Multi-temporal-spatial-scale cyanobacterial bloom early warning method for middle and large lake and reservoir water areas

Abstract

The invention discloses a multi-temporal-spatial-scale cyanobacterial bloom early warning method for middle and large lake and reservoir water areas, and belongs to the field of cyanobacterial bloom prediction and risk monitoring. The method comprises the following steps: generating a pixel-level FAI index based on target water area remote sensing data, resampling meteorological data into a pixel level, then constructing a spatial-temporal distribution data set, training an Autoformer-ST-GNN time sequence model to realize FAI index prediction, dividing cyanobacterial bloom levels according to the FAI index prediction, and obtaining a global change trend; water quality monitoring points are arranged in key areas to collect data, historical water quality and meteorological data are utilized to train a DMC-PatchTST model fused with a blue-green algae migration period, and multi-time-scale prediction of the density of blue-green algae at the monitoring points is achieved; and finally, combining the global trend with a monitoring point prediction result to construct a space-time multi-scale cyanobacterial bloom comprehensive early warning system. According to the method, multi-source data and multiple models are fused, so that cyanobacterial bloom time-space multi-scale comprehensive early warning is realized, and the method is accurate and comprehensive.

Description

Technical Field

[0001] The present invention belongs to the field of cyanobacterial bloom prediction and risk monitoring, and specifically relates to a multi-temporal and spatial scale cyanobacterial bloom early warning method for medium and large lake and reservoir waters. Background Art

[0002] Frequent algae blooms severely impact the sustainable development of aquatic ecosystems, posing a serious threat to drinking water safety, public health, and aquaculture. Therefore, accurate prediction and risk assessment of algae blooms are essential. Research on early warning technologies for algae blooms can help water environment managers obtain timely warning information before an outbreak occurs, enabling them to implement appropriate control measures and effectively reduce the frequency and scale of blooms.

[0003] Chinese invention patent CN117172411A discloses a method and system for automatically identifying and warning of cyanobacterial blooms in shallow lakes in real time, around the clock. By combining remote sensing data from geostationary satellites, video surveillance of the near-surface water surface, and data from water quality monitoring points, this system builds a ground-to-air cyanobacterial bloom monitoring platform. The method incorporates satellite remote sensing warnings, video surveillance warnings, and the fusion of meteorological and water quality monitoring data.

[0004] The above-mentioned comparative documents have the following defects: (1) Weak time series prediction capability: This early warning method is based on current or short-term observation data (satellite remote sensing data, real-time meteorological data, water quality data), and uses numerical threshold judgment to give the current algal bloom level assessment. It lacks long-term modeling of algal density changes, does not capture the characteristics of cyanobacteria bloom changes, and is difficult to complete the prediction of future algal density at multiple time scales (such as 3 days, 10 days). (2) Insufficient meteorological data drive: This method obtains meteorological data only for threshold judgment, does not explore its impact on the biological mechanism of cyanobacteria, and has low prediction accuracy. (3) Satellite remote sensing early warning only assesses the current status of water bloom characteristics by inverting the proportion of water bloom area, which is difficult to fully support the precise early warning function and the dynamic monitoring and prediction of water blooms in the entire region. Summary of the Invention

[0005] To address the above issues, the present invention proposes a multi-temporal and spatial scale cyanobacterial bloom early warning method for medium- and large-scale lakes and reservoirs. The technical solution is as follows:

[0006] A multi-temporal and spatial scale cyanobacterial bloom early warning method for medium- to large-scale lakes and reservoirs comprises the following steps:

[0007] Based on the remote sensing data of the target water area, the FAI index of each pixel in the water area is generated; the first meteorological data is collected and resampled to each pixel to form a global pixel-by-pixel meteorological-FAI index spatiotemporal distribution dataset; the historical spatiotemporal distribution dataset is used to train a time series prediction model composed of an Autoformer combined with an ST-GNN. After training, the FAI index of each pixel in the water area is predicted by acquiring remote sensing data and the first meteorological data in real time. The cyanobacteria bloom level is classified according to the FAI index, and the global change trend of the cyanobacteria bloom in the water area is obtained;

[0008] Water quality parameter monitoring points are set up at key monitoring locations within the target waters to obtain water quality data. A time series of algae indicators, consisting of historical water quality data from the monitoring points and secondary meteorological data, is used to train a time series prediction model, DMC-PatchTST, that incorporates the cyclical characteristics of cyanobacteria migration. This model, through real-time input of water quality data and secondary meteorological data, is used to predict the future cyanobacteria density at the monitoring points over multiple time scales.

