusion rate, ecological hazard, and social impact. The scores are then weighted and summed to obtain the pollution score, and corresponding early warning and dynamic adjustment strategies are implemented until the pollution is eliminated.

[0017] This invention provides a deep learning-based method and system for marine environmental monitoring and early warning, which has the following beneficial effects: This invention processes multi-source historical monitoring data and conducts predictive analysis to obtain a set of core feature standards within a preset period. It overcomes the shortcomings of existing solutions in anomaly detection, such as the lack of unified standards and high rates of false positives and false negatives. This invention achieves standardization and precision in anomaly detection, significantly improves the efficiency of anomaly detection, reduces the risk of false alarms and false negatives, enhances the accuracy of detection, lays a solid foundation for subsequent steps, and reduces the ineffective consumption of human and equipment resources.

[0018] This invention combines convolutional neural networks with converter neural networks and employs cross-attention fusion technology to process satellite remote sensing monitoring data, buoy matrix monitoring data, and a core feature standard set within a preset period. This overcomes the shortcomings of existing solutions, such as insufficient feature extraction for pollution identification, inadequate data fusion, inaccurate differentiation of pollution types, and inability to confirm the authenticity of pollution. It achieves accurate determination of the presence or absence of pollution, precise identification of pollution types, and deep feature fusion. This not only improves the accuracy of pollution identification and avoids misjudgments, but also clarifies the direction of pollution control, providing precise support for subsequent pollution control efforts.

[0019] This invention utilizes a combination of buoy contour generation technology and satellite kriging interpolation algorithm. It leverages a spacetime converter neural network to process abnormal buoy-collected data, satellite pollution plume trajectory data, real-time meteorological data, and a core feature standard set within a preset period. This overcomes the shortcomings of existing methods, such as significant deviations in pollution range delineation, inaccurate pollution source location, insufficient consideration of meteorological factors, and inadequate data coordination. It achieves accurate delineation of pollution range and preliminary accurate location of pollution sources, with deviations within a controllable range and adaptability to meteorological conditions. It achieves the goal of clearly defining pollution boundaries and accurately locating pollution sources, providing clear targets for subsequent UAV verification, reducing the difficulty of subsequent handling, and improving handling efficiency.

[0020] This invention constructs a complete process encompassing baseline prediction, dual monitoring, anomaly detection, pollution identification, range delineation and location, UAV calibration, diffusion analysis, and comprehensive early warning. This process effectively overcomes the shortcomings of existing solutions, such as the lack of a unified baseline, disconnect between stages, incomplete closed loop, fixed monitoring frequency, and poor practicality. It achieves full-process standardization, seamless connection between stages, dynamic response, closed-loop management, and significantly improved practicality. It achieves the goals of improving the efficiency of the entire monitoring and early warning process, ensuring precise connection between stages, dynamically adapting to pollution changes, reducing pollution treatment costs, and enhancing the overall capability of marine pollution prevention and control. Simultaneously, the synergistic cooperation of various technologies significantly improves the efficiency of the entire monitoring and early warning process, reduces connection errors between stages, and enables dynamic response to pollution changes, greatly enhancing practicality and operability, comprehensively strengthening the overall capability of marine pollution prevention and control. It has shown good results and broad application prospects. Attached Figure Description

[0021] Figure 1 This is a flowchart of a deep learning-based marine environmental monitoring and early warning method according to the present invention; Figure 2 This is a structural block diagram of a deep learning-based marine environmental monitoring and early warning system according to the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 are within the scope of protection of the present invention.

[0023] Example 1: Please see Figure 1 This embodiment provides a deep learning-based method for marine environmental monitoring and early warning, comprising the following steps: S1. Organize multi-source historical monitoring data, perform predictive analysis, and obtain the core feature standard set within the preset period; The multi-source historical monitoring data originates from a cloud database and encompasses historical satellite remote sensing data, historical meteorological data, and historical buoy matrix monitoring data. The satellite remote sensing data includes hyperspectral satellite imagery, thermal infrared satellite imagery, synthetic aperture radar satellite scattering coefficient data, and spatial distribution data of water quality parameters. This data primarily compensates for the spatial coverage limitations of the buoy matrix monitoring data. Further analysis is also conducted on chlorophyll fluorescence intensity, colored soluble organic matter absorption coefficient, oil film reflectance, and sea surface temperature. Meteorological data includes wind speed, precipitation, and sea surface temperature. Analysis of this meteorological data aims to capture the long-term impact of climate factors on water quality parameters, ensuring that subsequent analyses align with water quality variation patterns under different meteorological conditions. The buoy matrix monitoring data includes long-term monitoring sequences of phycobiliprotein concentration, microbial metabolic activity, suspended particulate matter particle size distribution, and volatile organic pollutant concentrations.

[0024] Performing predictive analytics includes: Data standardization and structured organization techniques are used to process multi-source historical monitoring data to obtain a standardized data sample set; First, the data is categorized and filtered to remove invalid data caused by equipment failure or extreme interference. Linear interpolation is used to supplement missing data to ensure data integrity. Second, the data is organized according to a two-dimensional structure combining time steps and spatial grids. The time steps are strictly matched with the set water quality benchmark period, and the spatial grids are divided according to the actual monitoring needs of the sea area. Each grid corresponds to a fixed range of latitude and longitude to ensure the consistency of the spatiotemporal dimensions of the data. Finally, the data of each spatial grid are sampled and integrated. Satellite remote sensing data and meteorological data of the same grid and the same time step are bound into a complete sample to obtain a standardized comprehensive data sample. The standardized comprehensive data sample is then summarized after removing outliers using the 3σ principle and normalizing the data to obtain a standardized data sample set.

