Water body algae bloom dynamic monitoring and early warning system based on data analysis

Abstract

The application relates to the technical field of environmental monitoring, and particularly discloses a water body algal bloom dynamic monitoring and early warning system based on data analysis. The system comprises a multi-source heterogeneous data fusion module, a space-time feature dynamic extraction module, a water bloom risk probability evolution module and an early warning decision and instruction generation module which are connected in sequence. Through the fusion of multi-source data and the extraction of space-time dynamic characteristics, the evolution of the water bloom risk probability is predicted by using a deep probability graph model based on physical constraints, and finally, the hierarchical early warning and structured response instructions are generated according to the risk level, so that the monitoring, early warning and closed-loop management of the algal bloom are realized.

Description

Technical Field

[0001] This invention belongs to the field of environmental monitoring technology, specifically relating to a dynamic monitoring and early warning system for algal blooms in water bodies based on data analysis. Background Technology

[0002] Aquatic environment monitoring and ecological protection is an important interdisciplinary field integrating environmental science and information technology. Its core objective is to achieve real-time assessment of water body health and timely early warning of abnormal events through continuous data collection and analysis, providing a scientific basis for water resource management and ecological restoration. Monitoring and early warning of algal blooms caused by eutrophication are key technological directions for ensuring drinking water safety and maintaining the balance of aquatic ecosystems.

[0003] In existing technologies, the monitoring of algal blooms mainly relies on manual on-site sampling combined with laboratory analysis, or the deployment of a limited number of fixed-point water quality sensors for data collection. These methods have the following problems: manual sampling is time-consuming and costly, and it is difficult to achieve simultaneous monitoring of large-scale water bodies, resulting in insufficient spatiotemporal resolution of the data and an inability to capture the rapid dynamic changes of algal blooms; while single-point sensor monitoring can provide continuous data, its coverage is extremely limited, making it difficult to reflect the spatial distribution pattern and migration and diffusion trend of algae in the entire water body.

[0004] While multi-source heterogeneous data such as satellite remote sensing and meteorological observation contain rich information, existing systems face challenges in data fusion. The lack of effective semantic association and collaborative analysis mechanisms between data from different sources and in different formats prevents the full exploitation of data value.

[0005] Existing early warning models are mostly based on static thresholds or simple historical data statistics, which are difficult to adapt to the complex mechanisms by which algal growth is affected by the nonlinear coupling of multiple environmental factors such as temperature, light, and nutrients. These models lack the ability to dynamically learn and iteratively optimize the deep patterns in time-series data, resulting in early warning results lagging behind the actual development of algal blooms. The accuracy and timeliness of these early warnings fail to meet the needs of refined management. Therefore, how to efficiently integrate multi-source heterogeneous monitoring data and construct an intelligent early warning system capable of real-time tracking and accurate prediction of algal bloom dynamics has become an urgent technical challenge in aquatic ecological management. Summary of the Invention

[0006] The purpose of this invention is to provide a dynamic monitoring and early warning system for algal blooms in water bodies based on data analysis, so as to solve the problems of insufficient spatiotemporal resolution of monitoring data, difficulty in fusion of multi-source heterogeneous data, and the inability of early warning models to adapt to the nonlinear dynamic changes of algal blooms, which lead to delayed early warning and low accuracy.

[0007] This invention provides a dynamic monitoring and early warning system for algal blooms in water bodies based on data analysis, comprising: The multi-source heterogeneous data fusion module is used to receive and process data from different monitoring sources and construct a multimodal data cube under a unified spatiotemporal reference. The spatiotemporal feature dynamic extraction module is connected to the multi-source heterogeneous data fusion module and is used to extract dynamic spatiotemporal features characterizing the occurrence and development of algal blooms from the unified multimodal data cube; the algal bloom risk probability evolution module is connected to the spatiotemporal feature dynamic extraction module and is used to simulate and predict the risk probability evolution process of algal blooms in each grid unit in the future period based on the extracted dynamic spatiotemporal features. The early warning decision and instruction generation module is connected to the algal bloom risk probability evolution module and is used to generate graded early warning instructions and specific response strategy suggestions based on the evolution results of the risk probability.

[0008] Preferably, the multi-source heterogeneous data fusion module includes a data access submodule, a spatiotemporal alignment submodule, and a semantic association submodule; The data access submodule is used to access satellite remote sensing data, ground-based fixed-point water quality sensor network data, meteorological observation station data, and data reported by manual patrols. The spatiotemporal alignment submodule is used to use geographic information systems and time series interpolation algorithms to unify the spatiotemporal resolution of all data sources from the data access submodule to a preset grid scale and time step. The semantic association submodule is used to establish physical and biochemical associations between different data modalities based on a preset domain knowledge graph, and to encode all associations in the form of a weighted adjacency matrix.