[0009] By combining the global change trend of cyanobacteria blooms in waters and the multi-time-scale prediction results of future cyanobacteria density at monitoring points, a comprehensive early warning of cyanobacteria blooms at multiple time and space scales can be achieved.

[0010] Furthermore, the calculation process of the FAI index includes:

[0011] Preprocess the remote sensing data to obtain the Rayleigh scattering reflectance data corresponding to each pixel;

[0012] The FAI index is obtained by taking the difference between the actual Rayleigh scattering reflectivity in the near-infrared band and the predicted Rayleigh scattering reflectivity in the near-infrared band.

[0013] Furthermore, the first meteorological data includes air pressure, temperature, wind speed, precipitation, and light.

[0014] Furthermore, the aforementioned use of the historical spatiotemporal distribution dataset to train a time series prediction model by combining Autoformer with ST-GNN includes:

[0015] Extract seasonal terms and trend terms from the time series data of each pixel;

[0016] Using the time series data X of each pixel i The graph structure G = (V, E, A) between the lake and reservoir water body pixels is constructed based on the association strength between the pixels. V is the pixel node set, E is the edge set based on spatial relations, and A is the adjacency matrix used to calculate the association strength between pixels. The calculation formula is:

[0017]

[0018] in, is the distance factor, (α+β·cosθ ij ) is the water flow direction factor under the influence of wind speed, d ij is the Euclidean distance between pixels, σ d is the distance attenuation parameter, θ ij is the angle between the pixel and the wind speed vector, α and β are the basic connection strength parameters and the wind speed direction influence weight respectively, d throshold is the distance threshold, σ d , α, and β are all learnable parameters;

[0019] Use ST-GNN to extract the features of each node in the graph structure G. Each node feature corresponds to a pixel feature, which is recorded as the first feature.

[0020] The splicing result of the seasonal term of the past period corresponding to each pixel and the first feature is input into the Autoformer model to predict the seasonal term of each pixel in the future period; the splicing result of the trend term of the past period corresponding to each pixel and the first feature is input into the Autoformer model to predict the trend term of each pixel in the future period; the two are added as the time series data of each pixel in the future period, and the time series data contains the pixel FAI index.

[0021] Furthermore, the method for extracting seasonal terms and trend terms from the time series data of each pixel is as follows:

[0022] Use moving average filter to filter the time series data X of each pixel i Perform time series decomposition and extract seasonal and trend terms:

[0023]

[0024] Among them, X i represents the time series data of the i-th pixel, Represents the trend term of the i-th pixel, which is pooled by moving average Extracted, k is the sliding window size of the moving average; Represents the seasonal term, which is obtained by the difference between the time series data and the trend term.

[0025] Furthermore, the water quality data includes water temperature, turbidity, total nitrogen, total phosphorus, cyanobacteria density and chlorophyll a.

[0026] Furthermore, the second meteorological data includes wind speed, precipitation, and sunlight.

[0027] Furthermore, the process of training a time series prediction model DMC-PatchTST that incorporates the periodic characteristics of cyanobacteria migration includes:

[0028] Autocorrelation function analysis:

[0029]

[0030] in, represents the characteristic value of the algae index time series at the monitoring point at time t, represents the characteristic value of the algae index time series at the monitoring point at time t+k, μ is the mean of the time series, T represents the time length, v(k) is the autocorrelation coefficient of lag k, and the point C where v(k) is the maximum value is found. ACF , as a candidate cycle;

[0031] Perform continuous wavelet transform (CWT) on algae indicators to identify multi-scale periodic features:

[0032]

[0033] Among them, ψ is the wavelet function, a and b are the scale parameter and displacement parameter respectively;

[0034] Establish a regression model between weather variables and algae migration period:

[0035]

[0036] Among them, C env is the period of dynamic estimation, β0 is the intercept coefficient, which represents the intrinsic basic migration period of cyanobacteria when all meteorological factors are standardized to zero, which is set to 24 hours in this method, and D met represents the number of meteorological characteristics, β i It represents the partial regression coefficient of the i-th meteorological characteristic, which quantifies the marginal impact of the meteorological factor on the cyanobacteria migration cycle. Its positive or negative value indicates the promotion (positive value) or inhibition (negative value) of the cycle.

