A method and system for extracting a lake phytoplankton phenology parameter based on multi-scale parameter migration and a dynamic threshold

Abstract

The present application relates to the technical field of data extraction, and discloses a lake phytoplankton phenology parameter extraction method and system based on multi-scale parameter migration and dynamic threshold, which comprises the following steps: obtaining remote sensing image data of a target area, and constructing multi-scale time sequence image data; dividing the growth cycle of plants, constructing a growth curve benchmark model and a parameter mapping model, to obtain the phenology morphological parameters of plants in the growth cycle, and identifying and extracting key phenology parameters of the growth cycle through a preset dynamic threshold. The present application improves the fitting stability of high-noise sparse time sequences, establishes a fitting parameter mapping model, avoids directly fitting high-noise data, and improves the robustness of the model; realizes dynamic adaptive adjustment of the phenology threshold, so that the threshold changes adaptively with the intensity of algal blooms, and improves the adaptability to different nutrient state lakes and different years; and supports lake algal bloom phenology monitoring in different climate zones, different nutrient levels and long time sequences.

Description

Technical Field

[0001] This invention relates to the field of data extraction technology, and in particular to a method and system for extracting phenological parameters of lake phytoplankton based on multi-scale parameter migration and dynamic threshold. Background Technology

[0002] Existing research has transplanted various time-series smoothing (e.g., filtering, fitting) and phenological parameter extraction methods (e.g., thresholding, rate of change, accumulation) from marine phenology. Different methods and threshold choices lead to significant differences in the extracted phenological parameters, and these methods were initially designed for open oceans with relatively uniform optical characteristics or terrestrial vegetation with relatively stable growth cycles. Phytoplankton, especially in eutrophic lakes, are characterized by short lifecycles and rapid environmental responses; their biomass (represented by remote sensing signals) changes more rapidly and fluctuates more dramatically, resulting in time-series signals with significantly higher noise levels compared to marine phytoplankton or surface vegetation. Therefore, directly applying existing models without adapting them to the rapid changes in lake phytoplankton and the high-noise data environment is the main reason for inaccurate and poorly universal phenological extraction results. Summary of the Invention

[0003] In view of the aforementioned existing problems, this invention is proposed. Therefore, this invention provides a method for extracting lake phytoplankton phenological parameters based on multi-scale parameter migration and dynamic thresholding. This addresses the challenge of effectively overcoming high noise interference in inland waters and accurately capturing the rapid changes in phytoplankton characteristics in an automated extraction method, thus overcoming the shortcomings of existing technologies in terms of accuracy, reliability, and automation of phenological parameter extraction.

[0004] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for extracting phenological parameters of lake phytoplankton based on multi-scale parameter migration and dynamic thresholding, comprising: Acquire remote sensing image data of the target area, and preprocess the remote sensing image data to construct multi-scale time series image data; Based on the multi-scale time series image data, the plant growth cycle is divided to identify the growth cycle, and a growth curve benchmark model and parameter mapping model are constructed to obtain the phenological morphological parameters of the plant during the growth cycle. The original pixel sequence of the image data is segmented into pixel-level growth cycles, and based on the trained parameter mapping model, key phenological parameters of the growth cycle are identified and extracted through a preset dynamic threshold.

[0005] As a preferred embodiment of the lake phytoplankton phenological parameter extraction method based on multi-scale parameter migration and dynamic thresholding described in this invention, the method includes: acquiring remote sensing image data of the target area and preprocessing the remote sensing image data, including: Collect raw remote sensing image data of the target area at different spatial resolutions, and extract the algae coverage features from the raw remote sensing image data; Global surface water data is collected, and a region is selected from the global surface water data as a reference water body mask. A dynamic water body mask algorithm is constructed using multispectral index and reflectance threshold to remove non-water body pixels and interfering water body pixels from the remote sensing image data, thereby obtaining the effective water body pixels in the remote sensing image data. The daily algae coverage index is calculated for the effective water body pixels, and multi-scale aggregation is performed on remote sensing image data with different spatial resolutions to obtain the algae coverage time series.

[0006] As a preferred embodiment of the lake phytoplankton phenological parameter extraction method based on multi-scale parameter migration and dynamic thresholding described in this invention, the method involves: multi-scale aggregation of remote sensing image data at different spatial resolutions to obtain the time series of algal bloom coverage, including: The original remote sensing image data with different spatial resolutions are divided into high-resolution and low-resolution image data. After extracting the algae coverage features, high-resolution original algae coverage time series and low-resolution algae coverage time series are obtained. The low-resolution image data is divided into several grid cells of the same size. The daily algae coverage value is determined based on the proportion of effective pixels in the grid cells, thus obtaining the low-resolution algae coverage time series. The low-resolution algae coverage time series was reconstructed by sequence difference to obtain the complete low-resolution algae coverage time series.