[0009] Preferably, the spatiotemporal feature dynamic extraction module includes a spatial pattern analysis submodule and a time series analysis submodule; The spatial pattern analysis submodule is used to calculate the spatial statistics of algae-related parameters within a preset grid area at each time step using a sliding window algorithm. The time series analysis submodule is used to apply an adaptive filtering algorithm to the time series data of each grid cell to separate the long-term trend term, seasonal periodic term, and high-frequency residual term, and to calculate the abrupt change point and persistence index of the residual term.

[0010] Preferably, the core of the algal bloom risk probability evolution module is a deep probabilistic graphical model based on physical constraints; The physical constraint-based deep probabilistic graphical model uses algal biomass as a latent state variable and extracted spatiotemporal features and environmental driving factors as observation variables to construct state transition equations and observation equations. The state transition equation incorporates a simplified kinetic model of algal growth, and its parameters are obtained through training with historical data. The observation equation describes the likelihood relationship between the hidden state and multi-source observation data; The operation process of the algal bloom risk probability evolution module is as follows: at the beginning of each prediction period, the posterior probability distribution of the hidden state is updated by using the observation data at the current moment as a condition through variational inference algorithm. Furthermore, by using the posterior distribution of the previous moment as the prior, and combining the state transition equation with the predicted values ​​of future environmental factors, multiple time steps are extrapolated forward to generate a spatiotemporal distribution map of future risk probabilities.

[0011] Preferably, the early warning decision and instruction generation module includes a threshold comparison submodule, a decision logic submodule, and an instruction encapsulation submodule; The threshold comparison submodule is used to compare the risk probabilities of each future time step and each spatial grid output by the algal bloom risk probability evolution module with the preset multi-level probability thresholds in real time. The decision logic submodule is used to perform hierarchical decision-making based on the comparison results; The instruction encapsulation submodule is used to generate a structured response instruction set containing suggested sampling points, suggested patrol routes, and suggested agent dosing areas and dosages based on the predicted location, diffusion direction and speed of the algal bloom core area and the preset emergency resource library and action plan library when an action-level early warning is triggered. The instruction is then sent to the corresponding execution terminal through the communication interface.

[0012] Preferably, the construction and updating process of the domain knowledge graph in the semantic association submodule is as follows: the initial association rules and weights defined by experts are loaded during system initialization; During system operation, the statistical correlation between different parameter sequences in historical data is continuously analyzed through association rule mining algorithms; When a strong relationship is discovered that is inconsistent with the initial rules or is a newly added relationship, a manual review process is triggered. Once confirmed, the strong association and its confidence level are added to the knowledge graph and used to update the weighted adjacency matrix of semantic associations.

[0013] Preferably, the adaptive filtering algorithm in the time series analysis submodule adopts a combination of empirical mode decomposition and ensemble empirical mode decomposition. The specific process is as follows: First, empirical mode decomposition is performed on the univariate time series of each grid cell to obtain a series of intrinsic mode functions; To address the potential mode aliasing problem in the decomposition results, ensemble empirical mode decomposition is introduced. By adding paired white noise and performing multiple decompositions and averaging, more stable intrinsic mode function components are obtained, thereby more accurately separating signal components at different time scales.

[0014] Preferably, the state transition equation is a discretized differential equation that describes the rate of change of algal biomass over time as a function of current biomass, water temperature, light intensity, and nutrient concentration. The specific parameterized form of the function is represented by a deep neural network. The input to the deep neural network is the current hidden state and environmental factors, and the output is the parameters of the state transition. It is trained end-to-end using the negative log-likelihood loss function between historical observation data and model prediction results.

[0015] Preferably, the setting of multi-level probability thresholds in the decision logic submodule adopts a dynamic adjustment mechanism; this dynamic adjustment mechanism is optimized based on the accuracy and false negative rate of historical warnings. The specific process is as follows: the system periodically compiles statistics on the actual occurrence of algal blooms after the issuance of warnings at all levels within the past cycle, and calculates the accuracy and recall rate of the warning at that level. The probability threshold for the corresponding level is dynamically adjusted based on the weighted harmonic mean of precision and recall. If the precision is too low, the threshold is raised to reduce false positives, and if the recall is too low, the threshold is lowered to reduce false negatives.

[0016] Preferably, the decision logic submodule performs the hierarchical decision-making process based on the comparison results as follows: When the risk probability exceeds the Level 1 threshold, a warning at the level of concern is generated, and the system is instructed to increase the data collection frequency. When the risk probability exceeds the Level 2 threshold, a warning level alert is generated, and the spatial coordinate range of the suspected algal bloom core area is automatically generated. When the risk probability exceeds the Level 3 threshold, an action-level warning is generated.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention effectively solves the heterogeneity problems in format, scale, and semantics of multi-source data such as satellite remote sensing, ground sensing, and meteorological data by constructing a multi-source heterogeneous data fusion module and designing a processing flow that includes spatiotemporal alignment and semantic association. The system can automatically integrate scattered and heterogeneous data into a unified spatiotemporal framework and establish their inherent physical and biochemical relationships, forming a high-quality, high-information-density multimodal data cube. This provides a reliable and consistent data foundation for subsequent in-depth analysis, improving data utilization efficiency and value.

