A method and system for simulating and predicting spatiotemporal dynamics of lake cyanobacterial blooms

Abstract

This invention discloses a method and system for spatiotemporal dynamic simulation and prediction of cyanobacterial blooms in lakes, comprising: acquiring multi-source environmental monitoring data containing meteorological, hydrodynamic, and water quality information of the target water area; performing simulation analysis using a pre-set ecological dynamics model to obtain a physical baseline chlorophyll concentration field; resampling and aligning the multi-source environmental monitoring data and the baseline concentration field to construct a multi-channel spatial feature tensor; inputting it into a pre-trained spatiotemporal prediction model to extract spatial distribution deviation features to obtain a residual prediction field; spatially superimposing and correcting the physical baseline field and the residual prediction field to obtain a composite chlorophyll concentration prediction field; and performing threshold screening and connected region aggregation processing to extract spatial distribution data of cyanobacterial bloom patches. This invention reduces the fitting difficulty of purely data-driven methods, compensates for the systematic bias of mechanistic models, and improves the accuracy and physical rationality of spatiotemporal evolution prediction of blooms.

Description

Technical Field

[0001] This invention relates to the field of water environment monitoring and numerical simulation prediction technology, and in particular to a method and system for spatiotemporal dynamic simulation prediction of cyanobacterial blooms in lakes. Background Technology

[0002] The stability of lake aquatic ecosystems depends on the accurate characterization of the spatiotemporal evolution of phytoplankton. Conducting high-frequency dynamic simulations and predictions of algal blooms can provide fundamental quantitative data support for predicting eutrophication trends, preventing water quality deterioration, and the preventative scheduling of water conservancy projects. This forms the technical foundation for ensuring the safe allocation of water resources and the health of the aquatic environment in a watershed.

[0003] Currently, the main technical means for monitoring and predicting algal blooms rely on satellite remote sensing inversion and ecological dynamics mechanism model simulation. On the one hand, high-frequency observation satellites can acquire water surface spectral data and thus invert chlorophyll distribution. However, due to cloud cover, atmospheric aerosol scattering, and interference from changes in solar altitude angle, the acquired remote sensing images often suffer from spatiotemporal data gaps and decreased fidelity, limiting the ability for high-frequency continuous monitoring. On the other hand, traditional ecological dynamics models extrapolate the evolution of algal blooms by constructing mathematical expressions encompassing physical, chemical, and biological processes. These models generally suffer from significant parameter uncertainties and high computational complexity, easily leading to high computational costs and large spatial deviations in simulation results under long-term extrapolation calculations.

[0004] In summary, existing single-technology approaches struggle to balance the continuity of data acquisition with the accuracy of spatial distribution prediction when dealing with complex and ever-changing aquatic environments. Effectively overcoming data distortion interference in large-scale environmental observations and breaking through the error accumulation bottleneck inherent in long-sequence mechanism deductions, thereby improving the robustness and adaptability of aquatic ecological dynamic prediction under complex conditions, are pressing technical problems in this field. Therefore, it is necessary to investigate a new method that can improve the accuracy and stability of predictions for complex lake aquatic environments. Summary of the Invention

[0005] Purpose of the invention: To provide a method and system for spatiotemporal dynamic simulation and prediction of cyanobacterial blooms in lakes, in order to solve the above-mentioned problems existing in the prior art.

[0006] Technical solution: A method for spatiotemporal dynamic simulation and prediction of cyanobacterial blooms in lakes, comprising: Acquire multi-source environmental monitoring data for the target water area, including meteorological, hydrodynamic, and water quality information; The physical baseline chlorophyll concentration field was obtained by simulating and analyzing multi-source environmental monitoring data using a pre-defined ecological dynamics model. Spatial grid resampling and alignment were performed on multi-source environmental monitoring data and physical baseline chlorophyll concentration fields to construct a multi-channel spatial feature tensor. The spatial distribution deviation features are extracted by inputting the multi-channel spatial feature tensor into a pre-trained spatiotemporal prediction model to obtain the residual prediction field. The physical baseline chlorophyll concentration field and the residual prediction field are spatially superimposed and corrected to obtain the composite chlorophyll concentration prediction field. Threshold screening and connected region aggregation were performed based on the composite chlorophyll concentration prediction field to extract and output spatial distribution data of cyanobacterial bloom patches.

[0007] A spatiotemporal dynamic simulation and prediction system for lake cyanobacterial blooms that integrates ecological dynamics models and spatiotemporal prediction models includes: a processor and a memory communicatively connected to the processor. The memory stores instructions that the computer executes; The processor executes computer execution instructions stored in memory to implement the methods described in the above embodiments.

[0008] Beneficial effects: This invention reduces the difficulty of fitting purely data-driven models, compensates for the systematic bias of mechanistic models, and improves the accuracy and physical rationality of predictions of the spatiotemporal evolution of algal blooms. Attached Figure Description

[0009] Figure 1 This is a schematic flowchart of a spatiotemporal dynamic simulation and prediction method for cyanobacterial blooms in lakes according to the present invention.

[0010] Figure 2 A schematic diagram illustrating the overall technology for model implementation.

[0011] Figure 3 This is a schematic diagram illustrating the effect of treating cyanobacterial bloom patches.

[0012] Figure 4 A flowchart illustrating the steps involved in constructing a multichannel spatial feature tensor. Detailed Implementation

[0013] In the embodiments of this application, some concepts are explained as follows: Ecological dynamics models refer to mechanistic models used for hydrodynamic and water quality simulation, such as the EFDC environmental fluid dynamics model, MIKE, or Delft3D, which can be implemented using open-source or commercial models. Spatiotemporal prediction model uniformly refers to a hybrid prediction model that integrates the physical model baseline and the deep learning feature mapping end. In specific applications, the deep learning feature mapping end can be implemented using the CNN-Attention architecture. The physical baseline chlorophyll concentration field refers to the spatial distribution of chlorophyll a concentration output by the ecological dynamics model.

[0014] Example 1: This example provides a method for simulating and predicting the spatiotemporal dynamics of cyanobacterial blooms in lakes by integrating ecological dynamics models and spatiotemporal prediction models, including the following steps: like Figure 1 As shown, in step 101, multi-source environmental monitoring data containing meteorological, hydrodynamic and water quality information of the target water area are obtained; Furthermore, the water quality information in the multi-source environmental monitoring data is specifically acquired in real time through physical monitoring equipment deployed in the target water area. The physical monitoring equipment includes a fluid inlet, a sensor mounting section, and optical and capacitive sensors mounted on the sensor mounting section. The optical and capacitive sensors are hardware-encapsulated through waterproof connectors to ensure the airtightness of underwater data acquisition.

[0015] Optical sensors are used to measure water turbidity and chlorophyll fluorescence intensity in real time, while capacitive sensors are used to measure water conductivity and temperature in real time.

[0016] Water quality data collected by sensors is transmitted to shore-based data receiving terminals via wired or wireless communication. After format conversion, the data is input into subsequent ecodynamic and spatiotemporal prediction models as part of the multi-source environmental monitoring data.

[0017] In addition to real-time water quality information, multi-source environmental monitoring data also includes regional digital elevation models, daily hydrological data, and hourly meteorological data such as rainfall, temperature, and wind speed.

[0018] In addition, high temporal resolution satellite remote sensing image data is required as the basis for inversion.

[0019] In practical applications, satellite remote sensing data is often affected by cloud cover and changes in solar altitude angle, and the quantitative chlorophyll concentration retrieval using a specific index depends on specific physical conditions.

[0020] Specifically, the remote sensing inversion is applicable to high-concentration conditions where surface cyanobacterial blooms have already formed or are about to form, and is not applicable to low-concentration conditions where algae are evenly distributed in the water column.

[0021] Step 102: Use a preset ecological dynamics model to simulate and analyze the multi-source environmental monitoring data to obtain the physical baseline chlorophyll concentration field; Specifically, the ecological dynamics model solves for flow field and water quality variables by setting upper and lower boundary conditions for the target area and using standard three-dimensional continuity equations, momentum equations, and water quality convection-diffusion equations.

[0022] To ensure the reliability of the simulation analysis, parameter calibration of the ecological dynamics model is necessary. In practice, a step-by-step calibration strategy is adopted, that is, first calibrating hydrodynamic parameters such as the bottom roughness coefficient, whose values ​​can be set between 0.02 and 0.04, and then calibrating water quality parameters. The calibration target is set as Nash efficiency coefficient > 0.6 and relative error < 30%.

[0023] After rigorous calibration, the ecodynamic model can output a physical baseline chlorophyll concentration field that reflects the physicochemical evolution of the water body, thus providing a physically constrained reference benchmark for subsequent spatiotemporal prediction models.

[0024] Step 103: Spatial grid resampling and alignment of multi-source environmental monitoring data and physical baseline chlorophyll concentration field to construct multi-channel spatial feature tensors; Because acquired meteorological, hydrodynamic, water quality, and remote sensing data often have different time spans and spatial resolutions, direct fusion faces technical obstacles due to dimensionality mismatch. Spatial interpolation algorithms are used to uniformly transform these heterogeneous data onto a grid with the same spatial resolution. After resampling and alignment, all variables are concatenated and combined along the spatial channel dimension to form a multi-channel spatial feature tensor. This tensor serves as the spatial coordinate framework for the physical model output, the multi-source measured data, and the standardized spatial input format for subsequent spatiotemporal prediction models.

[0025] Step 104: Input the multi-channel spatial feature tensor into the pre-trained spatiotemporal prediction model to extract the spatial distribution deviation features and obtain the residual prediction field; Unlike traditional end-to-end direct mapping methods that directly predict absolute chlorophyll concentration, spatiotemporal prediction models include a physical model baseline and a deep learning feature mapping endpoint. This allows the learning objective to be transferred to the spatial distribution differences between the coarse estimates of the physical model and the actual observations.

