A method for predicting the spatial distribution of algal blooms in lakes at an hourly scale based on geostationary satellites.

Abstract

This invention discloses a method for predicting the spatial distribution of algal blooms in lakes at an hourly scale based on geostationary satellites. It acquires hourly meteorological data such as wind speed, wind direction, and temperature, along with remote sensing data from the GOCI series satellites. Algal bloom coverage is extracted using an algal bloom identification index. Based on prior knowledge of wind-driven algal bloom migration and accumulation, new feature indicators are constructed using wind direction and remote sensing indices. Representative algal bloom samples are selected, and data on wind speed, temperature, and remote sensing indices for the same time period are extracted to construct a sample database. A machine learning model is trained and its accuracy is verified. The model is applied to a typical eutrophic lake to test its accuracy. Using the current algal bloom distribution and subsequent environmental data, the model iteratively predicts the hourly spatial distribution changes of algal blooms in the lake. This method can predict and warn of the spatial location and outbreak intensity of algal blooms in lakes at an hourly scale, assisting in the prevention and control of algal blooms in eutrophic lakes.

Description

Technical Field

[0001] This invention belongs to the field of satellite remote sensing and water environment analysis technology, specifically involving a method for predicting the spatial distribution of algal blooms in lakes on an hourly scale based on geostationary satellites. Background Technology

[0002] Lakes worldwide are facing widespread ecological and environmental problems such as eutrophication and harmful algal blooms. The proportion of eutrophic lakes and reservoirs is increasing year by year, and algal blooms are becoming more frequent, seriously threatening aquatic ecological security. Cyanobacteria' unique cellular structure and community characteristics endow them with the ability to migrate vertically within a water column to obtain light and food, making them the dominant species in phytoplankton communities. Under ideal meteorological and hydrological conditions, algal particles accumulate on the water surface, forming cyanobacterial blooms and triggering water crises.

[0003] The area of ​​algal blooms varies greatly due to environmental factors, and there are even instances where algal blooms accumulate in large quantities within hours and then disappear rapidly, which are difficult to capture using conventional methods. Remote sensing satellite data has advantages such as wide coverage, fast real-time processing, and periodicity, and has been widely used for long-term monitoring of algal blooms. Algal bloom prediction models can be divided into two categories: (1) mechanism-driven prediction models, which simulate and predict the ecological changes of algal blooms by studying hydrodynamics and algal bloom formation mechanisms. (2) data-driven prediction models, which use massive amounts of data to analyze the relationship between algal bloom indicators and various environmental factors to construct algal bloom prediction models. In general, mechanism-driven prediction models intuitively present the changes in mechanism components and realize the prediction of spatial variation of algal communities, but the calculation is complex and there are many parameters, requiring the construction of different simulation models according to the characteristics of lakes. Data-driven prediction models do not require complex calculations, saving time and effort, and can realize algal bloom prediction by extracting potential patterns from historical monitoring data. However, published studies usually focus on the time-series prediction of algal bloom indicators at sampling points or the average of the entire lake within the study area, lacking research on spatial variation prediction. Therefore, it is crucial to construct simple and effective prediction models based on prior knowledge of environmental factors driving algal blooms and migration, so as to predict the spatial location and intensity of algal blooms in the short term. Summary of the Invention

[0004] The purpose of this invention is to provide a method for predicting the spatial distribution of algal blooms in lakes on an hourly scale based on geostationary satellites.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A method for predicting the spatial distribution of algal blooms in lakes at an hourly scale based on geostationary satellites, the method comprising:

[0007] Acquire wind direction data for the study area and geostationary satellite remote sensing imagery data covering the study area;

[0008] Based on the remote sensing image data, calculate the algal bloom identification index and algal bloom coverage of the study area;

[0009] Using the wind direction data and the algal bloom coverage, the floating algae coverage index, distance index, and surrounding algae index of the image pixels are calculated, respectively, where:

[0010] The algae coverage index is the average algae bloom coverage within an n×n matrix containing the target image pixels, and the position of the target image pixels in the matrix is ​​determined based on wind direction.

[0011] The distance index represents the distance between a target image pixel and the nearest land pixel along the wind direction;

[0012] The surrounding algae index is the value of the center pixel of the matrix, which is characterized by the standard deviation of the algal bloom coverage of the pixels in the matrix.