[0018] 2. This invention utilizes a spatiotemporal feature dynamic extraction module, employing a combination of spatial pattern analysis and time series decomposition, to systematically extract dynamic features reflecting the spatial aggregation, heterogeneity, temporal trends, cycles, and abrupt changes of algal blooms from the fused data. This feature extraction method transcends static judgment based on single parameter thresholds, enabling more sensitive capture of early anomalous signals and spatial propagation patterns before algal blooms occur, providing crucial information input for risk probability prediction.

[0019] 3. This invention creatively proposes a probabilistic evolution module for algal bloom risk, the core of which is a deep probabilistic graphical model based on physical constraints. This deep probabilistic graphical model embeds the mechanism of algal growth into a probabilistic inference framework in the form of differential equations, while utilizing deep neural networks to learn complex nonlinear relationships, achieving a deep integration of the mechanism model and the data-driven model. This design enables the model not only to output the probability of future algal blooms but also to provide a measure of the probability's uncertainty, and its prediction results have clear physical interpretability. Through continuous variational inference and forward extrapolation, the system achieves dynamic and forward-looking simulation of algal bloom risk, improving the timeliness and accuracy of early warning.

[0020] 4. This invention achieves a closed loop from risk prediction to concrete action through an early warning decision-making and instruction generation module. The module provides intelligent hierarchical early warning based on dynamically adjusted multi-level probability thresholds and can automatically match and generate structured response instruction sets according to the predicted location and spread of algal bloom core areas. This design directly transforms early warning information into actionable guidelines, significantly shortening the time from risk discovery to countermeasures, improving the efficiency and accuracy of emergency response, and providing strong technical support for scientific decision-making and rapid intervention by water management departments. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the physical constraint-based depth probabilistic graphical model in this invention; Figure 3 This is a logical flow diagram of the multi-source heterogeneous data fusion module in this invention; Figure 4 This is a logical flowchart of the spatiotemporal feature dynamic extraction module in this invention; Figure 5 This is a logical flowchart of the early warning decision and instruction generation module in this invention. Detailed Implementation

[0022] Example 1: Reference Figures 1 to 5The present invention provides a dynamic monitoring and early warning system for algal blooms in water bodies based on data analysis. This system consists of four core functional modules: a multi-source heterogeneous data fusion module, a spatiotemporal feature dynamic extraction module, an algal bloom risk probability evolution module, and an early warning decision and command generation module.

[0023] The modules are interconnected in an orderly manner through pre-defined data interfaces and logical protocols, forming a complete closed-loop process from raw data acquisition to structured emergency command output. The entire system runs on a cloud-edge collaborative computing architecture, with the edge side responsible for real-time processing of high-frequency raw data and the cloud side responsible for global risk simulation and centralized decision-making, thus balancing response speed and computational depth.

[0024] First, the multi-source heterogeneous data fusion module serves as the system's data entry point and preprocessing hub; its internal structure is shown in the attached figure. Figure 3 As shown, this multi-source heterogeneous data fusion module comprises three sub-modules: a data access sub-module, a spatiotemporal alignment sub-module, and a semantic association sub-module. The data access sub-module is configured with multiple standardized communication interfaces, each used to receive data streams from different monitoring sources.

[0025] Specifically, satellite remote sensing data is accessed in raster image format through the standard service interface provided by the national or regional remote sensing data center, with a spatial resolution of 30 meters to 1000 meters and a temporal resolution of once per day or once every 3 days. The ground-based fixed-point water quality sensor network data is accessed via the Internet of Things protocol and includes parameters such as chlorophyll a concentration, dissolved oxygen, pH value, turbidity, water temperature, and conductivity. The sampling frequency is once every 5 minutes to once per hour. Meteorological observation station data is accessed through the meteorological department's open data interface, including temperature, air pressure, wind speed, wind direction, sunshine duration, precipitation, etc., with a time resolution of one per hour; Data from manual surveys is uploaded via a mobile application in the form of a structured form. The content includes qualitative and semi-quantitative information such as visual algae coverage area, water color, odor description, and on-site photos. The reporting frequency is dynamically adjusted according to the task plan, usually once or twice a week.