[0007] As a preferred embodiment of the lake phytoplankton phenological parameter extraction method based on multi-scale parameter migration and dynamic thresholding described in this invention, the method for obtaining phenological morphological parameters of plants during their growth cycle includes: Based on the algae coverage index, the growth cycle of the low-resolution algae coverage time series is segmented to obtain multiple independent growth cycles. Based on the independent growth cycle, curve fitting is performed on the low-resolution algae coverage time series of the independent growth cycle to obtain the fitting parameter set for each growth cycle. A parameter mapping model is constructed, and the model is trained using the parameter set of the fitting function to establish a mapping relationship between low-resolution data and high-resolution data.

[0008] As a preferred embodiment of the lake phytoplankton phenological parameter extraction method based on multi-scale parameter migration and dynamic thresholding described in this invention, the parameter mapping model includes: Based on the high-resolution original algal bloom coverage time series of the target area within each growth cycle, multiple statistical features of each pixel in the series are calculated. A parameter mapping model is constructed using a machine learning model. The high-resolution statistical features are used as the model input, and the low-resolution sequence of the fitted parameter set is used as the model output. The parameter mapping model is trained to establish the mapping relationship between the multiple statistical features and the fitted parameter set, so as to obtain the prediction fitting parameters of the high-resolution original algal cover time series.

[0009] As a preferred embodiment of the lake phytoplankton phenological parameter extraction method based on multi-scale parameter migration and dynamic thresholding described in this invention, the key phenological parameters for the growth cycle are identified and extracted as follows: By using a sliding window, the phenological segmentation index of the high-resolution original algal cover time series is calculated to divide the growth cycle. Based on each growth cycle segmented by each high-resolution pixel, the statistical features are extracted and input into the parameter mapping model to predict the fitting parameter set for each growth cycle of the pixel, thereby obtaining the fitting time series of the cycle. Based on the maximum amplitude of each growth cycle, the optimal phenological threshold is dynamically calculated, and the time data of plant growth are obtained from the fitted time series.

[0010] As a preferred embodiment of the lake phytoplankton phenological parameter extraction method based on multi-scale parameter migration and dynamic thresholding described in this invention, the method includes: calculating the phenological segmentation index of the high-resolution original algal cover time series and dividing the growth cycle, including: The high-resolution original algal bloom coverage time series for each year are aligned by year-to-day. The mean, standard deviation, and maximum value of algae coverage at each location on each day of the year are calculated using a preset sliding window and then fused to obtain the phenological segmentation index. The day corresponding to the minimum value of the phenological segmentation index is selected as the annual growth cycle segmentation point of the pixel. That is, the annual growth cycle of the pixel starts from the day of the year and ends on the day of the year of the following year.

[0011] As a preferred embodiment of the lake phytoplankton phenological parameter extraction method based on multi-scale parameter migration and dynamic thresholding described in this invention, the method includes: dynamically calculating the optimal phenological determination threshold and obtaining plant growth time data from the fitted time series, including: The maximum amplitude determined by the set of fitting parameters is used to construct a functional relationship between the maximum amplitude and the optimal active amplitude threshold. The optimal function form and function parameters are determined by the least squares method, and the optimal active amplitude threshold is calculated. Based on the optimal activity amplitude threshold, from the fitted time series, the date on which the algal bloom coverage value first rises above the threshold from the left baseline during the growth period is selected as the growth start date, and the date on which the algal bloom coverage value last falls to the threshold during the decline period is selected as the growth end date.

[0012] Secondly, the present invention provides a system for extracting phenological parameters of lake phytoplankton based on multi-scale parameter migration and dynamic thresholding, comprising: The acquisition and processing module is used to acquire remote sensing image data of the target area and preprocess the remote sensing image data to construct multi-scale time series image data. The mapping module is used to divide the plant growth cycle based on the multi-scale time series image data, so as to identify the growth cycle, construct a growth curve benchmark model and a parameter mapping model, and obtain the phenological morphological parameters of the plant during the growth cycle. The output module is used to perform pixel-level growth cycle segmentation on the original pixel sequence of the image data, and based on the trained parameter mapping model, identify and extract key phenological parameters of the growth cycle through a preset dynamic threshold.

[0013] Thirdly, the present invention provides a computer device, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of the method for extracting lake phytoplankton phenological parameters based on multi-scale parameter migration and dynamic threshold.

[0014] Fourthly, the present invention provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the method for extracting lake phytoplankton phenological parameters based on multi-scale parameter migration and dynamic thresholds.