[0013] A machine learning model is constructed that includes the algal bloom identification index, algal bloom coverage, floating algae coverage index, distance index, and surrounding algae index as input features. The machine learning model is trained to obtain an algal bloom prediction model, and the trained algal bloom prediction model is used to predict algal blooms in lakes.

[0014] As a preferred embodiment, the input features of the algal bloom prediction model also include environmental data and remote sensing indices related to algal bloom changes, which are screened using correlation analysis.

[0015] As a preferred embodiment, the input features of the algal bloom prediction model also include hourly environmental data, as well as land surface temperature, water depth, and algal bloom frequency data;

[0016] The environmental data includes wind speed, temperature, and solar radiation data.

[0017] Furthermore, in the input features, the environmental data is taken as the average value of the current time and the next time.

[0018] In a preferred embodiment, the position of the target image pixel in the matrix is ​​determined based on wind direction, including:

[0019] The wind direction is divided into eight zones, with each zone centered on the directions of due east, due west, due south, due north, northeast, northwest, southeast, and southwest.

[0020] Obtain the average wind direction of the target pixel at the current time and the next time, and determine the wind direction of the target pixel based on the wind direction interval in which the average wind direction is located;

[0021] The target pixel is located downwind of the center pixel in the matrix.

[0022] In a preferred embodiment, the matrix does not contain land pixels or pixels removed due to masking.

[0023] As a preferred implementation, the distance index is characterized based on a normalized index.

[0024] As a preferred implementation, during model training, images with wind speeds below a critical threshold are selected as training data; the critical threshold is the wind speed threshold for algal blooms.

[0025] In one preferred embodiment, the geostationary satellite remote sensing image data includes geostationary satellite remote sensing image data from different sources;

[0026] The geostationary satellite remote sensing image data from different sources, after passing consistency verification, are used to jointly construct the prediction model.

[0027] Furthermore, the method also includes, for geostationary satellite remote sensing image data from different sources that exhibit a linear relationship, selecting geostationary satellite remote sensing image data from one source as standard data, and then using the linear relationship to correct the geostationary satellite remote sensing image data from other sources before jointly constructing a prediction model.

[0028] In a preferred embodiment, the machine learning model is a random forest model. In this invention, the newly constructed metric has high feature importance in the random forest model, and the trained prediction model also has higher accuracy.

[0029] This invention, based on the wind-borne migration characteristics of algal blooms, constructs characteristic indicators reflecting algal bloom migration under different conditions using algal bloom indices and wind direction. These include a wind-direction-related floating algae coverage index, a distance index considering the ease of algal bloom formation along shorelines and incorporating wind direction and distance from shore, and a surrounding algae index related to changes in surrounding algal blooms. These newly constructed indicators are used as input features to train a prediction model, demonstrating good predictive performance for short-term spatial prediction of algal blooms. The data-driven algal bloom prediction model constructed using this invention can efficiently predict the location and intensity of lake algal blooms several hours in advance, making it an important technology for assisting in the management and control of eutrophic lakes and algal blooms, with significant scientific and practical value. Attached Figure Description

[0030] The accompanying drawings are provided to clearly and intuitively illustrate the process from model construction to model application, helping readers better understand this invention patent, improving the information delivery and readability of the document, and explaining the components and symbols in the drawings. Now, the various steps of the invention will be described through examples and with reference to the accompanying drawings, wherein:

[0031] Figure 1 It is the annual average statistical result of FAC data calculated from each GOCI image.

[0032] Figure 2 These are schematic diagrams of FACI and DI calculations. The left diagram is a schematic diagram of FACI calculation, and the right diagram is a schematic diagram of DI calculation.

[0033] Figure 3 It is the ranking result of the importance of input features in the random forest model and the deep neural network model.

[0034] Figure 4 These are the model validation results of the random forest model. (a) is the FAC data at the initial time step, (c) and (d) are the true and predicted FAC values ​​at the next time step, and (b) is the scatter validation after the synthesis of the 3*3 mean matrix.

[0035] Figure 5 This is a pixel-level spatial prediction accuracy demonstration of Taihu Lake algal blooms.

[0036] Figure 6 This is the prediction accuracy of the FAC lake mean. (a), (b), and (c) show the FAC mean after 1, 4, and 7 hours, respectively. Detailed Implementation

[0037] To more effectively illustrate the technical solution of the present invention, specific embodiments are provided and illustrated with accompanying drawings.