[0026] By inputting multi-channel spatial feature tensors into the deep learning network structure, the model focuses on extracting systematic physical deviations in the ecological dynamics model caused by factors such as grid resolution or terrain simplification, and outputs a residual prediction field characterizing the distribution and magnitude of these deviations. Through the residual learning mechanism, the fitting difficulty of the deep neural network itself is reduced, resulting in stronger spatial generalization ability under unknown data distributions.

[0027] Step 105: Spatially superimpose and correct the physical baseline chlorophyll concentration field and the residual prediction field to obtain the composite chlorophyll concentration prediction field. In this stage, the physical baseline chlorophyll concentration field derived from the ecological dynamics model in the previous step is added grid-by-grid to the residual prediction field output by the spatiotemporal prediction model. By feeding back the data-driven bias compensation value to the baseline derived from the physical mechanism, a deep integration of physical laws and measured data is achieved. This ensures that the composite chlorophyll concentration prediction field not only follows the macroscopic physical evolution laws of water quality dynamics but also corrects for numerical fitting distortions in local space.

[0028] Step 106: Based on the composite chlorophyll concentration prediction field, threshold screening and connected region aggregation are performed to extract and output the spatial distribution data of cyanobacterial bloom patches.

[0029] After obtaining the composite chlorophyll concentration prediction field, the preset chlorophyll concentration threshold is used to convert it into a binary spatial mask to accurately distinguish the algal bloom aggregation area from the normal water background.

[0030] Morphological noise reduction and boundary repair were performed on the identified independent algal bloom connectivity areas to filter out small isolated patches caused by observation noise or accidental water disturbance, and the edges of the main patches that were retained were smoothed.

[0031] The output spatial distribution data of cyanobacterial bloom patches can intuitively present the objective spatial distribution pattern of the bloom, providing a quantitative and structured decision-making basis for subsequent watershed environmental management and flood control and disaster relief scheduling.

[0032] Example 2: An explanation of the construction of multi-channel spatial feature tensors and the pre-construction spatial feature engineering process.

[0033] In one possible implementation, the following steps are included: like Figure 4 As shown, in step 201, multi-source environmental monitoring data with different time scales and spatial resolutions are resampled to a unified spatial resolution grid using a spatial interpolation algorithm; the physical baseline chlorophyll concentration field is aligned to a unified spatial resolution grid; and the resampled multi-source environmental monitoring data and the aligned physical baseline chlorophyll concentration field are spliced ​​together in the spatial channel dimension to obtain a multi-channel spatial feature tensor.

[0034] Multi-source environmental monitoring data includes meteorological parameters, hydrodynamic parameters, and water quality parameters obtained from external reanalysis meteorological datasets. The meteorological parameters include precipitation, evaporation, wind speed, radiation, temperature, and humidity.

[0035] The hydrodynamic and water quality parameters are output hourly by the aforementioned ecological dynamics model. The hydrodynamic parameters include water level and flow velocity, while the water quality parameters include dissolved oxygen, phosphate, total phosphorus, ammonia nitrogen, nitrate nitrogen, chemical oxygen demand, and biochemical oxygen demand.

[0036] Since the meteorological input grid, physical model grid, and remote sensing grid have different sizes and update frequencies, this embodiment employs an inverse distance weighted interpolation algorithm to eliminate distortion caused by spatial inelastic overlap. This algorithm projects and resamples all input variables from different sources onto a standardized spatial resolution grid, with a grid resolution set to 1 kilometer. The aligned basic independent variables form 15 spatial grid channels.

[0037] The physical baseline chlorophyll concentration field derived from the ecological dynamics model in the aforementioned steps is also aligned to this 1-kilometer resolution grid using the same coordinate framework. By performing a splicing operator on the channel dimension of the spatial grid, the 15 resampled basic independent variable channels are merged with the physical baseline chlorophyll concentration field channels to construct a unified multi-channel spatial feature tensor, establishing a consistent multi-dimensional feature input base for subsequent spatiotemporal prediction models.

[0038] Step 202: Extract the flow velocity field and wind speed field from the multi-source environmental monitoring data, and calculate the effective drift velocity field of the target water area based on the preset wind-induced surface drift coefficient. Based on the effective drift velocity field, a high-order numerical integration method is used to perform backward trajectory tracking on the spatial grid to determine the location of the drift source region and the corresponding backward tracking trajectory within a preset time window; A probability map of the drift source region is generated based on the kernel density estimation of the drift source region location, and a cumulative drift distance map is generated based on the backward tracking trajectory. The probability map of the drift source region and the cumulative drift distance map are used as additional spatial channels. They are spliced ​​and integrated with the resampled multi-source environmental monitoring data and the physical baseline chlorophyll concentration field in the spatial channel dimension to obtain a multi-channel spatial feature tensor.

[0039] The resampled multi-source environmental monitoring data is spliced ​​with the physical baseline chlorophyll concentration field in the spatial channel dimension to obtain a multi-channel spatial feature tensor. It also includes splicing and integrating the drift source region probability map and the cumulative drift distance map as additional spatial channels.

[0040] In step 202, the spatial dynamic characterization capability of the feature engineering is enhanced by introducing fluid dynamic transport constraints. Lake algal blooms exhibit significant surface drift and migration characteristics under the influence of wind and water flow.

[0041] To quantify this spatiotemporal transport process, the surface velocity field output from the ecodynamic model and the 10-meter wind speed field from meteorological data were extracted from multi-source environmental monitoring data. Using a preset wind-induced surface drift coefficient, the wind speed field and the velocity field were linearly converted and superimposed to construct an effective drift velocity field characterizing the actual motion state of the surface water.

[0042] The wind-induced surface drift coefficient represents the empirical proportion of wind energy converted into surface velocity, and its value range is limited to 0.01≤α≤0.03.

[0043] For example, the transverse component U of the effective drift velocity field drift =u s +α*u w ;where u s Let u be the transverse component of the surface velocity field, α be the preset wind-induced surface drift coefficient, and u be the transverse component of the surface velocity field. w This represents the lateral component of the 10-meter wind speed field.

[0044] Longitudinal component V of the effective drift velocity field drift =v s +α*v w ; where v s Let v be the longitudinal component of the surface velocity field, α be the preset wind-induced surface drift coefficient, and v w This represents the longitudinal component of the 10-meter wind speed field.

[0045] Based on the constructed effective drift velocity field, backward particle trajectory tracking is performed for each lake spatial grid point (x0, y0) at the target prediction time t. The tracking calculation is carried out within a preset time window, which is set to 6 hours to match the spatial transport scale of cyanobacterial patches at typical wind speeds, ranging from 3 km to 10 km.

[0046] The numerical integration adopts the fourth-order Runge-Kutta method, with a time step set to 0.5 hours to ensure the convergence and high accuracy of trajectory inverse extrapolation.

[0047] Tracking determined lateral source region coordinates x src =x0-integral(U drift ,td T,t x0 is the initial horizontal coordinate of the target grid point, integral is the definite integral operator, and U... drift Let d be the transverse component of the effective drift velocity field, t be the target prediction time, and d be the lateral component. T This is a preset time window.

[0048] Tracking determined longitudinal source region coordinates y src =y0-integral(V drift ,td T,t ); y0 is the initial longitudinal coordinate of the target grid point, V drift This represents the longitudinal component of the effective drift velocity field.

[0049] After calculating the backward trajectory source region locations of all grid points in the lake, the distribution of source region locations is spatially smoothed using the kernel density estimation function to generate a drift source region probability map that represents the contribution probability of each upstream grid point to the algal bloom at the current target point.

[0050] Simultaneously, by performing a definite integral over the entire time period on the instantaneous velocity vector of the moving particle along the tracked backward motion trajectory, the cumulative drift distance map of each grid area is calculated.

[0051] The cumulative drift distance D of each grid point drift =integral(sqrt(U drift 2 +V drift 2 ),td T,t );U drift For the transverse component of the effective drift velocity field, V drift The longitudinal component of the effective drift velocity field, t is the target prediction time, and d is the longitudinal component of the effective drift velocity field. T This is a preset time window.

[0052] The newly generated drift source region probability map and cumulative drift distance map are used as two new sets of additional feature channels, which are then fused and spliced ​​into the aforementioned multi-channel spatial feature tensor. This achieves the explicit encoding of the macroscopic mechanism evolution path into a spatial input tensor form that is easy for deep learning networks to parse.

[0053] Step 203: Perform Latin hypercube sampling on the core ecological parameters of the preset ecological dynamics model to generate multiple sets of parameters; By using multiple sets of parameters and multi-source environmental monitoring data to drive ecological dynamics models to extrapolate environmental fields, multiple sets of spatial concentration simulation fields are obtained. Based on the set discretization degree of multiple spatial concentration simulation fields, a normalized spatial confidence field is constructed; The resampled multi-source environmental monitoring data and the physical baseline chlorophyll concentration field are spliced ​​together in the spatial channel dimension to obtain a multi-channel spatial feature tensor. It also includes: splicing and integrating the spatial confidence field as an additional spatial channel to characterize the simulation confidence of each grid location.

[0054] In step 203, spatial quantitative measurement is performed on the uncertainty of the physical mechanism model to guide the dynamic adjustment of the weights of the subsequent deep network.

[0055] The ecological dynamics model contains many core ecological parameters that are highly sensitive to time and space, such as the maximum algal growth rate, phosphorus half-saturation constant, nitrogen half-saturation constant, algal mortality rate, and bottom roughness coefficient.

[0056] To quantitatively assess the spatial output fluctuations caused by the aforementioned mechanism parameters, a Latin hypercube sampling method can be employed to perform multidimensional uniform sampling within the compliant physical boundaries of the parameters, thereby generating 20 distinct parameter sets. These 20 parameter sets are then used to drive a parallel ecological dynamics model for long-term environmental field simulations. The simulation time span is selected from June to September, the peak period for cyanobacterial blooms, resulting in 20 simulated chlorophyll concentration fields with spatial distribution differences.