[0026] All incoming data is first temporarily stored in a distributed message queue, awaiting further processing. The spatiotemporal alignment submodule then starts, its core task being to unify the aforementioned heterogeneous data to a preset spatiotemporal grid reference. Spatially, the system adopts a universal transverse Mercator projection coordinate system and divides it into square grid cells with sides of 100 meters, covering the target water body and its surrounding buffer zone. Temporally, the system sets a basic time step of 6 hours, meaning each day is divided into 4 time slices.

[0027] For data with a temporal resolution greater than 6 hours (such as ground sensor data), linear interpolation or cubic spline interpolation algorithms are used for downsampling; for data with a temporal resolution less than 6 hours (such as some satellite data), nearest neighbor filling or forward prediction based on historical trends is used for upsampling. In the spatial dimension, remote sensing raster data is mapped to a 100-meter grid through bilinear resampling; point sensor data is diffused to the surrounding grid using inverse distance weighted interpolation; manually surveyed points are spatially smoothed using a Gaussian kernel function with an influence radius of 200 meters centered on their geographic coordinates.

[0028] After the above processing, all data are organized into a four-dimensional tensor structure, with dimensions of time step, grid row index, grid column index, and data modality type, forming a so-called "multimodal data cube".

[0029] Next, the semantic association submodule performs semantic-level association modeling on the various modalities within the data cube. The core of this submodule is a pre-defined domain knowledge graph, where nodes represent physical or biochemical variables, edges represent causal or correlational relationships between variables, and each edge carries weight and direction attributes. For example, the knowledge graph contains a directed edge from "ratio of remote sensing red light to near-infrared bands" to "chlorophyll a concentration" with a weight of 0.85, indicating a strong positive correlation between this spectral index and chlorophyll concentration; another edge points from "average daily air temperature" to "surface water temperature" with a weight of 0.72, reflecting the impact of meteorological conditions on the thermal state of water bodies.

[0030] The knowledge graph for this domain was constructed by domain experts during the system initialization phase based on long-term observation experience and literature review, containing approximately 200 nodes and 350 relational edges. During system operation, the semantic association submodule continuously performs association rule mining, specifically employing a method combining Pearson correlation coefficient within a sliding time window and Granger causality test to analyze the statistical dependencies between different parameter sequences within each grid cell. When a new association strength between a pair of variables is detected to be greater than the threshold of 0.6 and its duration is greater than 7 days, the system automatically generates an association rule change request, triggering a manual review process.

[0031] Once approved, the new relationship is incorporated into the knowledge graph, and the corresponding weighted adjacency matrix is ​​updated. Finally, the semantic association submodule outputs a sequence of weighted adjacency matrices that evolves over time. The rows and columns of each matrix correspond to the modal dimensions in the data cube, and the matrix element values ​​represent the semantic association strength, serving as an important input for subsequent modules.

[0032] After data fusion is completed, the spatiotemporal feature dynamic extraction module begins operation, and its internal logic flow is shown in the attached figure. Figure 4As shown, the spatiotemporal feature dynamic extraction module receives the multimodal data cube output by the multi-source heterogeneous data fusion module and extracts two types of key features from it: spatial pattern features and time series features. The spatial pattern analysis submodule traverses the entire target area using a sliding window at each time step. The window size is set to 5×5 grid cells (i.e., 500 meters × 500 meters), and the step size is 1 grid cell.

[0033] Within each window, the following statistics are calculated for core algae-related parameters (such as chlorophyll a concentration, remotely sensed algal bloom index, turbidity, etc.): the arithmetic mean of the parameters within the window, used to characterize the local algal biomass level; the standard deviation, reflecting the degree of spatial heterogeneity; and the Moran index. It is used to quantify spatial autocorrelation, and the calculation formula is: ; This represents the total number of grid cells within the window. Spatial weights are defined based on Euclidean distance (the closer the distance, the higher the weight). For the first The parameter values ​​of each grid cell, For the first The parameter values ​​of each grid cell, The mean, For all The sum of these parameters is also calculated. Furthermore, the gradient vector field of the parameter field is calculated, and the partial derivatives of each grid cell in the east-west and north-south directions are obtained using the central difference method, thus yielding the gradient magnitude and principal direction, which are used to identify the boundaries and diffusion trends of algal bloom patches. All spatial statistics are reorganized into feature tensors of the same dimensions as the original data cube, with each time step corresponding to a set of spatial feature maps.

[0034] Meanwhile, the time series analysis submodule decomposes the univariate time series for each grid cell. Taking chlorophyll a concentration as an example, its original sequence is... This time series analysis submodule first applies the empirical mode decomposition algorithm to adaptively decompose the sequence into several intrinsic mode function (IMF) components. and residuals ,satisfy .