[0037] Combining the above three results, the dynamic period C is estimated:

[0038] C=α1·C ACF +α2·C wavelet +α3·C env

[0039] Among them, α1, α2, and α3 are learnable weight coefficients whose sum is 1;

[0040] The dynamic period in the previous step is used as the basis for patch division, and a period position code is added to each patch. The period position code information is then embedded as a feature, and the time series prediction model DMC-PatchTST is used to predict the future cyanobacteria density at the monitoring point.

[0041] Furthermore, the levels of cyanobacteria blooms were divided according to the FAI index: level one FAI ≤ 0, level two 0 < FAI ≤ 0.006 and level three FAI > 0.006.

[0042] The beneficial effects of the present invention are:

[0043] First, multi-source remote sensing data is acquired to generate a global algal bloom distribution profile for the water area. Time series models are then used to predict changes in algal bloom distribution at different future times, enabling global prediction of algal blooms in medium- and large-scale waters. Next, water quality monitoring equipment is deployed at key monitoring points to generate water quality data, which is then input into the model to accurately predict algal density at each monitoring point over different timescales. Combining these two approaches creates a comprehensive, multi-spatiotemporal-scale early warning method for cyanobacterial blooms in medium- and large-scale lakes and reservoirs. This method integrates multi-source data with multiple models to provide accurate and comprehensive early warning of cyanobacterial blooms across multiple spatiotemporal scales. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 This is a flowchart of the multi-temporal and spatial scale cyanobacteria bloom early warning method for medium and large lakes and reservoirs;

[0045] Figure 2 This is an example of pixel-by-pixel data;

[0046] Figure 3 This is an example of the prediction results of algae density at the monitoring point. DETAILED DESCRIPTION

[0047] The present invention will be further described in detail below with reference to the accompanying drawings and examples, but the specific implementation manner of the present invention is not limited thereto.

[0048] This paper proposes a multi-temporal and spatial scale cyanobacterial bloom early warning method for medium and large lakes and reservoirs, focusing on the following aspects:

[0049] Integrate future weather forecast information to improve the foresight and accuracy of cyanobacteria bloom predictions;

[0050] A periodic detection module based on environmental factors was designed to detect the periodic response of cyanobacteria to environmental factors (light, precipitation, wind speed, air pressure, etc.), and was integrated into a data-driven time series prediction model. Actual test results showed that this module can improve the model's ability to predict algae density.

[0051] It can realize the prediction of cyanobacteria blooms at multiple temporal and spatial scales in lake and reservoir waters, which can not only realize the global and large-scale dynamic change monitoring of cyanobacteria blooms in lake and reservoir waters, but also set water quality monitoring stations according to needs to predict the future algal bloom level at that point to meet different environmental control needs.

[0052] Integrate multi-source data to overcome the limitations of a single data source under specific conditions.

[0053] like Figure 1 As shown in FIG, a multi-temporal and spatial scale cyanobacterial bloom early warning method for medium and large lakes and reservoirs mainly includes the following steps:

[0054] S1: Acquire multi-source remote sensing data, process the remote sensing data, generate a distribution of cyanobacteria bloom levels in the water area, collect meteorological data, resample it to each image, and construct a global meteorological-cyanobacteria bloom spatiotemporal distribution dataset. Design an Autoformer-based time series prediction model architecture combined with ST-GNN, input the data into the time series prediction model for training. After training, the model predicts the overall cyanobacteria bloom changes by acquiring current remote sensing data and meteorological data in real time, and then issues an early warning. This includes the following steps:

[0055] (1) Satellite remote sensing data acquisition: remote sensing data from different sources are obtained, including MODIS Tera / Aqua (Moderate Resolution Imaging Spectroradiometer Terra / Aqua, carried on Terra and Aqua satellites, with a resolution of 250m / 500m, open source, and an observation frequency of once a day) and GOCI-II (Geostationary Ocean Color Imager-II, a second-generation geostationary ocean color imager with a spatial resolution of 500m, open source, and an observation frequency of 10 times a day).