[0025] Spatiotemporal convolutional neural network technology is used to process spatially correlated data in the data sample set to obtain spatial dimension features; A three-layer 3×3 convolution kernel with a stride of 1 and "same" padding is used to ensure that the feature map size after convolution is consistent with the input. Convolution operations are performed to extract features from the spatial grid neighborhood data, capturing the correlation of water quality parameters in adjacent sea areas, the contour boundary features of water quality anomaly areas in satellite imagery, and the spatial gradient changes in meteorological data, thus initially extracting local spatial features. Subsequently, batch normalization is performed to normalize the local features extracted by convolution, adjusting the feature data to the same distribution range to avoid model training bias. Then, the ReLU activation function is used to nonlinearly transform the normalized features, retaining useful spatial features. Next, random deactivation regularization is performed, randomly discarding some feature nodes to prevent model overfitting. Finally, feature filtering is performed using a feature importance ranking algorithm to calculate the contribution of local features, eliminating redundant features with a contribution below a threshold (usually 0.6), retaining core spatial features, reducing the computational burden on subsequent layers, and obtaining spatial dimension features such as the transmission relationship of water quality parameters in adjacent sea areas, the spatial distribution pattern of water quality under seasonal coupling, and the spatial gradient features of meteorological parameters.

[0026] We use bidirectional long short-term memory neural network technology to process time-series data in the data sample set to obtain time-dimensional features; First, a bidirectional temporal traversal is performed. The bidirectional Long Short-Term Memory (LSTM) network is divided into two branches: forward and backward. The forward LSTM traverses from the beginning to the end of the time series, capturing long-term trends in water quality and meteorological parameters; the backward LSTM traverses from the end to the beginning, capturing short-term sudden changes and inverse correlation features. Second, through the input gate, forget gate, and output gate of the LSTM unit, the long-term and short-term dependencies of the time series data are accurately captured. The forget gate discards invalid historical features, the input gate receives current and useful historical features, and the output gate outputs the integrated features, solving the problem of long-term dependency loss in traditional recurrent neural networks and capturing both long-term and short-term changes in water quality. Next, the temporal features are integrated by concatenating and fusing the temporal features output from the two branches of the bidirectional LSTM. A weighted average algorithm is used to obtain preliminary temporal features containing long-term and short-term dependencies. Finally, L2 regularization is used to suppress overfitting, and the integrated temporal features are normalized to make them consistent with the magnitude of spatial features, thus obtaining temporal dimension features.

[0027] Dynamic weight allocation technology is used to process spatial and temporal features to obtain a weighted feature set adapted to different periods; An attention score calculation algorithm is used to calculate the importance scores of spatial and temporal features in different cycles. Since the importance varies in different seasons, the calculation is combined with the seasonal feature analysis results to consider the correlation between the features and common pollution types and core influencing factors in that cycle. Combined with the feature variance contribution, redundant features with small variance and no obvious fluctuations are eliminated. The remaining features are then sorted by importance score. Based on the sorting results and the cycle association mapping table, a weight multiplication strategy is implemented for key features to make the weight adjustment adapt to the set cycle. A weighted summation algorithm is used to fuse spatial and temporal features according to the adjusted weights. A multilayer perceptron is used to perform nonlinear transformation on the fused features to strengthen the expression of key features and suppress the interference of invalid features. Finally, a weighted feature set adapted to different set cycles is obtained.

[0028] The weighted feature set is processed using feature mapping and range calibration techniques to obtain the dynamic reference range of each type for each period and each spatial grid point, forming the core feature standard set within the preset period, and statistically divided into the satellite reference range set and the buoy matrix reference range set.

[0029] By employing a feature mapping algorithm, a one-to-one correspondence is established between the weighted feature set and the benchmark range of the element type. The key features in the weighted feature set directly affect the setting of the upper and lower limits of the benchmark range. During the mapping process, the statistical regularity of water quality parameters during the pollution-free period within the cycle is referenced simultaneously to obtain the benchmark range. The format is unified as cycle, spatial grid coordinates, element category, upper benchmark value, lower benchmark value, benchmark mean, and fluctuation range. At the same time, the cycle characteristics, common pollution types, and applicable sea areas corresponding to the benchmark are labeled, thereby forming the core feature standard set within the preset cycle, and statistically divided into satellite benchmark range set and buoy matrix benchmark range set.

[0030] To adapt to the dynamic changes in marine climate and ecological environment and avoid anomaly judgment bias caused by fixed benchmarks, a synchronous automatic update mechanism is set up according to a preset cycle. The specific operation is as follows: input the newly added satellite remote sensing monitoring data and buoy matrix monitoring data in the previous cycle to reflect the recent water quality change trend; adopt the incremental training method to fine-tune the parameters of the original scheme with the help of new samples, thereby saving training time and computing resources and ensuring the timeliness and accuracy of the benchmark range.

[0031] S2. Divide the region, collect and analyze satellite remote sensing monitoring data and buoy matrix monitoring data in each region in real time, obtain satellite score and matrix score, and process the edge depth model constructed by combining the core feature standard set input to obtain preliminary judgment results; Real-time acquisition and analysis of satellite remote sensing monitoring data includes first performing data cleaning to remove noise and eliminate cloud cover interference, then comparing it with the satellite baseline range set to calculate the deviation ratio of each satellite monitoring element, then scoring it according to the single-element scoring standard, and finally weighting and summing according to the importance of the elements to obtain the satellite anomaly score for each region. Satellite monitoring elements include chlorophyll fluorescence intensity, absorption coefficient of colored soluble organic matter, oil film reflectivity, sea surface temperature, and synthetic aperture radar scattering coefficient. The formula for calculating the deviation ratio of each satellite monitoring element is: Deviation ratio = |Measured value - Reference average| / Reference average × 100%. Given that sea surface temperature fluctuates relatively little, its deviation ratio needs to be determined separately. When the sea surface temperature exceeds the reference average by 0.5-2℃, it is calculated as 50%; when it exceeds the reference average by more than 2℃, it is calculated as 100%. When the synthetic aperture radar scattering coefficient exceeds the reference upper limit by less than 30%, it is calculated as 30%; when it exceeds the reference upper limit by more than 30%, it is calculated as 100%.