[0035] However, Empirical Mode Decomposition (EMD) is prone to mode aliasing when processing non-stationary signals, meaning that a single IMF component contains oscillations across multiple time scales. To overcome this problem, the system further introduces an ensemble EMF strategy: adding... (Total) White noise with an amplitude of 0.2 times the standard deviation of the original sequence was used for Empirical Mode Decomposition (EMD). Then, the arithmetic mean of the IMF components of the same order was taken to obtain more stable components. In practice, The value is 100.

[0036] After decomposition, the system classifies each IMF component according to its average period length: those with a period greater than 30 days are classified as long-term trend components, those with a period of 7 to 30 days are classified as seasonal period components, and those with a period of less than 7 days are classified as high-frequency residual components. For high-frequency residual components, the system calculates their abrupt change points—that is, the moments when the change amplitude within three consecutive time steps is greater than three times the standard deviation—and calculates the duration of the abnormal state after the abrupt change.

[0037] These time-scale features (trend slope, periodic amplitude, abrupt change location, and persistence indicators) are recorded grid-by-grid to form a time feature tensor. Finally, the outputs of the spatial pattern analysis submodule and the time series analysis submodule are concatenated and merged to form a complete dynamic spatiotemporal feature set, which serves as the direct input to the algal bloom risk probability evolution module.

[0038] The algal bloom risk probability evolution module is the core prediction engine of this system, and its principle framework is shown in the attached figure. Figure 2 As shown. This algal bloom risk probability evolution module constructs a physically constrained deep probabilistic graphical model, which includes algal biomass. Treating these as latent state variables, the aforementioned extracted dynamic spatiotemporal features and environmental driving factors (such as water temperature) are considered as latent state variables. Light intensity Nutrient concentration The hidden state is considered as an observed variable. The model contains two core equations: the state transition equation and the observation equation. The state transition equation describes the evolution of the hidden state over time, and its form is a discretized differential equation: ; The algal biomass at time t+1 The time step is 6 hours. It is a function parameterized by a deep neural network. These are its learnable parameters. This deep neural network adopts a three-layer fully connected structure, with the input layer receiving... Four scalars, 64 neurons in the hidden layer, ReLU activation function, and a single neuron in the output layer, representing the net growth rate of algae per unit time.

[0039] The design embeds the classical Monod equation for algal growth and mechanisms such as light suppression into the prior constraints of the network structure, while retaining the ability to learn complex nonlinear relationships through data-driven learning. The observation equation defines the likelihood relationship between the hidden state and multi-source observation data, assuming that the observation noise follows a Gaussian distribution, and its covariance matrix is ​​dynamically adjusted by the weighted adjacency matrix output by the semantic association submodule.

[0040] The model's operation consists of two phases: state inference and risk extrapolation. At the beginning of each prediction period (e.g., 00:00 daily), the system approximates the hidden state using variational inference algorithms, based on the latest observation data. posterior probability distribution The objective of variational inference is to minimize the KL divergence between the posterior distribution and the true distribution. The optimization process employs stochastic gradient descent, with an upper limit of 500 iterations or a change in the loss function of less than 10%. Stop when the time comes.

[0041] After obtaining the current posterior distribution, the system enters the risk extrapolation stage: using the previous posterior distribution as the initial prior, combined with the environmental factor predictions provided by the meteorological department for the next 72 hours (such as future water temperature, light intensity, etc.), 1000 possible algal biomass evolution trajectories are generated from the forward state transition equation through Monte Carlo sampling.

[0042] For each trajectory, determine at each future time step algal biomass at time Does it exceed the preset algal bloom threshold (e.g., biomass corresponding to a chlorophyll a concentration of 50 μg / L)? Calculate the proportion of all trajectories exceeding the threshold; this yields the percentage of that grid cell that exceeded the threshold. The probability of algal bloom occurring at any given time. This process is executed in parallel on all grid cells, ultimately generating a spatiotemporal distribution map of the risk probability that is updated every 6 hours over the next 12 time steps (i.e., 72 hours).

[0043] Finally, the early warning decision-making and instruction generation module executes a tiered response based on the risk probability distribution map, and its logical flow is shown in the attached figure. Figure 5 As shown. This early warning decision and instruction generation module includes a threshold comparison submodule, a decision logic submodule, and an instruction encapsulation submodule. The threshold comparison submodule maintains a three-level dynamic probability threshold system: Level 1 threshold... The corresponding alert level is set to an initial value of 0.3; the second-level threshold... The corresponding alert level warning has an initial value set at 0.6; the third-level threshold... The corresponding action-level warning has an initial value set to 0.85.

[0044] In each prediction period, the threshold comparison submodule iterates through all future time steps and spatial grids to determine the risk probability. The comparison is performed point by point with the three-level threshold. The decision logic submodule executes the corresponding action based on the comparison result.

[0045] If in any grid At any point in the future satisfy This triggers a high-level alert, instructing the multi-source heterogeneous data fusion module to temporarily increase the data collection frequency for the area to once per hour and extend the priority of satellite data revisit requests.