[0038] In this disclosure, embodiments of the invention will be described with reference to the accompanying drawings. However, the embodiments and drawings do not encompass all instances in which the invention is applicable. This is because the technical solutions and implementation processes described above, as well as the more detailed concepts and specific steps described below, can be flexibly implemented in various ways, and the technical content disclosed in this invention is not limited to any specific implementation. Furthermore, depending on the specific characteristics of the lake, some parts of the technical process can be added or removed to build specific prediction models for different lakes.

[0039] The environmental data used in this embodiment comes from the following sources:

[0040] Download hourly wind speed, wind direction, temperature and solar radiation data for the study area from 2011 to 2024 from the European Centre for Medium-Range Weather Forecasts (ECMWF, https: / / cds.climate.copernicus.eu / ).

[0041] The land surface temperature dataset (MODIS_LST, https: / / doi.org / 10.5067 / MODIS / MOD11A1.061) from 2011 to 2024 was downloaded from the Google Earth Engine (GEE) platform. The dataset has a spatial resolution of 1000 m and a temporal resolution of 1 day. Based on the MODIS Aqua dataset (MODIS / 061 / MYD09GA) from the GEE platform, the spatial resolution was resampled to 250 m. Using FAC≥5 as the criterion for algal blooms, the monthly algal bloom frequency (MODIS_Fre) at each pixel location was calculated.

[0042] The water depth data is calculated by subtracting the lake bottom elevation from the average water level.

[0043] The geostationary satellite remote sensing data used in the embodiments were obtained from GOCI data from May 2011 to March 2021 and GOCI-II data from January 2021 to October 2024, downloaded from the website (http: / / kosc.kiost.ac.kr / eng / ).

[0044] GOCI and GOCI-II have spatial resolutions of 500 meters and 250 meters, respectively, and a temporal resolution of 1 hour, providing 8 to 10 images per day. The GOCI data were processed using SeaDAS (7.5.3) software to remove water vapor, ozone absorption, and Rayleigh scattering signals, and the Rayleigh reflectance R in each band was calculated. rc For GOCI-II, R rc Data products are processed using G2AC (GOCI-II atmospheric correction). Red, green, and blue bands are used. rc Threshold combinations of values ​​to exclude clouds and solar flares (R) rc (490)>0.14 & R rc (555)>0.16 & R rc (660)>0.15) and cloud shadow (R rc (555)<0.07) interference, selected images with better quality and less cloud cover, and finally obtained 6731 images. Among them, images with wind speed below the critical threshold (3 m / s) and more obvious algal bloom changes were used as sample data for model training and validation, totaling 365 images.

[0045] Example 1

[0046] This embodiment illustrates the present invention's method for predicting the spatial distribution of algal blooms in lakes on an hourly scale based on geostationary satellites.

[0047] This embodiment predicts algal blooms in Taihu Lake based on machine learning and GOCI observations, as follows:

[0048] Using Taihu Lake as the study area, historical monitoring data such as wind speed, wind direction, and temperature were collected. Algal blooms were extracted using high temporal resolution GOCI data. New feature indices were constructed based on wind direction and remote sensing indices to outline typical algal bloom samples and generate a database for training a random forest model. After verifying the model's accuracy, the trained model was used for spatial prediction of algal blooms in Taihu Lake.

[0049] The implementation process of this invention will be illustrated step by step below with accompanying drawings.

[0050] 1) Calculate the FAI or AFAI index based on the geostationary satellite imagery, and then use the FAI or AFAI index as the independent variable to calculate the FAC value. The FAC index is the algal bloom coverage, which can intuitively represent the proportion of algal bloom in a single pixel.

[0051] In this embodiment, since the GOCI satellite lacks a shortwave infrared band, the AFAI index is used instead of the FAI index. The calculation formulas for the AFAI and FAI indices are as follows:

[0052] (1)

[0053] (2)

[0054] In the formula, λ1=660 nm, λ2=745 nm, λ3=865 nm, R rc This represents the Rayleigh reflectance for the corresponding waveband. Figure 1 The annual average statistical results of FAC data calculated for each GOCI image.

[0055] The GOCI and GOCI-II data have been verified to have a high degree of consistency. As a preferred approach, based on the linear relationship between GOCIAFAI and GOCI-II AFAI, the AFAI value calculated based on GOCI-II data has been linearly adjusted, and the adjusted GOCI-II AFAI is used for subsequent calculations.