[0056] Meteorological data collection: ERA5-LAND (ECMWF Reanalysis v5 Land, the fifth generation of land reanalysis data from the European Centre for Medium-Range Weather Forecasts, spatial resolution: 9 km, open source, temporal resolution: hourly) and LWQ 300m (LandWater Quality 300m, land water quality data with a resolution of 300 meters, open source, spatial resolution: 300 meters, temporal resolution: 10 days).

[0057] (2) Remote sensing data preprocessing: First, the MODIS data and GOCI data are automatically screened, and cloud-free remote sensing images covering all water areas are obtained through image classification algorithms. Then, the SeaDAS software is used for preprocessing, the corresponding geographical location and corresponding data are read, and geometric correction and spatial atmospheric correction are performed. Finally, the Rayleigh scattering reflectance (Rrc(λ)) data is obtained for the calculation of the FAI index.

[0058] By extracting and processing ERA5-land meteorological data, we can obtain hourly meteorological data with a resolution of approximately 9 km, including vector wind speed data, precipitation, sunlight, atmospheric pressure, and temperature. LWQ 300m water quality data provides lake nutrient status and turbidity data at a resolution of 300 m. After extraction, spatiotemporal interpolation methods are used to fill missing values.

[0059] (3) Inversion of cyanobacteria information: For the remote sensing data obtained in the previous step, the FAI (Floating Algae Index) value of each pixel point (pixel point is a commonly used term in this field) in the monitored lake and reservoir area is calculated. The pixel-by-pixel data example is as follows: Figure 2 shown.

[0060] The FAI index is a remote sensing index used to reflect cyanobacteria blooms in water bodies. The source is: Hu, Chuanmin. (2009). A novel ocean color index to detect floating algae in the global oceans. Remote Sensing of Environment, 113 (10), 2118-2129. Its formula is as follows:

[0061] FAI=R NIR -R′ NIR (1)

[0062] Among them, R NIR is the actual Rayleigh scattering reflectivity in the near-infrared band, R′ NIR It is the predicted Rayleigh scattering reflectivity of the near-infrared band obtained by linear interpolation based on the Rayleigh scattering reflectivity of the red band and the short-wave infrared band. The calculation formula is as follows:

[0063]

[0064] Among them, R red is the Rayleigh scattering reflectivity in the red light band, R SWIR is the Rayleigh scattering reflectivity in the shortwave infrared band, λ NIR is the wavelength of the near-infrared band, λ red is the wavelength of the red light band, λ SWiR is the wavelength of the shortwave infrared band; for MODIS data, R red 、R NIR 、R SWIR Select 645nm(λ red )、859nm(λ NIR ) and 1240nm(λ SWIR ) Rayleigh scattering reflectivity data at three wavelengths after atmospheric correction.

[0065] (4) Identification and extraction of cyanobacterial bloom pixels: The FAI index of each pixel of the remote sensing image is matched with the FAI algae bloom intensity level threshold of the monitored lake and reservoir waters. Pixels exceeding the threshold are defined as cyanobacterial bloom pixels. For example, for Taihu Lake, after investigation, its bloom risk can be divided into three levels: Level 0 (pure water pixel, FAI≤0), Level 1 (0<FAI≤0.006) and Level 2 (FAI>0.006). After calculating the FAI index, the bloom intensity of each pixel is determined and marked with 0, 1, and 2.

[0066] (5) Generate pixel-by-pixel FAI index-meteorological time series data set: Align the meteorological data (temperature, precipitation, vector wind speed, atmospheric pressure, and light) with the FAI index according to location and time to generate pixel-by-pixel FAI index-meteorological time series data. Each pixel contains the FAI index and the meteorological data information at the corresponding moment. An example is shown in Table 1.

[0067] Table 1 Minimum pixel data example

[0068]

[0069] Among them, the air pressure and wind speed use the daily average value, the precipitation uses the cumulative value, the light uses the average value during sunrise, and the temperature uses the daily maximum value. High-quality cloud-free remote sensing images are selected from the MODIS / GOCI-II daily data to calculate the FAI index.