[0046] like If the condition is met, a warning level alert will be triggered. The system will automatically identify all continuous grid regions that meet the condition, eliminate isolated noise points through morphological closing operations, generate polygonal boundary coordinates of suspected algal bloom core areas, and push them to the management platform's visualization interface.

[0047] like This triggers an action-level alert, and the instruction encapsulation submodule starts immediately.

[0048] The instruction encapsulation submodule integrates an emergency resource database and an action plan database. The emergency resource database contains the real-time location and status of available resources such as unmanned surface vessels, drones, mobile laboratories, and chemical storage points; the action plan database stores standard operating procedures for different water body types, algal bloom scales, and diffusion patterns.

[0049] When an action-level warning is triggered, the system first delineates a suggested sampling buffer zone within 500 to 2000 meters outside the core area, based on the predicted location of the algal bloom core area, the diffusion direction (determined by the principal gradient direction) and diffusion velocity (calculated from the centroid displacement of the core area over consecutive time steps). Within this buffer zone, the system invokes a spatial coverage algorithm to generate an optimal set of sampling points, ensuring that the distance between points is less than 300 meters and that the main water flow path is covered.

[0050] Simultaneously, combining digital elevation models and hydrodynamic simulation results, the shortest safety patrol route starting from the nearest dock is planned. For chemical dosing, the system calculates the theoretical dosage of copper sulfate or hydrogen peroxide based on the predicted peak algal biomass and water volume, and trims it with reference to the safety upper limits in the contingency plan library, ultimately generating a structured instruction that includes the dosing area polygon, the recommended dosage (unit: kg), and the dosing method (drone spraying or unmanned surface vessel targeted release). All instructions are encapsulated into a message body conforming to the JSON Schema standard and sent to the corresponding execution terminal, such as the unmanned surface vessel control center, environmental law enforcement APP, or water management dispatch platform, via HTTPS protocol.

[0051] It is worth noting that the aforementioned three-level probability thresholds are not statically fixed, but are continuously optimized through a feedback mechanism. The system has a built-in early warning performance evaluation unit that performs a retrospective analysis of all early warning events from the previous month at the beginning of each month.

[0052] For each warning level, the number of grid-time combinations of algal blooms that actually occurred within 72 hours of its issuance (true positives, TP), the number of algal blooms that actually occurred without a warning (false negatives, FN), and the number of algal blooms that did not occur despite a warning (false positives, FP) are calculated based on this. With recall rate And further calculate its weighted harmonic mean. , Set it to 2 to place greater emphasis on recall (i.e., reduce false negatives). It is a true positive. It was a false negative. False positive like If the threshold is less than the preset target value (e.g., 0.75), the threshold is adjusted according to the direction of the deviation: if the precision is too low (many false alarms), the corresponding threshold is increased by 0.05; if the recall is too low (many false negatives), it is decreased by 0.05. The adjusted threshold takes effect after confirmation by the system administrator to ensure that the early warning strategy always maintains the best match with the current ecological state of the water body.

[0053] Throughout the system's operation, the cloud-edge collaborative architecture played a crucial role. Edge computing nodes deployed on the lakeshore or floating platforms were equipped with high-performance embedded processors and solid-state storage, running lightweight versions of the multi-source heterogeneous data fusion module and the spatiotemporal feature dynamic extraction module.

[0054] These nodes only process the raw data within their jurisdiction radius (typically 5 kilometers). The calculated dynamic spatiotemporal features (approximately 5% of the raw data) are encrypted and compressed before being uploaded to the cloud via 4G / 5G or LoRaWAN networks. The cloud central server aggregates the feature data from all edge nodes, stitches them together into a global feature cube, and provides it to the algal bloom risk probability evolution module for global simulation.

[0055] This architecture avoids the bottleneck of remote transmission of massive amounts of raw data while ensuring global consistency in risk prediction. After the warning instruction is generated, the cloud accurately sends the instruction to the relevant edge nodes through the same communication link, and the local execution terminal responds directly, realizing an efficient closed loop of "edge perception - cloud decision-making - edge execution".

[0056] In summary, this embodiment constructs a high-precision, timely, and operable dynamic monitoring and early warning system for algal blooms through rigorous modular design, deeply integrated mechanism and data-driven model, and closed-loop intelligent decision-making mechanism. This system not only solves the technical challenges of multi-source data fusion and feature extraction, but also effectively transforms early warning information into management actions through probabilistic risk evolution and structured instruction generation, significantly improving the scientific rigor and proactiveness of water environment governance.