[0032] The satellite imagery covers a wide area, and the measured values ​​here are the average values ​​of individual monitored elements in the region.

[0033] When scoring according to the single-element scoring criteria, if the deviation ratio is 0%, 2 points are given; if the deviation ratio is between 1% and 30%, 4 points are given; if the deviation ratio is between 31% and 60%, 7 points are given; if the deviation ratio is between 61% and 100%, 10 points are given. Alternatively, the deviation ratio can be directly divided by 10 as the score for that element.

[0034] When weighting the summation according to the importance of the elements, the weight of chlorophyll fluorescence intensity is set at 20%, the weight of colored soluble organic matter absorption coefficient is 15%, the weight of oil film reflectivity is 25%, the weight of sea surface temperature is 15%, and the weight of synthetic aperture radar scattering coefficient is 25%. The higher the score, the more severe the anomaly.

[0035] The real-time acquisition and analysis of buoy matrix monitoring data includes first performing data cleaning to remove noise, then comparing it with the buoy matrix baseline range set to calculate the deviation ratio of each monitoring element in each buoy, then scoring according to the single-element scoring standard, and finally weighting and summing according to the importance of the elements to obtain the buoy anomaly score of each buoy, and calculating the average of the buoy anomaly scores of all buoys in the region to obtain the matrix score.

[0036] The buoy matrix consists of several buoys, with a certain number of buoys in each area for monitoring. Each buoy is equipped with several sensors, including a bio-optical sensor for collecting phycobiliprotein concentrations, primarily used for red tide identification; a microbial fuel cell sensor for collecting microbial metabolic activity, primarily used to detect the degree of organic pollution; a laser particle size analyzer for collecting the particle size distribution of suspended particulate matter, primarily used to distinguish between natural sediment and industrial pollutant particles; and a miniature volatile organic pollutant sensor for collecting the concentration of volatile organic pollutants.

[0037] To facilitate buoy matrix calculations, edge computing can be used. The edge device performs synchronous calculations on the data collected by all buoys in the entire buoy matrix, and then aggregates the data to the cloud. The edge device cleans and removes noise from the raw data, calculates the abnormal deviation degree of a single buoy, and calculates the abnormal score of a single buoy. Then, the cloud calculates the average of the abnormal scores of all buoys in the region to obtain the matrix score.

[0038] The calculation principle for the buoy anomaly score is the same as that for the satellite anomaly score, only the elements and their corresponding weights are different, so it will not be described in detail.

[0039] The edge depth model, constructed by combining the input of the core feature standard set, includes the following processing: Data normalization and integration techniques are used to process satellite remote sensing monitoring data, buoy matrix monitoring data, satellite scores, and matrix scores to obtain a standardized dataset; The processing here includes cleaning four types of data to remove interference, eliminating invalid data such as noise and equipment errors, standardizing outlier scores, and unifying and standardizing satellite and buoy outlier scores; aligning and integrating multi-source data to achieve matching of dual-source data in the spatiotemporal dimension; calculating balance indexes, verifying the balance coefficient of satellite data, and analyzing and transforming the dispersion of buoy point data.

[0040] Weights were allocated using a cross-attention mechanism, with satellite remote sensing data and buoy matrix monitoring data each receiving 40% weight and buoy matrix monitoring data 60%. Cross-correlation features were mined and integrated, and anomaly feature extraction, inference calculation, and balance deviation verification were carried out sequentially to obtain anomaly confidence and balance coefficient. Specifically, the satellite remote sensing monitoring data is processed using regional balance coefficient calculation technology. For the satellite remote sensing monitoring data, monitoring sub-regions are divided, the deviation ratio of each sub-region is calculated, and the regional balance coefficient is calculated. The balance coefficient = standard deviation of deviation ratio of each sub-region / mean of overall deviation ratio.

[0041] The system employs point dispersion analysis technology to process buoy matrix monitoring data. It calculates the average score of all buoys in the area, verifies the deviation rate between each buoy's score and the average score, counts the number of points with deviation rates exceeding a set value, and compares the results against corresponding conditions to obtain the buoy data balance coefficient. These conditions can be set according to requirements. For example, if the deviation rate of a single buoy is less than 50% and the number of buoys with deviation rates exceeding 30% accounts for less than 20% of the total number of buoys in the area, then the balance coefficient is 20%; otherwise, the balance coefficient is 60%.

[0042] The cross-attention mechanism is used to process satellite remote sensing monitoring data and buoy matrix monitoring data; weight allocation, cross-correlation feature mining, and fusion feature integration analysis and transformation are carried out; finally, a collaborative feature set that takes into account both macroscopic and microscopic anomalies is obtained and output; inference technology composed of a 3-layer perceptron is used to process the collaborative feature set, and feature mapping, nonlinear inference operations, and confidence calibration analysis and transformation are carried out in combination with the core feature standard set; the initial anomaly confidence is obtained and output to achieve a quantitative assessment of the degree of anomaly.

[0043] Anomaly threshold determination is performed based on anomaly confidence level and balance coefficient. If the anomaly confidence level is lower than the preset confidence threshold and the balance coefficient is lower than the set standard threshold, it is judged as normal; otherwise, it is judged as abnormal.

[0044] The lower the anomaly confidence level and the balance coefficient, the more normal the data is, and the lower the possibility of contamination.