[0070] (6) Time series prediction model design: A hybrid architecture combining Autoforme and ST-GNN (Spatio-Temporal Graph Neural Networks) is used. Autoformer processes the FAI index-meteorological time series data for each pixel, extracts the time series characteristics of the meteorological and FAI indexes, and captures periodic and trend patterns. The core of this model is the autocorrelation mechanism based on random process theory: by calculating the autocorrelation coefficients of the time series at different time delays, a periodic autocorrelation matrix is ​​generated. ST-GNN can process the spatial correlation between lake and reservoir pixels, establish a graph structure relationship between pixels, and simulate the spatial dynamic process of algal bloom spread.

[0071] First, the time series data X of each minimum pixel is i (At least one year) Use ST-GNN for pixel space modeling, first constructing the graph structure between lake and reservoir water pixels:

[0072] G=(V, E, A) (3)

[0073] Among them, V is the pixel node set, E is the edge set based on spatial relationships, and A is the adjacency matrix used to calculate the association strength between pixels.

[0074] Adjacency matrix A is calculated as follows:

[0075]

[0076] in, is the distance factor, where d ij is the Euclidean distance between pixels, σ d is the distance attenuation parameter. (α+β·cosθ ij ) is the water flow direction factor under the influence of wind speed, θ ij is the angle between the pixel and the wind speed vector, α and β are the basic connection strength parameter and the wind speed direction influence weight respectively. throshold is the distance threshold used to limit the sparsity of the graph and is set to 750m (3 pixels). d , α, and β are all learnable parameters.

[0077] Use ST-GNN to extract the first feature ai of each node in the graph structure G.

[0078] It should be noted that ST-GNN uses spectral domain graph convolution to capture the spatial structure of cyanobacteria migration:

[0079]

[0080] in, The adjacency matrix after adding the self-loop, for The degree matrix, H (l) is the node feature of the lth layer, W (l) are learnable weights.

[0081] The hybrid architecture integrates spatiotemporal information through a gating mechanism:

[0082] Z t =g t ×H t Autoformer +(1-g t )×H t ST-GNN (6)

[0083] The weights of temporal features and spatial features can be dynamically adjusted according to different situations, where the gating signal g t Computed via a feedforward network with low computational complexity:

[0084] g t =sigmoid(W g [H t Autoformer ;H t ST-GNN ]+b g ) (7)

[0085] Where W g 、b g are learnable parameters.

[0086] Define the comprehensive loss function:

[0087]

[0088] in Forecasting the intensity of algal blooms and predicting losses, Predict losses for spatial distribution, is the time series prediction loss, and Ω(W) is the regularization term.

[0089] Secondly, the moving average filter is used to extract the time series data X of each minimum pixel i Perform time series decomposition (at least one year) to extract seasonal and trend terms:

[0090]

[0091] Among them, X i represents the time series data of the i-th pixel, represents the trend term of the i-th pixel, represents seasonal terms;

[0092] Among them, the trend item is extracted by Avgpool (moving average pooling):

[0093]

[0094] Here, k is the sliding window size of the moving average, which is adjusted according to the growth cycle of cyanobacteria and is generally set to 1-5 months.

[0095] The seasonal term is obtained by taking the difference between the time series data and the trend term:

[0096]

[0097] Where k is the sliding window size, which is determined by the growth period of cyanobacteria.

[0098] The concatenation result of the seasonal term of the past period corresponding to each pixel and the ai is input into the Autoformer model (Source: Wu H, Xu J, Wang J, Long M. Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting. NeurIPS2021.) to predict the seasonal term of each pixel in the future period; the concatenation result of the trend term of the past period corresponding to each pixel and the ai is input into the Autoformer model to predict the trend term of each pixel in the future period; the two are added as the time series data of each pixel in the future period.

[0099] In this embodiment, the processed FAI index-meteorological time series data of lake and reservoir waters are divided into training set and test set in a ratio of 7:3, and the training set is input into the Autoformer-ST-GNN model for training.