[0057] At each specific spatial grid location (x, y), the above 20 sets of simulated concentrations are statistically analyzed, and their ensemble mean is calculated as a rough estimate of the physical baseline. The ensemble standard deviation is also calculated to characterize the uncertainty of that spatial location caused by parameter perturbation.

[0058] In order to transform this uncertainty into a feature map that can be directly read by a neural network, this embodiment introduces a reference uncertainty threshold as the denominator and normalizes the set standard deviation through a negative exponential mapping function to construct a spatial confidence field.

[0059] The reference uncertainty threshold was selected as the median of the ensemble standard deviations within the entire lake spatial domain.

[0060] Spatial confidence field C EFDC =exp(-σ EFDC / σ ref ); where exp is the natural exponential function, σ EFDC σ represents the ensemble standard deviation of multiple sets of spatial concentration simulation fields. ref This serves as a reference uncertainty threshold.

[0061] The value of the spatial confidence field constructed in this way is strictly limited to 0. <C EFDC Within the interval ≤1.

[0062] When the confidence value approaches 1, it indicates that the simulation output of the spatial grid location remains highly consistent under different parameter perturbations, and the calculation results of the physical mechanism model are highly reliable. When the confidence level approaches 0, it indicates that the outputs under different parameter combinations are seriously divergent, and the physical model has a weak simulation capability in this case.

[0063] To reduce the computational load during real-time prediction, the above parameter set perturbation evolution process is completed offline. The calculated spatial confidence field is then averaged over time and used as a quasi-static spatial prior template.

[0064] By incorporating the spatial confidence field as an additional feature channel into the multi-channel spatial feature tensor, the computational confidence of the physical model at each spatial grid location of the deep prediction network is explicitly communicated, thereby guiding the network's attention allocation mechanism to make targeted adjustments.

[0065] In some alternative implementations, for gridded resampling alignment of multi-source data, Kriging interpolation or bilinear interpolation can be used instead of inverse distance weighted interpolation to accommodate different lake boundary shoreline shapes.

[0066] In addition, alternative strategies such as Monte Carlo random sampling or orthogonal array sampling can be used to sample uncertainties in core ecological parameters to suit different hardware computing power configurations.

[0067] In another parallel optional design, the backward particle trajectory tracking stage can be further integrated with the buoyancy change formula of cyanobacteria cells, upgrading the trajectory simulation of the two-dimensional plane to the three-dimensional flow field tracking of water column transport.

[0068] After the complete feature engineering process described in steps 201 to 203, the final constructed multi-channel spatial feature tensor contains the sum of the number of channels for basic independent variables, physical baselines, drift features, and confidence. In some embodiments, this is specifically 15+1+2+1=19 spatial channels.

[0069] Example 3: Further explanation of the network mechanism, residual correction architecture, and autoregressive time series mechanism of the spatiotemporal prediction model.

[0070] In one possible implementation, the following steps are included: Step 301: The pre-trained spatiotemporal prediction model includes a physical model baseline and a deep learning feature mapping end; the physical baseline chlorophyll concentration field is provided by the physical model baseline, and the process of extracting spatial distribution deviation features is performed by the deep learning feature mapping end.

[0071] In this embodiment, the prediction architecture is functionally decoupled. The physical model baseline independently processes macroscopic fluid dynamics and matter conservation equations to generate coarse estimates with physical meaning.

[0072] Deep learning feature mapping eliminates the burden of learning physical conservation laws and focuses on extracting features specifically for spatial distribution deviations caused by parameter simplification and grid resolution limitations in physical models.

[0073] This architecture establishes a collaborative extraction mechanism that uses physical mechanisms to build the operational foundation and data-driven models to perform error compensation.

[0074] Step 302: Input the multi-channel spatial feature tensor into the pre-trained spatiotemporal prediction model to extract spatial distribution bias features, and obtain the residual prediction field, including: By utilizing the encoder and attention mechanism modules in the deep learning feature mapping end, feature extraction is performed on the multi-channel spatial feature tensor, and the residual prediction field representing the deviation of the physical model is output. Specifically, the deep learning feature mapping terminal receives a three-dimensional tensor input, which consists of a preset number of channel dimensions, grid rows, and grid columns.

[0075] The encoder consists of three sets of alternating convolutional and pooling layers. The first set of convolutional layers has 64 kernels of 3x3 dimensions, followed by stride-limited pooling operations to extract shallow environmental features. The second set of convolutional layers has 128 kernels, and the third set has 256 kernels for extracting high-level semantic features. Each convolutional layer is coupled with batch normalization and an activation function. An attention mechanism module is connected to the encoder's output.

[0076] This module calculates the contribution weight of each channel to physical deviation correction. The calculation logic for channel squeezing and feature recalibration is as follows: z c =(1 / (H*W))*sum(F c (i,j)); Among them, z c Let H be the scalar output of the c-th channel after global average pooling, H be the spatial row dimension of the input feature map, W be the spatial column dimension of the input feature map, and F be the scalar output of the c-th channel after global average pooling. c (i,j) represents the characteristic response value of the c-th channel at the planar coordinates (i,j), and sum is the summation operation for all spatial coordinate positions.

[0077] s c =Sigmoid(W2*ReLU(W1*z c )); Among them, s c W1 is the dynamic weight for attention activation of the c-th channel output. W1 is the dimensionality reduction weight matrix of the first fully connected network, where the dimensionality reduction ratio can be determined by those skilled in the art based on the number of input channels and computational efficiency requirements, typically taking a value of 16. W2 is the dimensionality increase weight matrix of the second fully connected network, where ReLU is the linear rectified activation function and Sigmoid is the normalized activation function that maps the output probability to the interval between 0 and 1.

[0078] F_ Tilde _ c =s c *F c ; Among them, F_ Tilde _ c This is the output feature map after attention channel enhancement.

[0079] The high-level feature map, after being enhanced by the attention channel weights, is then input into the decoder. It is then upsampled through the corresponding transposed convolutional block to gradually expand to the same spatial resolution dimension as the original lake grid. Finally, the residual prediction field is calculated by the single-channel output layer.

[0080] Step 303: Spatially superimpose and correct the physical baseline chlorophyll concentration field with the residual prediction field to obtain the composite chlorophyll concentration prediction field, including: The physical baseline chlorophyll concentration field provided by the physical model baseline end is added grid-by-grid to the residual prediction field output by the deep learning feature mapping end to obtain the composite chlorophyll concentration prediction field.

[0081] After extracting the residual prediction field, a matrix addition operation is performed on a spatial grid-by-grid basis to compensate the residual values ​​representing systematic physical deviations back into the physical baseline chlorophyll concentration field according to the corresponding coordinates.

[0082] In some lake area grids, the calculated residual value is negative, indicating that the physical model overestimates the concentration in this area. Performing the addition operation is equivalent to down-correction. In other grid regions, the calculated residuals are positive, and performing an addition operation is equivalent to up-compensation. The resulting composite chlorophyll concentration prediction field integrates the evolution profile of the physical baseline with the detailed deviations extracted by the network.

[0083] In an alternative implementation, for specific small-scale water area prediction tasks, the spatiotemporal prediction model can omit the architectural decoupling and physical residual extraction processes. This alternative directly inputs the constructed multidimensional spatial feature tensor into the encoder and decoder, and through end-to-end deep network computation, maps the input environmental parameters directly to the absolute chlorophyll concentration, directly outputting the final prediction field.

[0084] Step 304: Obtain the composite chlorophyll concentration prediction field at the previous prediction time, and calculate the difference between the composite chlorophyll concentration prediction field at the previous prediction time and the composite chlorophyll concentration prediction field at the even earlier prediction time to obtain the corresponding time change rate field. The composite chlorophyll concentration prediction field and the time change rate field of the previous prediction time are integrated as augmented input channels into the multi-channel spatial feature tensor of the current time to construct a time-series input feature with autoregressive feedback. For scenarios requiring continuous projection of algal bloom evolution over multiple time steps, static input at a single moment cannot capture the temporal continuity of algal reproduction. In this embodiment, the model uses the composite chlorophyll concentration prediction field output at time step T-1 as a feedback buffer.

[0085] The difference grid matrix between the prediction results at time T-1 and time T-2 is calculated to obtain the time rate of change field characterizing the local algal bloom propagation or decline trend.

[0086] At the start time of the continuous time period prediction, the chlorophyll concentration field of the most recent valid remote sensing observation is used as the initial input, and the initial time rate of change field is set to a zero matrix.

[0087] Starting from the second prediction time step, the model prediction output from the previous step is used as the feedback input.

[0088] The cached prediction field and the time rate of change field are treated as two new augmented channels and concatenated along the channel dimension into the multi-channel spatial feature tensor at time step T.

[0089] Based on the aforementioned autoregressive input mechanism, the network continuously acquires prior evolutionary inertia information when performing subsequent step-size extrapolation.

[0090] Step 305: Based on the preset maximum specific growth rate of algae and the current prediction time step, calculate the maximum allowable physical change boundary after the evolution of the current prediction time step. Using the predicted field of complex chlorophyll concentration at the previous prediction time as a benchmark, the predicted field of complex chlorophyll concentration at the current time is subjected to extreme value clipping using the maximum allowable physical change boundary, cutting off the predicted values ​​that exceed the physical growth limit, and thus obtaining the predicted field of complex chlorophyll concentration corrected by extreme value clipping.

[0091] In autoregressive closed-loop prediction iterations, the fitting perturbation of the neural network itself may cause the output value to diverge after multiple steps of cumulative calculation.

[0092] To establish a divergence prevention constraint mechanism, this embodiment applies a biophysical limit truncation to the output results.

[0093] A pre-configured maximum specific growth rate constant for algae in the target water area is used. Based on this constant and the concentration benchmark of the preceding buffer, a theoretically inviolable upper limit and lower limit range for phytoplankton density growth and decline is calculated.