[0045] S3. Analyze the preliminary judgment results. When the preliminary judgment results are abnormal, start the hybrid depth model to analyze the core feature standard set, satellite remote sensing monitoring data and buoy matrix monitoring data to obtain the pollution type and confidence level. The process of data analysis using a hybrid deep model is as follows: By employing data pruning and time-series processing techniques, we analyze the core feature standard set, satellite remote sensing monitoring data, and buoy matrix monitoring data to obtain a standard feature dataset. Based on the coordinates of the suspected contaminated area, targeted cropping operations were performed on the hyperspectral satellite full-area imagery and the synthetic aperture radar satellite scattering coefficient matrix. Image data of the suspected contaminated area and its surrounding 1-kilometer buffer zone were retained, while data from irrelevant areas were removed to reduce the computational load of the model. This ensured that the cropped image coordinates were consistent with the satellite monitoring coordinate system, resulting in satellite remote sensing data. The time step data of individual buoys were stitched together in chronological order to construct a single buoy time series. The time series of all buoys in the area were summarized, and the buoy coordinates and timestamps corresponding to each time series data were labeled to ensure accurate correlation between the time series data and spatial coordinates. Missing values ​​were filled in the stitched time series, and continuous invalid data caused by equipment failure were removed to obtain matrix time series data.

[0046] The min-max normalization algorithm is used to standardize the data so that it falls within the range of [0, 1], eliminating model bias caused by differences in data volume. Based on timestamps and spatial coordinates, satellite remote sensing data, matrix time series data, and core feature standard sets in the suspected contaminated area are fully aligned in the spatiotemporal dimension. After alignment, each set of data includes satellite spatial features, buoy time series features, and corresponding baseline parameters, forming a complete data sample and obtaining a standard feature dataset.

[0047] By using convolutional neural network technology to analyze satellite remote sensing image data and satellite baseline range set in standard feature dataset, a satellite pollution spatial feature set is obtained; Main steps: Local features of the images are extracted using four-layer convolution operations. For hyperspectral images, core spatial features such as oil film reflectance peaks are extracted. Features such as the scattering intensity of suspended particulate matter pollution plumes are extracted from the synthetic aperture radar scattering coefficient matrix, initially obtaining spatial feature maps for two types of satellite images. Max pooling is performed on the spatial feature maps to reduce their dimensionality and computational burden. After pooling, redundant spatial features are removed, retaining only key pollution-related features. A feature importance ranking algorithm is used to calculate the contribution of each spatial feature after pooling, removing redundant features with a contribution below 0.05. Combined with the spatial features selected by comparison with pre-defined benchmark parameters, abnormal spatial features exceeding the benchmark range are retained to enhance the expression of pollution features, ultimately obtaining the satellite pollution spatial feature set.

[0048] The converter neural network technology was used to analyze the buoy matrix monitoring data and the buoy matrix baseline range set in the standard feature dataset to obtain the buoy pollution time series feature set; The buoy pollution time series feature set includes the pollution spatial distribution profile, feature intensity, and anomaly gradient, with each feature labeled with its corresponding confidence level.

[0049] Main steps: The buoy time-series features are input into the converter encoder. A multi-head attention mechanism is used to capture time-series dependencies, focusing on abnormal fluctuations in parameter concentrations and time-series correlations between parameters, and initially extracting time-series feature vectors. The extracted time-series feature vectors are then subjected to nonlinear transformation to enhance the expression of abnormal time-series features. A time-series anomaly detection algorithm is used to mine abnormal trends and mark the corresponding pollution correlation. Combining the spatial features of satellite remote sensing data, the correlation between buoy time-series features and satellite spatial features is mined. If a correlation exists, the weight of the time-series feature is strengthened to improve feature correlation and reliability. Simultaneously, the persistence of abnormal trends is judged by combining the time-series changes in buoy micro-anomaly scores, resulting in a buoy pollution time-series feature set.

[0050] The buoy pollution time series feature set includes the time series fluctuation pattern of parameter concentration, abnormal trends, dual-source correlation characteristics, and pollution correlation degree. Each time series feature is labeled with its corresponding confidence level.

[0051] Cross-attention mechanism technology was used to analyze the spatial feature set of satellite pollution and the temporal feature set of buoy pollution, and the core identification feature set was obtained after feature fusion. Main steps: First, an attention score calculation algorithm is used to calculate the importance score and synergy of each feature, and the features are sorted from highest to lowest importance score. Second, based on the importance ranking, a weight multiplication strategy is implemented for synergistic features. Third, for features unique to a single data source, the base weight is maintained or the weight is reduced to avoid feature bias caused by interference from a single data source. Fourth, a weighted summation algorithm is used to fuse the ranked dual-source features according to the adjusted weights to obtain a preliminary fused feature set. This preliminary fused feature set is then input into a multilayer perceptron for nonlinear transformation to enhance the expression of key contamination features and suppress invalid interference features, ultimately yielding the core recognition feature set.

[0052] The core identification feature set includes pollution features, feature importance weights, collaborative association information, and preliminary association labels for pollution types. Each feature is labeled with a comprehensive confidence level.

[0053] Classify and infer the pollution type and pollution confidence level by classifying the core identification feature set.

[0054] Main steps: Input the core identification feature set into the multi-classifier and calculate the probability of each of the four common marine pollution types: oil pollution, red tide pollution, organic pollution, and suspended particulate matter pollution. Set the confidence threshold for pollution type to 85 points. If the probability of a certain type of pollution is greater than or equal to 0.85, it is clearly identified as that type of pollution, and the corresponding confidence score is output. If the probability of all pollution types is less than 0.85, the pollution type is uncertain and further confirmation is required based on UAV verification data. If the probability of two or more pollution types is greater than or equal to 0.85, it is identified as a compound pollution, and each type of pollution and its corresponding confidence score are labeled according to the probability from high to low.