[0100] (7) Pixel-by-pixel algal bloom prediction in lake and reservoir waters: After training the hybrid architecture model combining Autoformer and ST-GNN, real-time input of algal bloom intensity levels and meteorological data can be used to output global algal bloom dynamics predictions for lake and reservoir waters at different time scales. For example, if the time scale is set to 1 day, 3 days, or 10 days, the model outputs the algal bloom intensity for each pixel in the water area corresponding to the future time. The system assigns each pixel a color corresponding to the intensity level, generating a global algal bloom distribution map.

[0101] (8) Visual output: Different colors are assigned to each pixel according to its algal bloom level, with level 0 being blue, level 1 being yellow, and level 2 being red. This can generate an algal bloom risk prediction map that intuitively reflects the global algal bloom risk in the water area.

[0102] S2, set up water quality parameter monitors at key monitoring locations, obtain water quality parameter data including temperature, turbidity, total phosphorus, total nitrogen, chlorophyll a concentration, and algae density at fixed time intervals, and predict future algae density using the time series prediction model DMC-Pachtst.

[0103] (1) Acquisition and preprocessing of water quality parameter monitoring data: Read relevant water quality parameters of water quality monitoring points at a fixed time frequency of 4 hours, including temperature, turbidity, total phosphorus, total nitrogen, chlorophyll a concentration, and algae density, and process abnormal data (missing values, extreme values). In the missing value processing, short-term missing values ​​of no more than two time points are filled by linear interpolation, and missing values ​​of more than four time points are filled by cubic spline interpolation to achieve a smooth transition. In the extreme value processing, water quality data that obviously deviates from the range are processed. For example, if the temperature range is set to 0-45℃, the deviating data are processed by linear interpolation.

[0104] (2) Meteorological data alignment: The historical meteorological data in the remote sensing data are aligned with the historical water quality parameter data set in time to form an algae indicator time series data set. An example of the data from a monitoring point is shown in Table 2.

[0105] Table 2 Examples of time series of algae indicators

[0106]

[0107] (3) Data processing of algae indicator time series data sets: including data standardization and time series feature extraction.

[0108] Time series feature extraction: define N as the sample batch size, D as the feature dimension, T as the time series length, and the time series matrix as X∈R N×D×T , the features are divided into algae indicators X alg ∈R N×Dalg×T , historical water quality parameters X env ∈R N×Denv×T 、Meteorological dataX met ∈R N×Dmet×F .

[0109] (4) Constructing a dynamic period detection module: Considering the migration behavior of cyanobacteria, this module is used to detect its dynamic migration period Ct, including autocorrelation function analysis, wavelet analysis, and environmental factor correlation analysis.

[0110] Autocorrelation function (ACF) analysis reflects the intrinsic biological rhythm of cyanobacteria:

[0111]

[0112] Among them, X t+kalg represents the characteristic value of the algae index time series at the monitoring point at time t+k, ρ(k) is the autocorrelation coefficient of lag k, and μ is the mean of the time series. Find the maximum point C of ρ(k) ACF , as a candidate cycle.

[0113] Continuous wavelet transform (CWT) was performed on algae indicators to identify the driving force of the external environment on the algae density cycle:

[0114]

[0115] Among them, ψ is the wavelet function, a and b are the scale parameter and displacement parameter respectively.

[0116] A regression model was built between weather variables and the algal migration cycle to quantify the immediate impact of meteorological conditions on the cycle:

[0117]

[0118] Among them, C env is the period of dynamic estimation, β0 is the intercept coefficient, which represents the intrinsic basic migration period of cyanobacteria when all meteorological factors are standardized to zero, which is set to 24 hours in this method, and D met represents the number of meteorological characteristics, β i It represents the partial regression coefficient of the i-th meteorological characteristic, which quantifies the marginal impact of the meteorological factor on the cyanobacteria migration cycle. Its positive or negative value indicates the promotion (positive value) or inhibition (negative value) of the cycle.

[0119] Based on the multi-scale dynamic mechanism of cyanobacteria vertical migration, it is necessary to combine the above three methods to estimate the dynamic period:

[0120] C=α1·C ACF +α2·C wavelet +α3·C env (15)

[0121] α1, α2, and α3 are weight coefficients whose sum is 1 and are learnable.