[0094] If the model output matrix contains grid extrema falling outside the allowable interval, these extrema are forcibly overwritten as boundary values ​​of that interval. The control calculation logic for extrema clipping is as follows: Chl clip =clip(Chl pred Chl prev *exp(-μ max *dt),Chl prev *exp(μ max *dt)); Among them, Chl clipThis is the predicted composite chlorophyll concentration field output after extreme value physics clipping, where clip is the boundary limiting truncation operator function, and Chl is the output. pred Chl is the predicted field of composite chlorophyll concentration at the current moment calculated by a deep network. prev The baseline for the composite chlorophyll concentration prediction field at the previous prediction time is given by exp, which is the natural exponential function, and μ. max The pre-configured maximum specific growth rate of algae, dt is the current estimated single prediction time step, and μ max The same time unit should be used as dt to ensure that the exponential term μ max * dt is a dimensionless quantity. In this embodiment... ， μ max Set to 0.075 h -1 dt is set to 1 hour.

[0095] By using the above pruning rules, numerical spikes or precipitous drops caused by out-of-bounds model data space are eliminated, ensuring that the rate of change accumulated at each time step is subject to the mandatory envelope constraint of ecological physical laws.

[0096] By combining a dual-core driven architecture that integrates physical evolution baseline with deep learning residual correction, and an autoregressive time-series extreme value truncation mechanism based on the maximum specific growth rate of algae, the risk of dimensional divergence when the network directly fits the absolute concentration is eliminated. This also reduces the risk of sharp increases or precipitous drops in values ​​that violate biological common sense, which are prone to occur in continuous time step extrapolation, thus ensuring the physical objectivity and continuity of long-term prediction results.

[0097] Example 4: Further explanation of the offline training process and multi-objective loss function construction mechanism of the pre-trained spatiotemporal prediction model. In one possible implementation, the following steps are included: Step 401: Obtain multi-source environmental monitoring data for historical periods, the corresponding historical physical baseline chlorophyll concentration field, and the historical remote sensing inversion chlorophyll concentration field as the true label, and construct training sample pairs. Before performing offline training, the collected historical test dataset needs to be time-series divided and cleaned. The acquired historical data is divided into two parts according to the chronological order of occurrence. The data from the first 70% of the time period is used to build the training set, and the data from the last 30% of the time period is reserved as the validation set to avoid data leakage to the training stage in the future.

[0098] For remote sensing inversion images used as true labels, an effectiveness filtering operation is performed to retain only observation images with a spatial cloud coverage ratio of less than 30%, and to remove invalid samples with a large number of missing labels due to large-area cloud obstruction.

[0099] After the above quality control and time series division, the historical multi-source variable tensor and physical baseline at the corresponding time are used as the model input, and high-frequency remote sensing image data are used as the output label to form a normalized training sample pair for forward propagation and backward differentiation.

[0100] Step 402: Input the training sample pairs into the spatiotemporal prediction model to be trained and calculate the prediction output results; To improve the convergence stability of the model to be trained in the time series prediction task, a hybrid driving strategy combining teacher-mandated and free-running approaches is adopted in the input process.

[0101] Within the first 100 training iterations, a probability factor of 0.7 is configured to use real remote sensing chlorophyll image data to replace the model's predicted output from the previous time step as the input for the current time step; at the same time, a probability factor of 0.3 is configured to directly use the model's predicted output from the previous iteration step.

[0102] When the training iteration cycle exceeds 100, the probability factor using the real label as feedback input will be linearly decayed to 0, forcing the model to adapt to the closed-loop operating environment of its own accumulated error, thereby outputting a prediction output that conforms to the characteristics of time-series dynamic evolution.

[0103] Step 403: Based on the temperature, light and nutrient concentration data extracted from the training sample pairs, calculate the degree of limitation on the preset maximum algal growth rate, and then determine the theoretical net algal growth rate at the corresponding spatial grid position of the network. To incorporate biodynamic evolution principles into the neural network, this embodiment explicitly calculates the real-time proliferation potential of algae using input parameters such as temperature, light intensity, and nutrient concentrations (nitrogen and phosphorus). The network assesses the actual proliferation capacity under current physical conditions by calculating the nonlinear constraint coefficients of each environmental factor on the optimal growth state of the algae. The calculation relationships between temperature constraints, light constraints, nutrient constraints, and the final theoretical net algal growth rate are as follows: Temperature limiting factor f T =exp(-k T *(T env -T opt ) 2 ); where k T To set a preset temperature-induced attenuation factor while maintaining dimensional consistency, T env T is the real-time ambient temperature extracted from the spatial grid. opt This is the baseline value for the optimal growth temperature.

[0104] Light limiting factor f I =I rad / (I rad +K I ); where I radTo extract the photosynthetically active radiation intensity, K I This is the preset light half-saturation constant.

[0105] Combined limiting factor f of nutrients N =min(P con / (P con +K p ),N in / (N in +K N )); where min is the minimum value operation, P con To determine the effective phosphate concentration extracted, K p The preset phosphorus half-saturation constant, N in The concentration of extracted inorganic nitrogen, K N This is the preset nitrogen half-saturation constant.

[0106] Theoretical net growth rate of algae μ Net =μ max *f T *f I *f N -μ d -μ s ; Where, μ max The maximum algal growth rate under unconstrained ideal conditions, μ d μ is the preset death respiratory loss rate. s This is the preset physical settlement rate.

[0107] Step 404: Extract the predicted output results of adjacent observation times, calculate the time difference rate of change, and compare the time difference rate of change with the theoretical physical rate of change derived from the theoretical net algal growth rate to obtain the physical deviation penalty value. The concentration variation of the network output between two consecutive observation times is captured using time-difference operations. The difference between this variation obtained from network fitting and the physical increment calculated using the macroscopic theoretical algal net growth rate obtained in the previous step is then squared to calculate the penalty value for violating the biological growth theorem.

[0108] Physical deviation penalty value L phys =(1 / M)* ∑(((Chl pred2 -Chl pred1 ) / dt-μ net *Chl pred1 ) 2 ); Where M is the total number of effective spatial grids, ∑ is the spatial summation calculation, and Chl pred2 Chl is the predicted chlorophyll concentration for the next observation time. pred1This represents the predicted chlorophyll concentration at the previous observation time, where dt is the time interval between two consecutive observations. The time unit of dt should be consistent with the theoretical net algal growth rate μ. net The time unit used is kept consistent, μ net This represents the calculated theoretical net algal growth rate.

[0109] It should be noted that in the above formula, dt and μ net A consistent unit of time measurement should be used; when the system operates with hours as the basic time step, μ net μ in max μ d and μ s All should be converted to hours. -1 Substitute the unit into the calculation.

[0110] Step 405: Dynamically adjust the weight coefficients of the physical deviation penalty value in the multi-objective loss function according to the preset cosine annealing strategy to obtain a dynamically weighted physical constraint loss term. This physical constraint loss term is then incorporated into the multi-objective loss function to participate in the iterative update of the network weights. This ensures that the spatiotemporal prediction model to be trained is subject to strong physical constraints in the early stages of training.

[0111] To balance the gradient conflict between physical constraints and data fitting, this step guides the evolution path of network parameters by dynamically adjusting the weight coefficients. Initially, higher weight coefficients are configured to force the network parameters to optimize within a physically reasonable solution space. Subsequently, the physical intervention is gradually relaxed according to the decay curve of the cosine function, allowing the network to freely capture the microscopic differences in the data samples. The cosine annealing adjustment mechanism of the weight coefficients is as follows: λ _epoch =λ0*0.5*(1+cos(π*epoch) curr / epoch _max )); Where, λ _epoch λ0 represents the dynamic weight coefficients for the current training iteration cycle, λ0 is the preset initial maximum weight constant, cos is the cosine trigonometric function, and epoch represents the number of iterations. curr This is the index for the current training iteration epoch. _max This is the maximum total number of iterations configured.

[0112] Step 406: For each spatial grid in the historical remote sensing inversion chlorophyll concentration field, based on the cloud occlusion marker data corresponding to the historical remote sensing inversion chlorophyll concentration field, calculate the distance to the nearest cloud occlusion grid to determine the cloud proximity attenuation weight; based on the pre-stored target water area land boundary data, calculate the distance to the nearest land boundary grid to determine the water-land mixing transition weight; and based on the image acquisition time and grid geographic coordinates, calculate the solar zenith angle corresponding to the grid to determine the atmospheric path quality weight. For cloud edge regions, the inversion data often exhibits a positive shift due to the strong scattering effect of aerosols; for grids near the shoreline, pixel fidelity is compromised due to the cross-mixing of spectral signals from the water surface and land; in addition, an excessively large incident angle of sunlight can increase the transmission path of light in the atmosphere and lead to a decrease in the signal-to-noise ratio.

[0113] To offset the negative impact of these physical imaging interferences on model training, the absolute spatial distances from each grid to the cloud edges and land boundaries were extracted, and the celestial coordinate parameters at the imaging time were obtained to calculate three independent mass characterization coefficients. The calculation method is as follows: w cloud =1-exp(-d cloud / d c ); Among them, w cloud The cloud proximity decay weights are used for the spatial grid, exp is the natural exponential function, and d cloud d is the absolute planar distance from this grid to the nearest cloud-occluded grid. c This is the preset spatial attenuation scale constant for cloud radiation interference.

[0114] w land =min(1,d land / d l ); Among them, w land The water-land mixing transition weights are used for the spatial grid, min is the minimum value operation, and d land d represents the shortest distance from this grid to the lake-land boundary. l The preset spatial scale parameters for spectral mixing interference transition.

[0115] w sun =cos(θ s ); Among them, w sun Here, θ represents the atmospheric path quality weights for the spatial grid, cos is the cosine trigonometric function, and θ is... s The solar zenith angle corresponding to the target water area at the current image acquisition time.