[0055] A temperature scaling algorithm is used to calibrate the output probabilities of multi-class classification, correcting the confidence bias caused by model overfitting. The confidence of pollution type can effectively reflect the accuracy of type identification. The multi-class judgment results, corresponding confidence, feature support basis, spatial coordinates and timestamps are integrated to construct standardized pollution identification results.

[0056] S4. The concentration data is summarized using the buoy-based contour line generation method to generate pollution concentration contour lines and upload the coordinates. A preliminary pollution range map is drawn using the Kriging interpolation algorithm combined with satellite imagery. Based on the preliminary pollution range map, the preliminary coordinates of the pollution source are calculated using the reverse deduction algorithm to obtain the pollution range and the preliminary location coordinates of the pollution source. The steps for generating the pollution range and preliminary location coordinates of the pollution source are as follows: The real-time collected buoy matrix monitoring data is verified, and invalid concentration data caused by wind, waves, and equipment errors are removed. Then, the data collected by all buoys that are initially judged to be abnormal are summarized to construct an abnormal dataset. The contamination gradient is obtained by calculating the extent by which all data in the anomalous dataset exceeds the baseline range of the buoy matrix. Pollution gradient = (measured abnormal concentration - baseline upper limit) / baseline upper limit, with intervals of 0.1 units of pollution gradient.

[0057] The pollution gradient contour lines were generated using linear interpolation, key coordinate points were selected, and the data was transmitted to the satellite data processing center. Linear interpolation is used to generate pollution concentration contour lines. Each contour line corresponds to a unique concentration gradient value, and the pollution gradient interval is clearly marked. At the same time, the key coordinate points of each contour line are selected.

[0058] The key coordinate points were calibrated, and the Kriging interpolation algorithm was used to fill the monitoring blind spots by combining satellite image features to draw a preliminary pollution range map. After receiving the contour line key coordinates and concentration gradient data uploaded by the buoys, the satellite data processing center first calibrates the coordinates. Then, it uses the Kriging interpolation algorithm to combine the key coordinates of the buoys with the features of the satellite macroscopic image and performs concentration gradient interpolation work in the blank areas between adjacent buoys to fill the monitoring blind spots.

[0059] Based on the interpolation results, the pollution boundary is delineated using the set pollution gradient as the critical value, and the coordinates of the core pollution area are determined. Based on the interpolation results, a preliminary pollution range map is drawn, dividing the pollution boundary and core pollution area, and marking the location of buoy monitoring points and the corresponding pollution gradient. After the drawing is completed, it is compared with the range of macroscopic anomalies in the satellite and the interpolation deviation is corrected to ensure that the preliminary pollution range can cover the macroscopic anomaly area monitored by the satellite. Finally, the preliminary delineation error is controlled within ≤1 kilometer to meet the basic accuracy requirements for subsequent UAV calibration.

[0060] By summarizing satellite pollution plume trajectories, anomalous datasets, real-time meteorological data, and prior baseline data, spatiotemporal alignment and preprocessing are completed, and data affected by natural factors are removed to obtain the analysis dataset. The dataset was processed and analyzed using a time-space converter neural network, and the pollution diffusion starting point was calculated using a reverse inference algorithm to obtain the preliminary location coordinates of the pollution source. First, the temporal feature extraction method is used to capture the temporal variation trend of the pollution plume trajectory and the dynamic variation law of the concentration gradient. Then, the spatial feature fusion method is used to fuse spatial information such as the distribution of high concentration areas of buoys and ocean current meteorological conditions. Subsequently, the reverse inference algorithm is launched, taking the currently monitored pollution plume terminal and the centroid of the high concentration area as the starting point, and inversely calculating along the pollution diffusion trajectory. The diffusion velocity deviation is corrected by combining dynamic conditions such as ocean current wind speed. Finally, the coordinates of the pollution diffusion starting point are calculated as the preliminary location of the pollution source, and the confidence level of the location is output.

[0061] By combining real-time satellite remote sensing monitoring data with anomaly datasets, the specific type of pollution source can be determined.

[0062] The pollution source locations are classified as follows: If satellite imagery shows significant oil film reflection characteristics at the location, and the concentration gradient of volatile organic pollutants monitored by buoys is higher than the preset value, the pollution source is identified as a ship oil spill or an oil leak from an offshore platform; if satellite imagery shows no significant oil film characteristics at the location, but the concentration gradient of phycobiliproteins monitored by buoys is higher than the preset value, the pollution source is identified as a red tide; if the location is near the coastline, and buoy monitoring shows significantly higher microbial metabolic activity, the pollution source is identified as a shore-based sewage outlet; if there are industrial sewage pipeline markers around the location, and the particle size distribution of suspended particulate matter matches the characteristics of industrial pollution, the pollution source is identified as an industrial wastewater discharge.

[0063] S5. Send the pollution range and preliminary location coordinates of the pollution source to the verification platform, execute the on-site verification strategy, and receive the on-site verification data; The on-site verification strategy uses drone verification. Based on the pollution boundary in the preliminary pollution range map, the drone swarm conducts a full-coverage scan with a resolution of 10 meters to collect multispectral data and concentration data. The data is then compared with the preliminary pollution range map to eliminate natural interference areas that are misjudged as pollution areas and to supplement any missed diffusion areas, thereby improving the accuracy of the verification.

[0064] Focusing on the initial pollution source coordinates, a fixed-point low-altitude scan was performed using a 5-meter resolution. Simultaneously, concentration detection and multispectral sampling were initiated. The obtained data were compared with the initial location data to verify the actual coordinates, emission status, and specific type of the pollution source. At the same time, the trajectory of the moving pollution source was tracked and the coordinates were updated.