[0057] Example 2: Building upon Example 1, this example enhances the deep probabilistic graphical model in the algal bloom risk probability evolution module to address the sudden disturbances to algal bloom dynamics caused by extreme weather events. In actual operation, it was found that extreme weather events such as typhoons, heavy rainfall, or sustained high temperatures significantly alter the water's mixing state, nutrient input, and light conditions, leading to discontinuous abrupt changes in algal growth rates. The state transition equation in Example 1, lacking explicit modeling of such abrupt changes, may result in biased risk predictions.

[0058] Therefore, this embodiment introduces an external shock factor into the state transition equation. This system quantifies the instantaneous impact of extreme events on algal growth. The external impact factor is generated by a separate extreme event detection and quantification submodule. This submodule continuously monitors meteorological station data and short-term numerical weather prediction products. An extreme event is identified when any of the following conditions are detected: cumulative precipitation greater than 50 mm within the next 24 hours; wind speed greater than 10 m / s for more than 6 consecutive hours; or daily maximum temperature greater than the 95th percentile for three consecutive days. Once an extreme event is identified, the submodule retrieves the corresponding impact function from a pre-defined impact response library based on the event type and intensity. For example, the impact function for a heavy rainfall event is a negative pulse, with an amplitude proportional to the rainfall intensity and a duration of 24 hours, reflecting the inhibitory effect of rainwater erosion and water dilution on algae; while the impact function for a sustained high-temperature event is a positive step, with an amplitude correlated to the degree of overheating and a duration of up to 72 hours, reflecting the promoting effect of high temperatures on algal metabolism.

[0059] External shock factor When directly superimposed on the right side of the state transition equation, the corrected equation is: ; Furthermore, to improve the model's generalization ability in sparse data regions, this embodiment also introduces spatial smoothing constraints into the observation equation. Specifically, when calculating the observation likelihood of a certain grid cell, not only is the observation data of that grid cell itself considered, but also the observation information of its 8 neighboring grids is weighted and fused, with the weights determined by the inverse squared distance. This effectively alleviates the problem of local data loss caused by individual sensor failures or satellite cloud obstruction, making the risk probability distribution map smoother and more reasonable.

[0060] In the early warning decision-making process, this embodiment adds an independent early warning channel for extreme events themselves. When the extreme event detection submodule is triggered, regardless of the current probability of algal bloom risk, the system will generate a "meteorological disturbance warning," alerting managers that the algal bloom situation may change drastically within the next 72 hours and suggesting the early deployment of mobile monitoring equipment to key sections. This warning is displayed in parallel with the algal bloom risk warning, providing more comprehensive information support for integrated assessment.

[0061] Through the above enhancements, this embodiment improves the robustness and prediction accuracy of the system under complex meteorological conditions, and is particularly suitable for large lakes and reservoirs located in monsoon regions or climate change sensitive areas.

[0062] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0063] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A dynamic monitoring and early warning system for algal blooms in water bodies based on data analysis, characterized in that: include: The multi-source heterogeneous data fusion module is used to receive and process data from different monitoring sources and construct a multimodal data cube under a unified spatiotemporal reference. The spatiotemporal feature dynamic extraction module is connected to the multi-source heterogeneous data fusion module and is used to extract dynamic spatiotemporal features characterizing the occurrence and development of algal blooms from the unified multimodal data cube. The algal bloom risk probability evolution module is connected to the spatiotemporal feature dynamic extraction module and is used to simulate and predict the risk probability evolution process of algal blooms in each grid unit in the future period based on the extracted dynamic spatiotemporal features. The early warning decision and instruction generation module is connected to the algal bloom risk probability evolution module and is used to generate graded early warning instructions and specific response strategy suggestions based on the evolution results of the risk probability.

2. The water body algal bloom dynamic monitoring and early warning system based on data analysis according to claim 1, characterized in that, The multi-source heterogeneous data fusion module includes a data access submodule, a spatiotemporal alignment submodule, and a semantic association submodule; The data access submodule is used to access satellite remote sensing data, ground-based fixed-point water quality sensor network data, meteorological observation station data, and data reported by manual patrols. The spatiotemporal alignment submodule is used to unify the spatiotemporal resolution of all data sources from the data access submodule to a preset grid scale and time step by employing geographic information systems and time series interpolation algorithms. The semantic association submodule is used to establish physical and biochemical associations between different data modalities based on a preset domain knowledge graph, and to encode all associations in the form of a weighted adjacency matrix.

3. The water body algal bloom dynamic monitoring and early warning system based on data analysis according to claim 2, characterized in that, The spatiotemporal feature dynamic extraction module includes a spatial pattern analysis submodule and a time series analysis submodule; The spatial pattern analysis submodule is used to calculate the spatial statistics of algae-related parameters within a preset grid area at each time step using a sliding window algorithm. The time series analysis submodule is used to apply an adaptive filtering algorithm to the time series data of each grid cell to separate the long-term trend term, seasonal periodic term, and high-frequency residual term, and to calculate the abrupt change point and persistence index of the residual term.