[0116] Step 407: Perform a grid-by-grid product operation on the cloud proximity attenuation weight, the water-land mixing transition weight, and the atmospheric path quality weight to obtain the pixel-level data quality weight of a single spatial grid. After extracting the penalty weights corresponding to the three environmental obstacle factors, a comprehensive evaluation matrix covering all grid regions of the entire lake was calculated through a scalar product operation. The resulting pixel-level data quality weights quantify the label credibility of each grid.

[0117] Step 408: The prediction residuals corresponding to each grid are weighted and penalized using pixel-level data quality weights. When the ratio of the prediction residuals to the physical baseline chlorophyll concentration field exceeds a preset allowable ratio coefficient, a regularization constraint is applied to obtain the pixel-level remote sensing quality weighted residual loss term, which is then incorporated into the multi-objective loss function to participate in the iterative update of the network weights.

[0118] Based on the pixel-level data quality weights calculated above, a residual fitting loss function with anti-interference correction is constructed. Simultaneously, to prevent the deep network from drastically modifying the underlying physical baseline to forcibly fit individual extreme outliers, a regularization boundary constraint is further introduced. When the proportion of the change in the model's predicted correction value relative to the original physical baseline exceeds the safe tolerance range, additional penalty pull is applied to control the magnitude of the modification. The mechanism of the regularization constraint is as follows: L _reg =(1 / M)*∑(ReLU(abs(R _pred )-β*Chl EFDC ) 2 ); Among them, L _reg Here, R represents the regularization constraint loss value, M is the total number of computational grids, ∑ represents spatial summation, ReLU is the linear rectified activation function, abs is the absolute value operation, and R... _pred The residual prediction value is calculated by the deep network, and β is a preset allowable modification ratio coefficient, which is taken as 0.8 in this embodiment. EFDC This is the baseline concentration value output from the physical model's baseline.

[0119] Step 409: Based on the difference between the predicted output and the true label, the network weights of the spatiotemporal prediction model to be trained are iteratively updated using a preset multi-objective loss function until the convergence condition is met, thus obtaining the pre-trained spatiotemporal prediction model.

[0120] By combining the residual fitting loss with spatial pixel weights, the regularization constraint loss representing the correction magnitude truncation, and the ecological mechanism constraint loss that introduces a physical penalty mechanism based on the time rate of change, a complete overall evaluation system for multi-objective loss functions is constructed.

[0121] An optimizer algorithm based on adaptive moment estimation is used to calculate the overall loss, and the error gradients of all convolutional kernels and fully connected layer parameters at the feature map end are calculated layer by layer through a backpropagation mechanism. Forward inference and backward parameter update loops are continuously executed until the overall evaluation loss matrix value stabilizes and no longer shows a decreasing trend. Upon reaching this convergence condition, the weight parameter sets of each layer's synapses are locked, and the current frozen network output is stored as a pre-trained spatiotemporal prediction model for real-time applications.

[0122] A multi-dimensional spatial pixel weighting mechanism encompassing cloud proximity attenuation, land-water mixing transition, and atmospheric path quality is introduced. By applying a dynamic gradient attenuation penalty to inferior observation labels, the misleading effect of edge spectral anomalies on deep network parameter updates is reduced, enhancing the prediction system's robustness under harsh optical imaging conditions.

[0123] Example 5: Based on the above examples, this example further illustrates the post-processing procedure for threshold screening and connected region aggregation based on the composite chlorophyll concentration prediction field. In one possible implementation, the following steps are included: Step 501: Binarize the composite chlorophyll concentration prediction field based on the preset ecological concentration threshold to generate an algal bloom patch mask; mark the connected regions of the algal bloom patch mask, count the number of pixels in each connected region, and remove isolated noise patches with a pixel count lower than the preset area threshold to obtain the filtered algal bloom patch mask. Morphological dilation was used to fill the internal cavities of the remaining patches in the screened cyanobacterial bloom mask and connect adjacent patches. Morphological erosion was used to remove the protruding areas at the edges of the patches, thus obtaining spatial distribution data of morphologically intact cyanobacterial bloom patches.

[0124] In step 501, the continuous concentration numerical field is converted into an ecological spatial characteristic region with physical boundaries. The system obtains a preset ecological concentration threshold, which is established based on the critical index for algal blooms in a specific lake.

[0125] In the Taihu Lake waters application scenario, the ecological concentration threshold is set at 20 micrograms per liter, which ecologically corresponds to the early warning limit for cyanobacterial blooms when the cyanobacterial cell density reaches 500,000 cells per milliliter. The calculated concentration value of each spatial grid in the composite chlorophyll concentration prediction field is compared with the ecological concentration threshold one by one. If the predicted concentration value of a spatial grid is ≥20 micrograms per liter, the state value of that grid point is assigned a value of 1, indicating that it belongs to the cyanobacterial bloom coverage grid; if the predicted concentration value of that grid point is <20 micrograms per liter, it is assigned a value of 0, indicating that it belongs to the normal water background. Through this binarization operator, the continuous field data is converted into a two-dimensional grid matrix composed of 0s and 1s, thereby generating a bloom patch mask.

[0126] Due to local fluctuations in remote sensing reflectance or edge perturbations in neural network output, the binarized matrix often contains sporadic, small, high-value noise points that lack macro-level ecological governance significance. To eliminate such spatial perturbations, this embodiment utilizes a two-dimensional bidirectional connected component search algorithm to identify spatially adjacent independent closed grid sets with all state values ​​of 1 within the algal bloom patch mask, thus marking connected regions. An 8-connectivity criterion is used during the search. For each marked independent closed grid set, the total number of pixel grids it contains is calculated. The system is configured with a preset area threshold to intercept small-scale random noise points. The preset area threshold is set to 9 grid pixels, corresponding to a physical coverage area of ​​9 square kilometers at a 1 km grid resolution. The total number of pixel grids in each independent closed grid set is compared with the preset area threshold. If the total number of pixel grids in a connected region is <9, the region is determined to be transient environmental noise or a fragmented small patch, and the state values ​​of all its internal grids are forcibly updated from 1 to 0, resulting in removal. Real algal bloom aggregation regions with a total number of pixel grids ≥9 are retained.

[0127] Morphological geometric regularization was performed on the preserved main interconnected regions. Due to water mixing and local network fitting errors, real cyanobacterial bloom patches may contain pseudo-pores with a state value of 0, or small discontinuous protrusions at the patch edges, which would reduce the accuracy of patch spatial boundary quantification.

[0128] This embodiment introduces a morphological operator composed of specific structural elements. Specifically, the structural elements are 3x3 square matrix structural elements.

[0129] Morphological dilation is performed on the filtered connected regions using structuring elements. This operation expands the outer boundaries of patches by applying local maximum filtering at the mesh boundaries, filling in tiny voids within the patches and closing and connecting spatially close adjacent patches.

[0130] Morphological erosion is performed on the expanded patches using the same structuring element. This morphological erosion operation, through local minimum filtering, causes the patch boundaries to shrink proportionally inward, eliminating small, spiky protrusions at the patch edges caused by noise, while maintaining the area balance of the patch's core region.

[0131] Through the combined processing of expansion and corrosion described above, the spatial geometric integrity and boundary smoothness of the connected algal bloom patches are ensured, and the spatial distribution data of the morphologically complete cyanobacterial bloom patches are finally output.

[0132] In some alternative implementations, the ecological concentration threshold can be dynamically and adaptively mapped based on the season or water temperature at the time of prediction. During periods of high water temperature in summer, the threshold can be adjusted to 15 micrograms per liter to enhance the sensitivity of the warning.

[0133] Similarly, the preset area threshold can be flexibly scaled according to the scale of the managed watershed. In addition, in the morphological regularization stage, the shape of the structural elements, in addition to squares, can also be replaced with 3×3 circular or cross-shaped structural elements to adapt to complex and meandering lake shoreline boundaries with higher curvature variations.

[0134] Compared with using an ecological dynamics model or a data-driven method alone, the technical solutions of the above embodiments have technical improvements in the following aspects: First, the residual learning framework enables deep learning networks to learn only the bias distribution of the physical model rather than the complete mapping of the absolute concentration field, which reduces the fitting complexity of the network and thus achieves better spatial generalization ability under limited training sample conditions. Second, the pixel-level remote sensing quality weighted loss mechanism weakens the misleading effect of low-quality observations near cloud edges and shorelines on network parameter updates, and enhances the anti-interference capability of the prediction system under harsh optical imaging conditions. Third, the autoregressive extreme value pruning mechanism based on the maximum specific growth rate of algae ensures that the continuous time step extrapolation results conform to the biological and physical limits and suppresses the numerical divergence phenomenon in multi-step iteration.

[0135] Example 6: In one possible implementation, the following steps are included: Step 601: Perform structured mapping on the spatial distribution data of cyanobacterial bloom patches, the physical baseline chlorophyll concentration field, and the extracted multi-source environmental monitoring data to obtain the structured mapping results; In this step, the underlying business structure parsing and physical spatiotemporal alignment of cross-modal data are performed.

[0136] The spatial distribution data of cyanobacterial bloom patches generated in the preceding steps are a two-dimensional image mask array, the physical baseline chlorophyll concentration field is a continuous floating-point tensor, and the multi-source environmental monitoring data includes one-dimensional time-series observation sequences and digital elevation parameters acquired by optical and capacitive sensors encapsulated by through-wall waterproof connectors. To eliminate the heterogeneity of different data carriers in storage and computation, the system constructs a unified spatiotemporal grid index mapping table.

[0137] For spatial distribution data of cyanobacterial bloom patches, the geographic coordinates of the vertices, the centroid coordinates of the patches, and the geometric attributes of the area of ​​their closed circumscribed polygons are extracted and encapsulated as vector feature entities. For hydrological flow velocity and water level in multi-source environmental monitoring data, their instantaneous scalar values ​​are extracted and associated with the corresponding watershed cross-section nodes, and encapsulated as hydrodynamic feature entities.