[0065] S6. Analyze the field verification data, pollution range and preliminary location coordinates of pollution sources, real-time satellite remote sensing monitoring data, real-time buoy matrix monitoring data and real-time meteorological data to obtain the pollution diffusion trend in the future period; When analyzing field verification data, pollution range and preliminary location coordinates of pollution sources, real-time satellite remote sensing monitoring data, real-time buoy matrix monitoring data, and real-time meteorological data, fluid dynamics simulation technology is used to process these data to obtain preliminary prediction results of pollution diffusion trends.

[0066] Using three-dimensional fluid dynamics simulation technology, precise data verified by UAVs, real-time meteorological data, and pollution type data are processed to analyze the fluid movement patterns in the sea area and the diffusion characteristics of pollutants in the fluid, thereby obtaining preliminary prediction results of pollution diffusion trends and clarifying the physical basis of pollution diffusion.

[0067] The spatiotemporal converter neural network technology is used to process satellite remote sensing monitoring data and buoy matrix monitoring data to analyze the spatiotemporal dependence of pollution diffusion, nonlinear variation characteristics, and sources of deviation in preliminary prediction results, thereby obtaining deviation correction parameters.

[0068] By employing result fusion and error calibration techniques, the preliminary prediction results and deviation correction parameters are processed to analyze the magnitude and causes of the deviation between the preliminary prediction results and the actual pollution diffusion patterns, thereby deriving the pollution diffusion trend in the future period.

[0069] The pollution diffusion trend in the future period covers the pollution diffusion range, concentration peak changes and diffusion speed in the next 6 hours, 12 hours and 24 hours. It is presented in the form of visual charts and text reports, clearly defining the pollution boundary and core area concentration in each period, marking the areas exceeding the standard, and presenting the concentration peak change curve.

[0070] S7. Based on the preset comprehensive evaluation system, the indicators are evaluated according to four categories: abnormal pollution concentration, diffusion rate, ecological hazard, and social impact. The pollution score is obtained by weighted summation, and the corresponding early warning strategy and dynamic adjustment strategy are implemented until the pollution is eliminated.

[0071] For example, in scoring: abnormal pollution concentrations are scored based on the multiple of the exceedance: 10 points for exceedance up to 1 times, 20 points for exceedance 1-2 times, and 30 points for exceedance more than 2 times; if the pollution concentration shows a continuous upward trend, an additional 5 points are added; diffusion speed is scored based on diffusion distance per hour: 5 points for diffusion distance less than or equal to 0.5 kilometers per hour, 15 points for diffusion distance between 0.5 and 1 kilometer per hour, and 25 points for diffusion distance exceeding 1 kilometer per hour; if the diffusion direction points towards ecologically sensitive areas or densely populated areas, an additional 5 points are added; ecological hazards are scored based on the overlap between the polluted area and the ecologically sensitive area: 5 points for no overlap, 15 points for overlap between 10% and 30%, and 25 points for overlap exceeding 30%; social impact is scored based on the distance between the polluted area and the human activity area: 5 points for a distance of more than 5 kilometers, 12 points for a distance between 1 and 5 kilometers, and 20 points for a distance of less than 1 kilometer.

[0072] The values ​​1-2, 0.5-1, 10%-30%, and 1-5 mentioned above all include the upper limit.

[0073] The corresponding early warning strategies are as follows: When the pollution score is higher than 80 points, it is judged as Level 1 pollution, an emergency response is initiated, operations in the polluted area are suspended, cleaning equipment is deployed, high-frequency monitoring is carried out and early warning information is issued, and a support request signal is sent; when the pollution score is between 60 and 80 points, it is judged as Level 2 pollution, a special response is initiated, activities in the polluted area are restricted, the responsible parties are notified to rectify the situation, medium- and high-frequency monitoring is implemented and early warning information is issued; when the pollution score is lower than 60 points, it is judged as Level 3 pollution, a routine response is initiated, monitoring is strengthened, the pollution spread trend is tracked, the responsible parties are notified to conduct inspections and warning information is issued.

[0074] Example 2: like Figure 2 As shown, a deep learning-based marine environmental monitoring and early warning system comprises five layers: a data support layer, a data acquisition layer, an application analysis layer, an application verification layer, and an analysis and early warning layer. The data support layer forms the foundation of the system and includes a historical analysis module. The data acquisition layer primarily provides real-time data and includes a historical analysis module. The application acquisition layer mainly analyzes the data to determine whether pollution has occurred. The application analysis layer includes a pollution determination module and an analysis and location module. The application verification layer primarily verifies and determines whether pollution exists and includes an on-site verification module. The analysis and early warning layer further analyzes the data, determines the pollution situation, and outputs relevant information. The analysis and early warning layer includes a pollution prediction module and a detection and early warning module. Historical analysis module: Organizes multi-source historical monitoring data, performs predictive analysis, and obtains a set of core feature standards within a preset period; Anomaly detection module: Divides regions, collects and analyzes satellite remote sensing monitoring data and buoy matrix monitoring data in each region in real time, obtains satellite scores and matrix scores, and processes the edge depth model constructed by combining the input of the core feature standard set to obtain preliminary detection results; Pollution determination module: Analyzes preliminary determination results. When the preliminary determination result is abnormal, the hybrid depth model is activated to analyze the core feature standard set, satellite remote sensing monitoring data and buoy matrix monitoring data to obtain the pollution type and confidence level. Analysis and positioning module: The buoy-based contour line generation method is used to summarize concentration data, generate pollution concentration contour lines and upload coordinates. The Kriging interpolation algorithm is used in combination with satellite imagery to draw a preliminary pollution range map. Based on the preliminary pollution range map, the reverse deduction algorithm is used to calculate the preliminary coordinates of the pollution source, and the pollution range and the preliminary positioning coordinates of the pollution source are obtained. Field verification module: Sends the pollution range and preliminary location coordinates of the pollution source to the verification platform, executes the field verification strategy, and receives the field verification data; Pollution prediction module: Analyzes on-site verification data, pollution range and preliminary location coordinates of pollution sources, real-time satellite remote sensing monitoring data, real-time buoy matrix monitoring data and real-time meteorological data to obtain the pollution diffusion trend in the future period; Detection and early warning module: Based on a preset comprehensive evaluation system, it evaluates the scores of four categories of indicators: abnormal pollution concentration, diffusion rate, ecological hazard, and social impact. The scores are then weighted and summed to obtain the pollution score, and corresponding early warning and dynamic adjustment strategies are implemented until the pollution is eliminated.