[0056] 2) Constructing new feature indicators:

[0057] Wind speed and direction are key factors influencing the horizontal and vertical movement of algal blooms. Wind-induced disturbances drive cyanobacteria to migrate and accumulate in windward directions. Driven by wind, algae are more likely to accumulate near the shore, leading to algal blooms. Therefore, based on the above prior knowledge, this invention constructs a new feature index that combines remote sensing indicators with environmental factors for modeling.

[0058] In this embodiment, three new characteristic indices are constructed using wind direction and the FAC index: the Floating Algae Cover Index (FACI), the Distance Index (DI), and the Algae Around Index (AAI). Their calculation methods are as follows:

[0059] ① Floating Algae Coverage Index (FACI);

[0060] The wind direction range of 0~360° is divided into 8 equally spaced intervals, as shown in Table 1:

[0061] Table 1 Wind Direction Zones

[0062]

[0063] Based on the average wind direction at the current and next time moments, a corresponding statistical region is selected, and the average FAC in the 3×3 matrix is ​​obtained and assigned to the target cell. For example... Figure 2 As shown, Figure 2 (Left) The average wind direction is northwest. Algal blooms tend to migrate from northwest to southeast under the influence of wind. Therefore, the corresponding statistical range is a 3x3 matrix of the upper left corner of the target pixel (i.e., the target pixel is located in the lower right corner, or southeast corner). The formula for calculating the algal bloom coverage index is as follows:

[0064] (3)

[0065] In the formula, n is the number of effective pixels contained in the 3×3 matrix corresponding to the target pixel. Let f be the FAC value of the i-th pixel. The matrix does not include land pixels or pixels that are masked.

[0066] ② Distance index DI;

[0067] The DI value represents the distance between a target pixel and the land. The determination of this distance in the DI value is related to wind direction (hereinafter referred to as wind-direction land distance). The wind-direction land distance of a target pixel is represented by the distance between the target pixel and the nearest land pixel in the wind direction, using a normalized value in the formula. For example... Figure 2 As shown, Figure 2 (Right) The average wind direction is northeast. The distance between the target pixel and the nearest land pixel in the wind direction is the same as the distance between the target pixel and the nearest land pixel in the due northeast direction. The formula for calculating DI is as follows:

[0068] (4)

[0069] In the formula, Distance iDistance is the distance between target pixel i in the current image and the nearest land pixel along the wind direction. max Distance of all target pixels calculated for the current image i The maximum value among the values.

[0070] ③ Surrounding algae index (AAI);

[0071] Calculate the standard deviation of the FAC within the 3×3 matrix to which the target pixel belongs in 1), and use it as the AAI value of the center pixel of that matrix, as follows:

[0072] (5)

[0073] 3) The model is trained using environmental factors, AFAI index, AFAI gradient (AFAI_Fradient), FAC index and FACI, DI and AAI feature indices constructed in 2) as input features and algal bloom index as output feature. In this embodiment, the algal bloom index is selected as FAC index.

[0074] In this embodiment, two machine learning models, Random Forest (RF) and Deep Neural Network (DNN), which can handle high-dimensional data and capture complex nonlinear relationships, are selected for spatial prediction of cyanobacterial blooms in Taihu Lake.

[0075] The input environmental factors are wind speed, air temperature, solar radiation data, water depth, algal bloom frequency (MODIS_Fre), and land surface temperature (MODIS_LST). Among them, wind speed, air temperature, and solar radiation are the average values ​​of the current time and the next time, water depth data are kept constant, algal bloom frequency is the algal bloom frequency of the previous month, and land surface temperature is the value of the current time.

[0076] Typical algal bloom samples and randomly distributed water samples within the lake area were manually sketched. Feature data from the same locations were extracted to construct a sample database. The dataset was divided into training and test sets in a 9:1 ratio for model training and validation. The random forest model was trained using the Scikit-Learn package in Python 3.7, and the DNN model was constructed by adding input, hidden, and output layers sequentially using the Keras library. R was then used... 2 The accuracy of the model was evaluated using RMSE, PRD, MAPE, and UPD, and the results are shown in Table 2.