[0138] Through the above parsing and transformation operations, the originally independent pixel matrix and time-series signal are uniformly mapped into a relational data record table with standardized attribute fields, thereby completing the structured leap from algorithm tensor output to a general database storage format.

[0139] Step 602: Based on the structured mapping results, construct a multi-dimensional collaborative watershed intelligent knowledge base to provide comprehensive decision support for water environment anomaly early warning and ecological flood control and disaster relief.

[0140] After completing the structured mapping of the underlying data, a multi-dimensional collaborative method for constructing a smart knowledge base for watershed flood control is implemented through association analysis and rule extraction.

[0141] Extract information on physical hydraulic engineering facilities in the target water area, including the topological node coordinates of pumping stations, reservoir gates, and drinking water intakes.

[0142] Using a graph database engine, the physical facility nodes, hydrodynamic feature entities formed by the mapping of the preceding steps, and algal bloom patch vector feature entities are used as vertices in the graph structure, and the hydraulic connectivity and spatial distance between vertices are used as edges in the graph structure to construct a knowledge graph of water environment elements.

[0143] Based on this, the historical scheduling operation records are correlated with the corresponding physical baseline chlorophyll concentration field change data to generate a set of coupled control rules for hydrodynamics and ecological environment.

[0144] The constructed knowledge graph is integrated with the coupled control rule set to form a complete intelligent knowledge base for watershed flood control. This knowledge base is used in practical engineering scenarios to guide water conservancy scheduling decisions that take into account ecological security.

[0145] After obtaining real-time input parameters, a decision-making simulation is performed by querying the knowledge base. The simulation process is implemented as follows: When a flood control and discharge task instruction is received, the knowledge base engine extracts the current flow field change trend and retrieves the spatial distribution data of cyanobacterial bloom patches in the downstream waters.

[0146] To determine whether the surge in hydrodynamic flow velocity caused by the flood discharge will push existing algal bloom patches into key drinking water intake protection areas.

[0147] If a spatial conflict rule is found, the knowledge base engine will output a comprehensive decision support solution that includes instructions such as truncation of the flow limit or opening of alternative drainage channels.

[0148] By leveraging the aforementioned multi-dimensional information collaborative access mechanism, the risk of secondary water pollution spread is avoided in traditional flood control scheduling, which only considers reservoir safety and ignores the risk of secondary water pollution spread. This achieves closed-loop control of watershed flood control and ecological disaster relief.

[0149] In some alternative implementations, for the structured mapping step, spatial hashing algorithms can be used to replace the basic spatiotemporal grid index in order to improve the retrieval response speed for fluid sensing data of tens of billions or more.

[0150] In another alternative implementation, the underlying architecture of the intelligent knowledge base can be independent of graph databases. Instead, it can use a relational database with spatial query extension plugins to store entity locations and attributes, and solidify coupling control rules through an offline expert decision tree model to adapt to some edge-side physical flood control terminals with limited computing power.

[0151] In summary, the backdrift trajectory of the Lagrange wind field, characterizing fluid transport, and the spatial confidence field, quantifying the uncertainty of the physical model, were extracted beforehand. Furthermore, a partial differential penalty term for algal net growth rate, constrained by temperature, light, and nutrients, was deeply embedded in the bottom layer of the network training. This mechanism achieves a deep synergy between macroscopic hydrodynamic evolution priors and microscopic ecological baseline mechanisms, significantly improving the model's accuracy in spatial distribution extrapolation in areas lacking high-frequency effective observation labels and its robust generalization ability under unseen conditions.

[0152] Compared to using an ecological dynamics model or a data-driven model alone, this application shows a significant improvement in the accuracy and temporal continuity of chlorophyll concentration spatial distribution prediction.

[0153] Taking a certain lake as an example, this paper describes a method for spatiotemporal dynamic simulation and prediction of cyanobacterial blooms in lakes that integrates ecological dynamics models and CNNs, including the following steps: (1) Collect digital elevation models, hydrological and meteorological data, water quality data, and remote sensing imagery of the lake. Obtain hourly-scale spatial data of lake hydrodynamics and water quality by constructing a lake ecological dynamics model. Invert hourly-scale spatial data of lake chlorophyll a based on high temporal resolution remote sensing images. Specifically: (1-1) Basic Data Collection: Acquisition of lake topographic data, daily-scale hydrological data (flow and water level), hourly-scale meteorological data (rainfall, temperature, wind speed, etc.), monthly-scale water quality data, and hourly-scale high-frequency satellite remote sensing imagery. In this example, the high-frequency remote sensing imagery acquired is GOCI (Geostationary Ocean Color Imager) remote sensing data.

[0154] (1-2) Construction of Eco-hydrodynamic Model: Based on the collected basic data, the upper and lower boundaries of the region were set, and the open-source model EFDC was used to establish an ecological dynamic mechanism model to simulate and output hourly hydrodynamic and water quality parameters. The governing equations of the hydrodynamic and water quality models are existing technologies and will not be detailed here.

[0155] (1-3) Chlorophyll a satellite remote sensing inversion: The reflectance data of each spectral band of the high-frequency satellite remote sensing data are obtained, and then the AFAI (Alternative Floating Algae Index) index is calculated to reflect the algal biomass characteristics of the lake. The measured chlorophyll a concentration data is used to perform a quadratic polynomial correction on the AFAI index results, so as to obtain the chlorophyll a distribution results obtained by remote sensing inversion.

[0156] The formula for calculating the AFAI index is as follows: ; in , ,and The data are reflectance data for GOCI at 660nm, 745nm, and 865nm, respectively.

[0157] (2) The acquired independent variable data (meteorological, hydrodynamic, and water quality) and dependent variable data (remotely sensed chlorophyll a) are resampled to the same spatial resolution, and a CNN-Attention model integrating convolutional neural networks and channel attention mechanisms is constructed to achieve spatiotemporal simulation prediction of chlorophyll a. Specifically: (2-1) The hourly-scale independent variable data (meteorological, hydrodynamic, and water quality) and dependent variable data (retrieved chlorophyll a from remote sensing) were resampled to the same spatial resolution using the inverse distance weighting (IDW) interpolation method. In this example, the independent variable data included meteorological data (precipitation, evaporation, wind speed, radiation, temperature, and humidity) obtained from ERA5 and hydrodynamic (water level, flow velocity) and water quality (dissolved oxygen, phosphate, total phosphorus, ammonia nitrogen, nitrate nitrogen, chemical oxygen demand, and biochemical oxygen demand) data obtained from the ecohydrodynamic model. The dependent variable data was lake chlorophyll a data retrieved from GOCI remote sensing. All data were resampled to a spatial resolution of 1 km.

[0158] (2-2. Construct model training data based on the image data of the independent and dependent variables, dividing it into training and validation sets in a 7:3 ratio, and establish a convolutional neural network-attention mechanism (CNN-Attention) model to perform spatiotemporal simulation prediction of chlorophyll a in lakes. Specific steps include:) ① Input the model training data into the CNN encoder, and extract high-level semantic feature maps through alternating operations of convolutional layers and pooling layers; ② Construct an attention mechanism module, which calculates the attention weight matrix based on the feature dimension and spatial dimension of the high-level semantic feature map, and outputs the attention contribution value of each feature through a normalized exponential function, as shown in the following formula: ; in and For trainable weight matrix, For feature dimension, F The input feature data.

[0159] ③ Perform element-wise multiplication between the attention weight matrix and the feature map output by the CNN to generate attention-enhanced features, and input them into a fully connected layer or decoder to calculate and output the prediction result of lake chlorophyll a.

[0160] (3) Threshold extraction was performed based on the chlorophyll a dataset to extract cyanobacterial bloom patches in lakes. Noise and small patches were removed through patch aggregation and other processing to obtain hourly spatial distribution data of cyanobacterial bloom patches in lakes; specifically including: (3-1) Based on the distribution results of chlorophyll a concentration in lakes output by the CNN-Attention model, a chlorophyll a concentration threshold is set to generate a binary mask for extracting cyanobacterial bloom patches in lakes.

[0161] (3-2) Connectivity labeling is performed on the extracted cyanobacterial bloom patches to identify independent cyanobacterial bloom patches. A pixel count threshold for the patches should be set to exclude noise and small-scale cyanobacterial bloom aggregation areas. Then, morphological operations are performed, such as expanding the patch boundaries, filling small holes and connecting adjacent patches through dilation, and removing small protrusions at the edges through erosion to ensure the integrity of the patch shape, thereby obtaining ecologically significant hourly-scale lake cyanobacterial bloom patch data.

[0162] According to one aspect of this application, the channel attention calculation process is as follows: ; ; ; in For the first Global average pooling results for each channel. , For two fully connected layers, weights The compression ratio is... This is the Sigmoid activation function.

[0163] It should be noted that in the Taihu Lake cyanobacterial bloom research scenario, the research objective is to predict algal bloom outbreaks (high concentration, surface aggregation state), therefore the AFAI method is valid in the application scenario of this technical solution. However, it is not applicable to low-concentration situations where algae are uniformly distributed in the water column.

[0164] According to one aspect of this application, the specific architecture design of the CNN-Attention model is as follows: The model input layer receives dimensions of The three-dimensional tensor, in which Input the number of channels (6 meteorological parameters + 2 hydrodynamic parameters + 7 water quality parameters). and This represents the number of rows and columns of the lake raster after resampling.

[0165] The encoder consists of three sets of convolutional blocks, each containing two convolutional layers and one max-pooling layer: the first set uses 64... Convolution kernel, stride 1, padding 1, pooling window The second group used 128. Convolution kernels, parameters as before; the third group uses 256 kernels. The convolutional kernels have the same parameters as before. Each convolutional layer is followed by a BatchNorm layer and a ReLU activation function.

[0166] The attention module is located after the encoder output and processes the feature map output by the encoder. Flattened into a sequence Through learnable projection matrices Generate query, key, and value matrices, calculate attention output, and then restore the spatial dimensions.