[0075] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0076] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0077] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A deep learning-based method for marine environmental monitoring and early warning, characterized in that, Includes the following steps: By analyzing historical monitoring data from multiple sources and performing predictive analysis, a set of core feature standards for a preset period is obtained. The region is divided, and satellite remote sensing monitoring data and buoy matrix monitoring data in each region are collected and analyzed in real time to obtain satellite scores and matrix scores. The edge depth model constructed by combining the input of the core feature standard set is then processed to obtain preliminary judgment results. Analyze the preliminary judgment results. When the preliminary judgment results are abnormal, start the hybrid depth model to analyze the core feature standard set, satellite remote sensing monitoring data and buoy matrix monitoring data to obtain the pollution type and confidence level. The concentration data was summarized using the buoy-based contour line generation method to generate pollution concentration contour lines and upload the coordinates. A preliminary pollution range map was drawn using the Kriging interpolation algorithm combined with satellite imagery. Based on the preliminary pollution range map, the preliminary coordinates of the pollution source were calculated using a reverse extrapolation algorithm to obtain the pollution range and the preliminary location coordinates of the pollution source. Send the pollution range and preliminary location coordinates of the pollution source to the verification platform, execute the on-site verification strategy, and receive on-site verification data; By analyzing on-site verification data, pollution range and preliminary location coordinates of pollution sources, real-time satellite remote sensing monitoring data, real-time buoy matrix monitoring data and real-time meteorological data, the pollution diffusion trend in the future period can be obtained. Based on a pre-set comprehensive assessment system, the indicators are scored according to four categories: abnormal pollution concentration, diffusion rate, ecological hazard, and social impact. The scores are then weighted and summed to obtain the pollution score. Corresponding early warning strategies and dynamic adjustment strategies are then implemented until the pollution is eliminated.

2. The deep learning-based marine environmental monitoring and early warning method according to claim 1, characterized in that: The multi-source historical monitoring data covers historical satellite remote sensing monitoring data, historical meteorological data, and historical buoy matrix monitoring data. Among them, the satellite remote sensing monitoring data includes hyperspectral satellite imagery, thermal infrared satellite imagery, synthetic aperture radar satellite scattering coefficient data, and spatial distribution data of water quality parameters, as well as chlorophyll fluorescence intensity, colored soluble organic matter absorption coefficient, oil film reflectance, and sea surface temperature obtained through further analysis. The meteorological data includes wind speed, precipitation, and sea surface temperature. The buoy matrix monitoring data includes long-term monitoring sequence data of phycobiliprotein concentration, microbial metabolic activity, suspended particulate matter particle size distribution, and volatile organic pollutant concentration.

3. The deep learning-based marine environmental monitoring and early warning method according to claim 2, characterized in that: Performing predictive analytics includes: Data standardization and structured organization techniques are used to process multi-source historical monitoring data to obtain a standardized data sample set; Spatiotemporal convolutional neural network technology is used to process spatially correlated data in the data sample set to obtain spatial dimension features; We use bidirectional long short-term memory neural network technology to process time-series data in the data sample set to obtain time-dimensional features; Dynamic weight allocation technology is used to process spatial and temporal features to obtain a weighted feature set adapted to different periods; The weighted feature set is processed using feature mapping and range calibration techniques to obtain the dynamic reference range of each type for each period and each spatial grid point, forming the core feature standard set within the preset period, and divided into satellite reference range set and buoy matrix reference range set.

4. The deep learning-based marine environmental monitoring and early warning method according to claim 1, characterized in that: Real-time acquisition and analysis of satellite remote sensing monitoring data includes first performing data cleaning to remove noise and eliminate cloud cover interference, then comparing it with the satellite baseline range set to calculate the deviation ratio of each satellite monitoring element, then scoring it according to the single-element scoring standard, and finally weighting and summing according to the importance of the elements to obtain the satellite anomaly score for each region. The real-time acquisition and analysis of buoy matrix monitoring data includes first performing data cleaning to remove noise, then comparing it with the buoy matrix baseline range set to calculate the deviation ratio of each monitoring element in each buoy, then scoring according to the single-element scoring standard, and finally weighting and summing according to the importance of the elements to obtain the buoy anomaly score of each buoy, and calculating the average of the buoy anomaly scores of all buoys in the region to obtain the matrix score.

5. The deep learning-based marine environmental monitoring and early warning method according to claim 4, characterized in that: The edge depth model processing, constructed by combining the input of the core feature standard set, includes: processing satellite remote sensing monitoring data, buoy matrix monitoring data, satellite scores, and matrix scores using data normalization and integration techniques to obtain a standardized dataset; allocating weights using a cross-attention mechanism, mining and integrating cross-correlation features, and sequentially performing anomaly feature extraction, inference operations, and balance deviation verification to obtain anomaly confidence and balance coefficients; and performing anomaly threshold judgment based on the anomaly confidence and balance coefficients. If the anomaly confidence is lower than a preset confidence threshold and the balance coefficient is lower than a set standard threshold, it is judged as normal; otherwise, it is judged as abnormal.