[0077] Table 2 Performance evaluation of RF and DNN models used for spatial prediction of algal blooms

[0078]

[0079] It can be seen that the RF model has better predictive performance. Figure 3The importance contributions of 12 input features in model training are shown. Among them, the FAC and AFAI indices, used as indicators for algal bloom identification, predictably rank first and second in importance in both models. In the feature importance proportion of the RF model, FACI, DI, and AAI rank 4th, 5th, and 7th respectively, contributing 27% to the RF model, effectively assisting in pixel-level algal bloom spatial prediction in Taihu Lake. In the feature importance proportion of the DNN model, FACI, DI, and AAI rank 5th, 8th, and 12th respectively, contributing 14.4% to the DNN model. As shown in Table 1, the RF model, with a higher contribution from the newly constructed indices, exhibits better predictive performance than the DNN model, demonstrating the effectiveness of the newly constructed indices.

[0080] 4) Apply the trained RF model to specific images to verify its accuracy. The image from September 18, 2017 was selected. Figure 4 (a) is the FAC truth data at the current moment. Figure 4 (c) and (d) represent the true and predicted values ​​of FAC at the next time step, respectively. Figure 4 (b) is a scatter plot of the true and predicted values.

[0081] 5) Apply the trained model to GOCI and GOCI-II images, and predict the spatial location and outbreak intensity of Taihu Lake algal blooms hourly through iterative model calculations. Figure 5 For pixel-level spatial prediction of Taihu Lake algal blooms, Figure 6 This represents the prediction accuracy of the overall FAC mean.

Claims

1. A method for predicting the spatial distribution of algal blooms in lakes at an hourly scale based on geostationary satellites, characterized in that, The method includes: Acquire wind direction data for the study area and geostationary satellite remote sensing imagery data covering the study area; Based on the remote sensing image data, calculate the algal bloom identification index and algal bloom coverage of the study area; Using the wind direction data and the algal bloom coverage, the floating algae coverage index, distance index, and surrounding algae index of the image pixels are calculated, respectively, where: The algae coverage index is the index including target image pixels. n × n The mean algal bloom coverage within the matrix range, wherein the position of the target image pixel in the matrix is ​​determined based on wind direction, including: The wind direction is divided into eight zones, with each zone centered on the directions of due east, due west, due south, due north, northeast, northwest, southeast, and southwest. Obtain the average wind direction of the target pixel at the current time and the next time, and determine the wind direction of the target pixel based on the wind direction interval in which the average wind direction is located; The target pixel is located downwind of the center pixel in the matrix; The distance index represents the distance between a target image pixel and the nearest land pixel along the wind direction; The surrounding algae index is the value of the center pixel of the matrix, which is characterized by the standard deviation of the algal bloom coverage of the pixels in the matrix. A machine learning model is constructed that includes the algal bloom identification index, algal bloom coverage, floating algae coverage index, distance index, and surrounding algae index as input features. The machine learning model is trained to obtain an algal bloom prediction model, and the trained algal bloom prediction model is used to predict algal blooms in lakes.

2. The method according to claim 1, characterized in that, The input features of the algal bloom prediction model also include environmental data and remote sensing indices related to algal bloom changes, which are screened using correlation analysis.

3. The method according to claim 1, characterized in that, The input features of the algal bloom prediction model also include hourly environmental data, as well as land surface temperature, water depth, and algal bloom frequency data. The environmental data includes wind speed, temperature, and solar radiation data.

4. The method according to claim 1, characterized in that, The matrix does not contain land pixels or pixels removed due to masking.

5. The method according to claim 1, characterized in that, The distance index is characterized based on a normalized index.

6. The method according to claim 1, characterized in that, During model training, images with wind speeds below a critical threshold are selected as training data; the critical threshold is the wind speed threshold for algal blooms.

7. The method according to claim 1, characterized in that, The geostationary satellite remote sensing image data includes geostationary satellite remote sensing image data from different sources; The geostationary satellite remote sensing image data from different sources, after passing consistency verification, are used to jointly construct the prediction model.

8. The method according to claim 7, characterized in that, The method further includes, for geostationary satellite remote sensing image data from different sources that exhibit a linear relationship, selecting geostationary satellite remote sensing image data from one source as standard data, and then using the linear relationship to correct the geostationary satellite remote sensing image data from other sources before jointly constructing a prediction model.

9. The method according to claim 1, characterized in that, The machine learning model is a random forest model.

Images