[0122] Cycle-aware patch partitioning: The adopted patchst model (NIE Y, NGUYEN NH, SINTHONG P, et al. A Time Series is Worth 64 Words: Long-term Forecasting with Transformers[C] / / International Conference on Learning Representations (ICLR 2023). Kigali, Rwanda, 2023.) has a specific parameter patch, i.e., patch. The present invention uses the dynamic cycle of the previous step as the basis for patch partitioning.

[0123] (5) Add periodic position coding to each patch, then embed the periodic position coding information as features, and use the time series prediction model DMC-PatchTST to predict the future cyanobacteria density at the monitoring point.

[0124] In this embodiment, the constructed meteorological-water quality time series dataset is divided into a training set and a test set according to a ratio of 7:3. The training set is input into the DMC-PatchTST model for training, and the model parameters are optimized using the Adam optimizer.

[0125] When predicting the algae density at a monitoring site, historical meteorological data and historical water quality data are input into the model in real time to obtain the algae density prediction results for the monitoring site at three time scales: 4 hours, 3 days, and 10 days in the future, and to evaluate the algal bloom risk level at the predicted time. Figure 3 This is an example of the prediction results for the same time scale. Based on the algae density prediction results, the algal bloom level at the corresponding prediction time in the future can be output. The algal bloom level corresponding to different algal densities is based on the Technical Specifications for Remote Sensing and Ground Monitoring and Evaluation of Algal Blooms issued by the Ministry of Ecology and Environment of China. The classification standards are as follows in Table 3:

[0126] Table 3 Bloom degree classification standards

[0127]

[0128] S3, system integration: Automated integration of the above two methods. By acquiring water quality data, meteorological data, and remote sensing data in real time, multi-scale spatiotemporal prediction of algal blooms in medium and large water areas can be achieved, and graded warning information can be generated based on the prediction results at multiple spatiotemporal scales.

[0129] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be considered as equivalent replacement methods and are included in the scope of protection of the present invention.

Claims

1. A multi-temporal and spatial scale cyanobacterial bloom early warning method for medium and large lakes and reservoirs, characterized by: The following steps are involved: Based on the remote sensing data of the target water area, the FAI index of each pixel in the water area is generated; the first meteorological data is collected and resampled to each pixel to form a global pixel-by-pixel meteorological-FAI index spatiotemporal distribution dataset; the historical spatiotemporal distribution dataset is used to train a time series prediction model composed of an Autoformer combined with an ST-GNN. After training, the FAI index of each pixel in the water area is predicted by acquiring remote sensing data and the first meteorological data in real time. The cyanobacteria bloom level is classified according to the FAI index, and the global change trend of the cyanobacteria bloom in the water area is obtained; Water quality parameter monitoring points are set up at key monitoring locations within the target waters to obtain water quality data. A time series of algae indicators, consisting of historical water quality data from the monitoring points and secondary meteorological data, is used to train a time series prediction model, DMC-PatchTST, that incorporates the cyclical characteristics of cyanobacteria migration. This model, through real-time input of water quality data and secondary meteorological data, is used to predict the future cyanobacteria density at the monitoring points over multiple time scales. By combining the global change trend of cyanobacteria blooms in waters and the multi-time-scale prediction results of future cyanobacteria density at monitoring points, a comprehensive early warning of cyanobacteria blooms at multiple time and space scales can be achieved.

2. The multi-temporal and spatial scale cyanobacterial bloom early warning method for medium and large lakes and reservoirs according to claim 1 is characterized in that: The calculation process of the FAI index includes: Preprocess the remote sensing data to obtain the Rayleigh scattering reflectance data corresponding to each pixel; The FAI index is obtained by taking the difference between the actual Rayleigh scattering reflectivity in the near-infrared band and the predicted Rayleigh scattering reflectivity in the near-infrared band.

3. The multi-temporal and spatial scale cyanobacterial bloom early warning method for medium and large lakes and reservoirs according to claim 1 is characterized in that: The first meteorological data includes air pressure, temperature, wind speed, precipitation, and sunlight.