4. The water body algal bloom dynamic monitoring and early warning system based on data analysis according to claim 3, characterized in that, The core of the algal bloom risk probability evolution module is a deep probabilistic graphical model based on physical constraints; The physical constraint-based deep probabilistic graphical model uses algal biomass as a latent state variable and extracted spatiotemporal features and environmental driving factors as observation variables to construct state transition equations and observation equations. The state transition equation incorporates a simplified kinetic model of algal growth, and its parameters are obtained through training with historical data. The observation equation describes the likelihood relationship between the hidden state and multi-source observation data; The operation process of the algal bloom risk probability evolution module is as follows: at the beginning of each prediction period, the posterior probability distribution of the hidden state is updated by using the observation data at the current moment as a condition through variational inference algorithm. Furthermore, by using the posterior distribution of the previous moment as the prior, and combining the state transition equation with the predicted values ​​of future environmental factors, multiple time steps are extrapolated forward to generate a spatiotemporal distribution map of future risk probabilities.

5. The water body algal bloom dynamic monitoring and early warning system based on data analysis according to claim 4, characterized in that, The early warning decision and instruction generation module includes a threshold comparison submodule, a decision logic submodule, and an instruction encapsulation submodule. The threshold comparison submodule is used to compare the risk probabilities of each future time step and each spatial grid output by the algal bloom risk probability evolution module with the preset multi-level probability thresholds in real time. The decision logic submodule is used to perform hierarchical decision-making based on the comparison results; The instruction encapsulation submodule is used to generate a structured response instruction set containing suggested sampling points, suggested patrol routes, and suggested agent dosing areas and dosages based on the predicted location, diffusion direction and speed of the algal bloom core area and the preset emergency resource library and action plan library when an action-level early warning is triggered. The instruction is then sent to the corresponding execution terminal through the communication interface.

6. The water body algal bloom dynamic monitoring and early warning system based on data analysis according to claim 5, characterized in that, The construction and updating process of the domain knowledge graph in the semantic association submodule is as follows: During system initialization, the initial association rules and weights defined by experts are loaded; During system operation, the statistical correlation between different parameter sequences in historical data is continuously analyzed through association rule mining algorithms; When a strong relationship is discovered that is inconsistent with the initial rules or is a newly added relationship, a manual review process is triggered. Once confirmed, the strong association and its confidence level are added to the knowledge graph and used to update the weighted adjacency matrix of semantic associations.

7. The water body algal bloom dynamic monitoring and early warning system based on data analysis according to claim 6, characterized in that, The adaptive filtering algorithm in the time series analysis submodule adopts a combination of empirical mode decomposition and ensemble empirical mode decomposition. The specific process is as follows: First, empirical mode decomposition is performed on the univariate time series of each grid cell to obtain a series of intrinsic mode functions; To address the potential mode aliasing problem in the decomposition results, ensemble empirical mode decomposition is introduced. By adding paired white noise and performing multiple decompositions and averaging, more stable intrinsic mode function components are obtained, thereby more accurately separating signal components at different time scales.

8. The water body algal bloom dynamic monitoring and early warning system based on data analysis according to claim 7, characterized in that, The state transition equation is specifically a discretized differential equation, which describes the rate of change of algal biomass over time as a function of current biomass, water temperature, light intensity, and nutrient concentration. The specific parameterized form of the function is represented by a deep neural network. The input to the deep neural network is the current hidden state and environmental factors, and the output is the parameters of the state transition. It is trained end-to-end using the negative log-likelihood loss function between historical observation data and model prediction results.

9. The water body algal bloom dynamic monitoring and early warning system based on data analysis according to claim 8, characterized in that, The decision logic submodule uses a dynamic adjustment mechanism for setting multi-level probability thresholds; this dynamic adjustment mechanism is optimized based on the accuracy and false negative rate of historical warnings. The specific process is as follows: the system periodically compiles statistics on the actual occurrence of algal blooms after the issuance of warnings at all levels within the past period, and calculates the accuracy and recall rate of the warning at that level. The probability threshold for the corresponding level is dynamically adjusted based on the weighted harmonic mean of precision and recall. If the precision is too low, the threshold is raised to reduce false positives, and if the recall is too low, the threshold is lowered to reduce false negatives.

10. The water body algal bloom dynamic monitoring and early warning system based on data analysis according to claim 9, characterized in that, The decision logic submodule performs hierarchical decision-making based on the comparison results as follows: When the risk probability exceeds the Level 1 threshold, a warning at the level of concern is generated, and the system is instructed to increase the data collection frequency. When the risk probability exceeds the Level 2 threshold, a warning level alert is generated, and the spatial coordinate range of the suspected algal bloom core area is automatically generated. When the risk probability exceeds the Level 3 threshold, an action-level warning is generated.

Images