[0167] The decoder section employs three sets of deconvolution blocks (transposed convolution), gradually restoring the original spatial resolution through upsampling, and finally... The chlorophyll a prediction map of a single channel output by the convolutional layer.

[0168] The model was trained using the Adam optimizer with an initial learning rate of The batch size is 16, and the training epochs are 200. The loss function used is mean squared error (MSE). ; in The chlorophyll a concentration predicted by the model. This represents the chlorophyll a concentration retrieved via remote sensing. This represents the total number of valid pixels.

[0169] According to one aspect of this application, the chlorophyll a concentration threshold is set at 20 μg / L (corresponding to a cyanobacterial cell density of approximately 5 × 10⁻⁶) based on the ecological standards for cyanobacterial blooms in Taihu Lake. 5 cells / L, which is the warning threshold for algal blooms in Taihu Lake. The threshold for the number of patch pixels is set to 9 pixels (corresponding to...). Areas smaller than this (km²) are considered noise and removed. Morphological operations are performed using... The square structuring element undergoes one iteration for the dilation operation and one iteration for the erosion operation.

[0170] According to one aspect of this application, the training samples are constructed using a time-by-time one-to-one mapping: for each time with valid GOCI observations... The input tensor is composed of the 15 independent variable channel raster data at that moment. Using remote sensing inversion of chlorophyll a raster data at the corresponding time point as labels This constitutes a training sample pair. Only GOCI images with cloud coverage below 30% were included in the training set. The training and validation sets were divided chronologically (the first 70% of the time period was the training set, and the last 30% was the validation set) to avoid data leakage.

[0171] According to one aspect of this application, the EFDC model parameter calibration employs a step-by-step calibration method: First, the hydrodynamic parameters are calibrated, with the bottom roughness coefficient (Manning coefficient) ranging from 0.02 to 0.04 and the horizontal turbulent viscosity coefficient ranging from 1.0 to 5.0 m² / s, verified by comparing simulated water level and flow velocity with measured values; second, the water quality parameters are calibrated, including the maximum algal growth rate and half-saturation constant, using monthly measured water quality concentrations for calibration. The calibration target is a Nash-Sutcliffe efficiency coefficient (NSE) > 0.6 and a relative error < 30%.

[0172] According to one aspect of this application, step (2-1) further comprises: Based on the standardization of multi-source data, a wind-driven Lagrangian drift feature map construction process is introduced to explicitly encode the dynamic information of spatial transport of algal blooms as additional input features of CNN.

[0173] The specific implementation process is as follows: Based on the hourly surface velocity field output by the EFDC model. 10m wind speed field of ERA5 meteorological data Construct an effective surface drift velocity field:

[0174] ; in This is the wind-induced surface drift coefficient, with a value ranging from 0.01 to 0.03 (an empirical value, indicating that the surface water drift speed is approximately 1% to 3% of the wind speed).

[0175] Based on this effective drift velocity field, for each grid point in the lake Perform backward Lagrangian trajectory tracking and calculate the target prediction time. Before Location of the drift source region within the time window: ; ; in The backward tracking time window was set to 6 hours based on the Taihu Lake scale (the drift distance in 6 hours under typical wind speeds in Taihu Lake is approximately 3-10 km, which matches the scale of algal bloom patches). The integration was performed using the fourth-order Runge-Kutta method with a time step of 0.5 hours.

[0176] Kernel density estimation is performed on the backward trajectory source region locations of all grid points to generate a drift source region probability density map. This characterizes the probability of each spatial location contributing to the algal bloom at the target point. Simultaneously, it calculates the cumulative drift distance map for each grid point. : ; Will and These two additional channels are concatenated to the original 15-channel input tensor to form 17-channel input data. The input is used to train and predict the CNN-Attention model.

[0177] According to one aspect of this application, step (2-2) further comprises: In the training process of the CNN-Attention model, in addition to the data-driven mean squared error loss, a physical constraint loss term based on algal ecological dynamics is introduced to construct a hybrid loss function, thereby achieving a deep integration of physical mechanisms and data learning.

[0178] The process of constructing the physical constraint loss is as follows: Based on the algal biomass change equation in the EFDC ecological dynamics model, the time change rate of algal chlorophyll a concentration should satisfy: ; in The specific growth rate of algae, Rate of death / respiratory consumption, For the rate of settlement loss, is the horizontal diffusion coefficient.

[0179] Algal growth rate Limited by the combined effects of temperature, light, and nutrients: ; ; ; ; in The maximum specific growth rate (taken as 1.5~2.0d) - ¹), The optimal growth temperature is 25~28°C. The temperature influence coefficient (taken as 0.004~0.008°C) - ²), Photosynthetically active radiation, Let be the light half-saturation constant. This refers to the phosphate concentration. The value is the phosphorus half-saturation constant (taken as 0.01~0.02 mg / L). Inorganic nitrogen concentration, The value is the nitrogen half-saturation constant (taken as 0.02~0.05 mg / L).

[0180] Of the parameters mentioned above, temperature ,illumination Nutrient concentration , All of these can be obtained directly from the CNN input channel, therefore the net growth rate It can be calculated from the input data.

[0181] For two adjacent valid observation times and (interval The physical constraint loss is defined as: ; in Effective pixel count For the model at time For pixels The predicted value is given here, ignoring the diffusion term (at a resolution of 1 km, the diffusion effect is relatively small compared to the growth / drift term).

[0182] The model's total loss function is: ; in The physical constraint weights are initially set to 0.1 and adjusted according to the cosine annealing strategy with each training epoch. ; This strategy involves strong physical constraints in the early stages of training (guiding the model to learn physically consistent predictions), and gradually relaxing the constraints in the later stages to allow the model to freely fit the details of the data.

[0183] According to another aspect of this application, step (2-2) can also be: To capture the continuity and autocorrelation of the temporal evolution of cyanobacterial blooms, an autoregressive feedback mechanism of the preceding bloom state is introduced into the CNN-Attention model, using the model prediction results of the previous time step as additional input features for the current time step.

[0184] The specific implementation method is as follows: Define the time. The model input is an augmented tensor. : ; in The original input features, This represents the chlorophyll a concentration field predicted by the model at the previous time step. This represents the chlorophyll a rate of change field at the previous time step (characterizing the growth / decline trend of algal blooms). The augmented number of input channels is... .

[0185] The training phase employs a hybrid approach combining teacher-mandated strategies and free-flowing training: in the first 100 epochs, real chlorophyll a data from remote sensing observations is used with a 70% probability. (Teacher-forced) The model uses its own previous prediction with a 30% probability (free run); in the next 100 epochs, the teacher-forced ratio is gradually reduced to 0% to allow the model to adapt to the cumulative effect of its own prediction error.

[0186] In the prediction phase (time periods without remote sensing labels), the model is run autoregressively: the first time step is initialized using chlorophyll a data from the most recent valid GOCI observation. In subsequent steps, the prediction results from the previous step are used as feedback inputs.

[0187] To prevent the error from accumulating infinitely in autoregressive predictions, physical constraint pruning is applied after each autoregressive step: ; This means limiting the rate of change in a single step to no more than the maximum possible rate of growth / decay in ecology, thus ensuring the physical rationality of the prediction results.

[0188] According to one aspect of this application, step (1-2) further comprises: Based on the hourly hydrodynamic and water quality parameters output by the EFDC ecodynamic model, the algal biomass simulation results of the EFDC eutrophication module are further extracted to obtain the spatial field of chlorophyll a concentration simulated by EFDC. Simultaneously, a spatial confidence field output by the EFDC model is constructed using a parameter ensemble perturbation method to characterize the reliability of the EFDC simulation results at different spatial locations in the lake.

[0189] Specifically: Regarding the core ecological parameters of the EFDC model, including the maximum algal growth rate... Phosphorus half-saturation constant Nitrogen half-saturation constant Algal mortality rate Manning coefficient for bottom roughness Within a reasonable range of values, Latin hypercube sampling is performed to generate... Group parameter combination ( Each set of parameters independently drives the EFDC model during a representative period, namely the peak period of cyanobacterial blooms from June to September, obtaining... Spatial field of chlorophyll a concentration: ; Take the set mean as a rough estimate of chlorophyll a in EFDC: ; Take the set standard deviation as the model output uncertainty: ; The uncertainty is normalized into a spatial confidence field: ; in For reference uncertainty threshold, the entire lake spatial domain is taken. The median makes the confidence level exist The intervals are approximately uniformly distributed. The closer the value is to 1, the higher the consistency of the EFDC simulation at that position and the greater the reliability of the output; the closer the value is to 0, the greater the discrepancy in the EFDC output under different parameter combinations and the lower the reliability.

[0190] To reduce computational costs, the ensemble simulation only needs to be run during representative periods, and the resulting spatial confidence field is averaged over time and used as a quasi-static spatial template for all periods.

[0191] According to one aspect of this application, step (2-1) further comprises: Based on the completion of multi-source data standardization processing, the mean of the EFDC chlorophyll a set is... and EFDC spatial confidence field These two additional channels are concatenated with the original 15-channel independent variable data to form a 17-channel input tensor. Channel 16 is the coarse estimate of chlorophyll a from EFDC (providing a physical baseline reference), and channel 17 is the spatial confidence level (providing information on the confidence level of the physical baseline).

[0192] According to one aspect of this application, step (2-2) may also be: In the construction and training of the CNN-Attention model, the learning objective of the model is changed from directly predicting the absolute concentration of chlorophyll a to predicting the spatial residual field between the simulated EFDC value and the remote sensing observation value. A residual learning framework of physical baseline + data-driven correction is constructed to achieve deep integration of physical model and deep learning.

[0193] Specifically: Define the EFDC spatial residual field As training labels for CNN: ; in To retrieve chlorophyll a concentration using GOCI remote sensing, To improve the coarse estimate of chlorophyll a in the EFDC ensemble mean obtained in Scheme 4. This residual characterizes the systematic bias of the EFDC physical model at various spatial locations.