4. The multi-temporal and spatial scale cyanobacterial bloom early warning method for medium and large lakes and reservoirs according to claim 1 is characterized in that: The aforementioned method of using a historical spatiotemporal distribution dataset to train a time series prediction model using an Autoformer combined with an ST-GNN includes: Extract seasonal terms and trend terms from the time series data of each pixel; Using the time series data X of each pixel i The graph structure G = (V, E, A) between the lake and reservoir water body pixels is constructed based on the association strength between the pixels. V is the pixel node set, E is the edge set based on spatial relations, and A is the adjacency matrix used to calculate the association strength between pixels. The calculation formula is: in, is the distance factor, (α+β·cosθ ij ) is the water flow direction factor under the influence of wind speed, d ij is the Euclidean distance between pixels, σ d is the distance attenuation parameter, θ ij is the angle between the pixel and the wind speed vector, α and β are the basic connection strength parameters and the wind speed direction influence weight respectively, d throshold is the distance threshold, σ d , α, and β are all learnable parameters; Use ST-GNN to extract the features of each node in the graph structure G. Each node feature corresponds to a pixel feature, which is recorded as the first feature. The splicing result of the seasonal term of the past period corresponding to each pixel and the first feature is input into the Autoformer model to predict the seasonal term of each pixel in the future period; the splicing result of the trend term of the past period corresponding to each pixel and the first feature is input into the Autoformer model to predict the trend term of each pixel in the future period; the two are added as the time series data of each pixel in the future period, and the time series data contains the pixel FAI index.

5. The multi-temporal and spatial scale cyanobacterial bloom early warning method for medium and large lakes and reservoirs according to claim 4 is characterized in that: The method for extracting seasonal terms and trend terms from the time series data of each pixel is: Use moving average filter to filter the time series data X of each pixel i Perform time series decomposition and extract seasonal and trend terms: Among them, X i represents the time series data of the i-th pixel, Represents the trend term of the i-th pixel, which is pooled by moving average Extracted, k is the sliding window size of the moving average; Represents the seasonal term, which is obtained by the difference between the time series data and the trend term.

6. The multi-temporal and spatial scale cyanobacteria bloom early warning method for medium and large lakes and reservoirs according to claim 1 is characterized in that: The water quality data include water temperature, turbidity, total nitrogen, total phosphorus, cyanobacteria density and chlorophyll a.

7. The multi-temporal and spatial scale cyanobacteria bloom early warning method for medium and large lakes and reservoirs according to claim 1 is characterized in that: The second meteorological data includes wind speed, precipitation, and sunlight.

8. The multi-temporal and spatial scale cyanobacterial bloom early warning method for medium and large lakes and reservoirs according to claim 1 is characterized in that: The process of training a time series prediction model DMC-PatchTST that combines the periodic characteristics of cyanobacteria migration includes: Autocorrelation function analysis: in, represents the characteristic value of the algae index time series at the monitoring point at time t, represents the characteristic value of the algae index time series at the monitoring point at time t+k, μ is the mean of the time series, T represents the time length, ρ(k) is the autocorrelation coefficient of lag k, and the maximum value point C of ρ(k) is found. ACF , as a candidate cycle; Perform continuous wavelet transform (CWT) on algae indicators to identify multi-scale periodic features: Among them, ψ is the wavelet function, a and b are the scale parameter and displacement parameter respectively; Establish a regression model between weather variables and algae migration period: Among them, C env is the period of dynamic estimation, β0 is the intercept coefficient, which represents the intrinsic basic migration period of cyanobacteria when all meteorological factors are standardized to zero, and D met represents the number of meteorological characteristics, β i represents the partial regression coefficient of the i-th meteorological feature; Combining the above three results, the dynamic period C is estimated: C=α1·C ACF +α2·C wavelet +α3·C env Among them, α1, α2, and α3 are learnable weight coefficients whose sum is 1; The dynamic period in the previous step is used as the basis for patch division, and a period position code is added to each patch. The period position code information is then embedded as a feature, and the time series prediction model DMC-PatchTST is used to predict the future cyanobacteria density at the monitoring point.

9. The multi-temporal and spatial scale cyanobacteria bloom early warning method for medium and large lakes and reservoirs according to claim 8, characterized in that: The levels of cyanobacteria blooms are divided according to the FAI index: level one FAI≤0, level two 0<FAI≤0.006 and level three FAI>0.006.

Images