6. The deep learning-based marine environmental monitoring and early warning method according to claim 5, characterized in that: The process of data analysis by the hybrid depth model involves using data pruning and time series processing techniques to analyze the core feature standard set, satellite remote sensing monitoring data, and buoy matrix monitoring data to obtain a standard feature dataset. By using convolutional neural network technology to analyze satellite remote sensing image data and satellite baseline range set in the standard feature dataset, a satellite pollution spatial feature set is obtained. By using converter neural network technology to analyze buoy matrix monitoring data and buoy matrix baseline range set in the standard feature dataset, a buoy pollution temporal feature set is obtained. By using cross-attention mechanism technology to analyze the satellite pollution spatial feature set and the buoy pollution temporal feature set, the core identification feature set is obtained after feature fusion. Classification reasoning is performed on the core identification feature set to obtain the pollution type and pollution confidence level.

7. A deep learning-based marine environmental monitoring and early warning method according to claim 6, characterized in that: The steps for generating the pollution range and preliminary location coordinates of the pollution source are as follows: The real-time collected buoy matrix monitoring data is verified, invalid concentration data is removed, and then all data collected by buoys that are initially judged to be abnormal are summarized to construct an abnormal dataset. The contamination gradient is obtained by calculating the extent by which all data in the anomalous dataset exceeds the baseline range of the buoy matrix. The pollution gradient contour lines were generated using linear interpolation, key coordinate points were selected, and the data was transmitted to the satellite data processing center. The key coordinate points were calibrated, and the Kriging interpolation algorithm was used to fill the monitoring blind spots by combining satellite image features to draw a preliminary pollution range map. Based on the interpolation results, the pollution boundary is delineated using the set pollution gradient as the critical value, and the coordinates of the core pollution area are determined. By summarizing satellite pollution plume trajectories, anomalous datasets, real-time meteorological data, and prior baseline data, spatiotemporal alignment and preprocessing are completed, and data interfered with by natural factors are removed to obtain the analysis dataset. The dataset was processed and analyzed using a time-space converter neural network, and the pollution diffusion starting point was calculated through a reverse inference algorithm to obtain the preliminary location coordinates of the pollution source. By combining real-time satellite remote sensing monitoring data with anomaly datasets, the specific type of pollution source can be determined.

8. A deep learning-based marine environmental monitoring and early warning method according to claim 7, characterized in that: When analyzing field verification data, pollution range and preliminary location coordinates of pollution sources, real-time satellite remote sensing monitoring data, real-time buoy matrix monitoring data, and real-time meteorological data, fluid dynamics simulation technology is used to process the data to obtain preliminary prediction results of pollution diffusion trends. Spatiotemporal converter neural network technology is used to process satellite remote sensing monitoring data and buoy matrix monitoring data to obtain deviation correction parameters. Result fusion and error calibration technology is used to process the preliminary prediction results and deviation correction parameters to obtain the pollution diffusion trend for future periods.

9. A deep learning-based marine environmental monitoring and early warning method according to claim 8, characterized in that: The corresponding early warning strategies are as follows: When the pollution score is higher than 80 points, it is determined to be Level 1 pollution, an emergency response is initiated, activities in the polluted area are suspended, high-frequency monitoring is carried out and early warning information is issued, and a support request signal is sent out; when the pollution score is between 60 and 80 points, it is determined to be Level 2 pollution, a special response is initiated, activities in the polluted area are restricted and rectification is carried out, medium- and high-frequency monitoring is implemented and early warnings are issued; when the pollution score is lower than 60 points, it is determined to be Level 3 pollution, a routine response is initiated, monitoring work is strengthened, the trend of pollution spread is tracked, patrols are carried out and warning information is issued.

10. A deep learning-based marine environmental monitoring and early warning system, characterized in that, include: Historical analysis module: sorts through multi-source historical monitoring data, performs predictive analysis, and obtains a set of core feature standards within a preset period; Anomaly detection module: Divides regions, collects and analyzes satellite remote sensing monitoring data and buoy matrix monitoring data in each region in real time, obtains satellite scores and matrix scores, and processes the edge depth model constructed by combining the input of the core feature standard set to obtain preliminary detection results; Pollution determination module: Analyzes preliminary determination results. When the preliminary determination result is abnormal, the hybrid depth model is activated to analyze the core feature standard set, satellite remote sensing monitoring data and buoy matrix monitoring data to obtain the pollution type and confidence level. Analysis and positioning module: The buoy-based contour line generation method is used to summarize concentration data, generate pollution concentration contour lines and upload coordinates. The Kriging interpolation algorithm is used in combination with satellite imagery to draw a preliminary pollution range map. Based on the preliminary pollution range map, the reverse deduction algorithm is used to calculate the preliminary coordinates of the pollution source, and the pollution range and the preliminary positioning coordinates of the pollution source are obtained. Field verification module: Sends the pollution range and preliminary location coordinates of the pollution source to the verification platform, executes the field verification strategy, and receives the field verification data; Pollution prediction module: Analyzes on-site verification data, pollution range and preliminary location coordinates of pollution sources, real-time satellite remote sensing monitoring data, real-time buoy matrix monitoring data and real-time meteorological data to obtain the pollution diffusion trend in the future period; Detection and early warning module: Based on a preset comprehensive evaluation system, it evaluates the scores of four categories of indicators: abnormal pollution concentration, diffusion rate, ecological hazard, and social impact. The scores are then weighted and summed to obtain the pollution score, and corresponding early warning and dynamic adjustment strategies are implemented until the pollution is eliminated.

Images