[0194] The CNN-Attention model takes a 17-channel input tensor as input, including the coarse estimate of chlorophyll a from EFDC and the spatial confidence field, and outputs a single-channel residual prediction field. ; Final chlorophyll a predicted value Obtained by superimposing the physical baseline and the data correction term: ; In this framework, the CNN does not need to learn the complete mapping from environmental variables to chlorophyll a from zero; it only needs to learn the spatial bias patterns of the EFDC physical model. Since EFDC biases are mainly driven by grid resolution, terrain simplification, and parameter uncertainty, they have spatial continuity and structured characteristics, making them easier to learn than directly learning the absolute value of chlorophyll a. Simultaneously, the spatial confidence channel in the model input provides the CNN with prior guidance on where large corrections are needed (low confidence region) and where only fine-tuning is required (high confidence region).

[0195] The model training loss function consists of a residual fitting term and a residual magnitude regularization term: ; The residual fitting term uses pixel-level quality-weighted mean square error: ; in Effective pixel count For the first The quality weight of remote sensing labels per pixel.

[0196] Pixel-level quality weights The following three factors are considered together: (a) Cloud proximity effect weight: The farther away from the cloud marker pixel, the higher the label quality. ; in For pixels Euclidean distance (in pixels) to the nearest cloud marker pixel. Pixel is the attenuation scale parameter.

[0197] (b) Shoreline proximity effect weight: The farther the pixel is from the land, the smaller the water-land mixing effect. ; in The distance from one pixel to the nearest land pixel. Pixels are transition scale parameters.

[0198] (c) Solar zenith angle weight: The larger the solar altitude angle (the smaller the zenith angle), the shorter the atmospheric path, and the higher the quality of remote sensing data. ; in For this pixel at time The zenith angle of the sun.

[0199] The overall pixel quality weight is the product of three factors: ; The residual magnitude regularization term uses the formula in the above embodiment or existing methods to prevent the CNN from over-correcting the physical model, causing the prediction results to lose their physical meaning.

[0200] A spatiotemporal dynamic simulation and prediction system for lake cyanobacterial blooms that integrates ecological dynamics models and spatiotemporal prediction models includes: a processor and a memory communicatively connected to the processor. The memory stores instructions that the computer executes; The processor executes computer execution instructions stored in memory to implement the methods described in the above embodiments.

Claims

1. A method for spatiotemporal dynamic simulation and prediction of cyanobacterial blooms in lakes, characterized in that, include: Acquire multi-source environmental monitoring data for the target water area, including meteorological, hydrodynamic, and water quality information; The physical baseline chlorophyll concentration field was obtained by simulating and analyzing multi-source environmental monitoring data using a pre-defined ecological dynamics model. Spatial grid resampling and alignment were performed on multi-source environmental monitoring data and physical baseline chlorophyll concentration fields to construct a multi-channel spatial feature tensor. The spatial distribution deviation features are extracted by inputting the multi-channel spatial feature tensor into a pre-trained spatiotemporal prediction model to obtain the residual prediction field. The physical baseline chlorophyll concentration field and the residual prediction field are spatially superimposed and corrected to obtain the composite chlorophyll concentration prediction field. Threshold screening and connected region aggregation were performed based on the composite chlorophyll concentration prediction field to extract and output spatial distribution data of cyanobacterial bloom patches. Spatial grid resampling and alignment were performed on multi-source environmental monitoring data and physical baseline chlorophyll concentration fields to construct a multi-channel spatial feature tensor, including: Spatial interpolation algorithms are used to resample multi-source environmental monitoring data with different time scales and spatial resolutions to a unified spatial resolution grid. The resampled multi-source environmental monitoring data and the aligned physical baseline chlorophyll concentration field are spliced ​​together in the spatial channel dimension to obtain a multi-channel spatial feature tensor. Before stitching the resampled multi-source environmental monitoring data with the physical baseline chlorophyll concentration field in the spatial channel dimension to obtain the multi-channel spatial feature tensor, the following steps are also included: The flow velocity field and wind speed field are extracted from multi-source environmental monitoring data, and the effective drift velocity field of the target water area is calculated based on the preset wind-induced surface drift coefficient. Based on the effective drift velocity field, a high-order numerical integration method is used to perform backward trajectory tracking on the spatial grid to determine the location of the drift source region and the corresponding backward tracking trajectory within a preset time window; A probability map of the drift source region is generated based on the kernel density estimation of the drift source region location, and a cumulative drift distance map is generated based on the backward tracking trajectory. The pre-trained spatiotemporal prediction model consists of a physical model baseline and a deep learning feature mapping. The physical baseline chlorophyll concentration field is provided by the physical model baseline end, and the process of extracting spatial distribution deviation features is performed by the deep learning feature mapping end. When performing simulation predictions for continuous time periods, before inputting the multi-channel spatial feature tensor into a pre-trained spatiotemporal prediction model to extract spatial distribution bias features, the following steps are also included: The composite chlorophyll concentration prediction field and the corresponding time change rate field of the previous prediction time are obtained and integrated into the multi-channel spatial feature tensor of the current time as augmented input channels to construct a time-series input feature with autoregressive feedback. After obtaining the predicted field of complex chlorophyll concentration, the following is also included: Based on the preset maximum specific growth rate of algae and the current prediction time step, calculate the maximum allowable physical change boundary after the evolution of the current prediction time step. Using the composite chlorophyll concentration prediction field at the previous prediction time as a benchmark, the current composite chlorophyll concentration prediction field is subjected to extreme value pruning using the maximum allowable physical change boundary, cutting off prediction values ​​that exceed the physical growth limit.

2. The method according to claim 1, characterized in that, The spatial distribution bias features are extracted by inputting a multi-channel spatial feature tensor into a pre-trained spatiotemporal prediction model to obtain the residual prediction field, including: By utilizing the encoder and attention mechanism modules in the deep learning feature mapping end, feature extraction is performed on the multi-channel spatial feature tensor, and the residual prediction field representing the deviation of the physical model is output. The physical baseline chlorophyll concentration field and the residual prediction field are spatially superimposed and corrected to obtain the composite chlorophyll concentration prediction field, including: The physical baseline chlorophyll concentration field provided by the physical model baseline end is added grid-by-grid to the residual prediction field output by the deep learning feature mapping end to obtain the composite chlorophyll concentration prediction field.

3. The method according to claim 1, characterized in that, Before stitching the resampled multi-source environmental monitoring data with the physical baseline chlorophyll concentration field in the spatial channel dimension to obtain the multi-channel spatial feature tensor, the following steps are also included: Latin hypercube sampling is performed on the core ecological parameters of the pre-defined ecological dynamics model to generate multiple sets of parameters; By using multiple sets of parameters and multi-source environmental monitoring data to drive ecological dynamics models to extrapolate environmental fields, multiple sets of spatial concentration simulation fields are obtained. Based on the set discretization degree of multiple spatial concentration simulation fields, a normalized spatial confidence field is constructed; The resampled multi-source environmental monitoring data and the physical baseline chlorophyll concentration field are spliced ​​together in the spatial channel dimension to obtain a multi-channel spatial feature tensor. It also includes splicing and integrating the spatial confidence field as an additional spatial channel to characterize the simulation confidence of each grid location.

4. The method according to claim 1, characterized in that, Threshold screening and connected region aggregation were performed based on the composite chlorophyll concentration prediction field to extract and output spatial distribution data of cyanobacterial bloom patches, including: The composite chlorophyll concentration prediction field is binarized based on a preset ecological concentration threshold to generate an algal bloom patch mask. Connected regions are marked on the algal bloom patch mask, and the number of pixels in each connected region is counted. Isolated noise patches with a number of pixels lower than a preset area threshold are removed to obtain the filtered algal bloom patch mask. Morphological dilation was used to fill the internal cavities of the remaining patches in the screened cyanobacterial bloom mask and connect adjacent patches. Morphological erosion was used to remove the protruding areas at the edges of the patches, thus obtaining spatial distribution data of morphologically intact cyanobacterial bloom patches.

5. The method according to claim 1, characterized in that, The pre-trained spatiotemporal prediction model is obtained through the following steps: Acquire multi-source environmental monitoring data for historical periods, the corresponding historical physical baseline chlorophyll concentration field, and the historical remote sensing inversion chlorophyll concentration field as a true label, and construct training sample pairs; The training sample pairs are input into the spatiotemporal prediction model to be trained, and the prediction output is calculated. Based on the difference between the predicted output and the true label, the network weights of the spatiotemporal prediction model to be trained are iteratively updated using a pre-set multi-objective loss function until the convergence condition is met, thus obtaining the pre-trained spatiotemporal prediction model.

6. The method according to claim 5, characterized in that, The preset multi-objective loss function includes a physical constraint loss term based on the ecological dynamics mechanism of algae; The calculation process of the physical constraint loss term based on algal ecodynamics includes: Based on the temperature, light and nutrient concentration data extracted from the training sample pairs, the degree of limitation on the preset maximum algal growth rate is calculated, and then the theoretical net algal growth rate at the corresponding spatial grid position of the network is determined. The predicted output results of adjacent observation times are extracted to calculate the time difference rate of change, and the time difference rate of change is compared with the theoretical physical rate of change derived from the theoretical net algal growth rate to obtain the physical deviation penalty value. The weight coefficient of the physical deviation penalty value in the multi-objective loss function is dynamically adjusted according to the preset cosine annealing strategy, so that the spatiotemporal prediction model to be trained is subject to physical constraints in the early stage of training.

7. A spatiotemporal dynamic simulation and prediction system for lake cyanobacterial blooms integrating ecological dynamics models and spatiotemporal prediction models, characterized in that, include: The processor, and the memory that is communicatively connected to the processor; The memory stores instructions that the computer executes; The processor executes computer execution instructions stored in memory to implement the method as described in any one of claims 1 to 3.

Images