[0015] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention improves the fitting stability of high-noise sparse time series by fusing a relatively complete time series at the 5km scale with 250m pixel-level time series statistical features to establish a fitting parameter mapping model, avoiding direct fitting of high-noise data and improving model robustness; it achieves dynamic adaptive adjustment of phenological thresholds by establishing a functional relationship between the maximum amplitude and the optimal threshold, allowing the threshold to adapt to changes in algal bloom intensity, thus improving adaptability to lakes with different trophic states and different years; it unifies the uncertainty assessment standard for phenological parameters by introducing spatial variation coefficients and spatial autocorrelation indicators to quantify the spatial consistency of model results and enhance the comparability of results from different regions and methods; it supports cross-regional, long-term lake algal bloom phenological monitoring, applicable to typical lakes in different climate zones and with different trophic levels, providing reliable technical support for early warning and ecological management of algal blooms. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of the overall process of a method for extracting phenological parameters of lake phytoplankton based on multi-scale parameter migration and dynamic threshold, according to an embodiment of the present invention.

[0018] Figure 2 This is a schematic diagram illustrating the specific process of a method for extracting phenological parameters of lake phytoplankton based on multi-scale parameter migration and dynamic threshold, as described in one embodiment of the present invention.

[0019] Figure 3 This is a comparative experimental data chart of a method for extracting phenological parameters of lake phytoplankton based on multi-scale parameter migration and dynamic threshold, as described in one embodiment of the present invention. Detailed Implementation

[0020] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0021] Example 1, referring to Figures 1-2Table 1, as an embodiment of the present invention, provides a method for extracting lake phytoplankton phenological parameters based on multi-scale parameter migration and dynamic thresholding, comprising: S1. Acquire remote sensing image data of the target area, and preprocess the remote sensing image data to construct multi-scale time series image data; S2, based on multi-scale time series image data, divide the plant growth cycle to identify the growth cycle, construct a growth curve benchmark model and a parameter mapping model to obtain the phenological morphological parameters of the plant during the growth cycle. S3 performs pixel-level growth cycle segmentation on the original pixel sequence of the image data, and based on the trained parameter mapping model, identifies and extracts key phenological parameters of the growth cycle through a preset dynamic threshold.

[0022] It should be noted that this invention improves the fitting stability of high-noise sparse time series by fusing relatively complete time series at the 5km scale with 250m pixel-level time series statistical features to establish a fitting parameter mapping model, avoiding direct fitting of high-noise data and improving model robustness. It also achieves dynamic adaptive adjustment of phenological thresholds by establishing a functional relationship between the maximum amplitude and the optimal threshold, allowing the threshold to adapt to changes in algal bloom intensity, thus improving adaptability to lakes with different trophic states and different years. Furthermore, it unifies the uncertainty assessment criteria for phenological parameters by introducing spatial variation coefficients and spatial autocorrelation indicators to quantify the spatial consistency of model results and enhance the comparability of results from different regions and methods. Finally, it supports cross-regional, long-term lake algal bloom phenological monitoring, applicable to typical lakes in different climate zones and with different trophic levels, providing reliable technical support for early warning and ecological management of algal blooms.

[0023] Preferably, in step S1, remote sensing image data of the target area is acquired, and the remote sensing image data is preprocessed, including: Collect raw remote sensing image data of the target area at different spatial resolutions and extract the algae coverage features from the raw remote sensing image data; Global surface water data is collected, and regions are selected from the global surface water data as reference water body masks. A dynamic water body mask algorithm is constructed using multispectral index and reflectance threshold to remove non-water body pixels and interfering water body pixels in remote sensing image data, thereby obtaining effective water body pixels in remote sensing image data. The daily algae coverage index was calculated for effective water body pixels, and multi-scale aggregation of remote sensing image data with different spatial resolutions was performed to obtain the time series of algae coverage.

[0024] This embodiment uses the floating algae cover (FAC) index calculated from MODIS Aqua satellite data as an example. High- and low-resolution image data are collected: a high-resolution 250m raw pixel sequence and a low-resolution 5km but relatively complete and smooth grid cell sequence. Specifically, MODIS Aqua MYD09GQ (250m resolution, including red and near-infrared bands) and MYD09GA (500m resolution, including blue, green, and shortwave infrared bands) data from 2003 to 2024 for the study area (such as Taihu Lake, Chaohu Lake, Dianchi Lake, and Hulun Lake) are obtained from the Google Earth Engine platform. The 500m band is resampled to 250m to spatially align the 500m and 250m data. Combined with JRC global surface water data, a dynamic water body mask is constructed using spectral indices and reflectance thresholds to eliminate non-water body pixels and outliers.

[0025] The MODIS MYD09GQ dataset provides red and near-infrared reflectance at a resolution of 250 meters for core calculations. The MODIS MYD09GA dataset provides blue, green, shortwave infrared 1, and shortwave infrared 2 reflectance at a resolution of 500 meters. The data is bilinearly resampled to a resolution of 250 meters to unify the spatial scale.

[0026] Furthermore, the algae cover characteristics are extracted from the original remote sensing image data, and the daily FAC value is calculated, expressed as: ; in, This indicates the reflectance value in the near-infrared band, with a center wavelength of 859 nm. This indicates the reflectivity value in the red light band, with a center wavelength of 645nm; This represents the reflectance value of the shortwave infrared band 1, with a center wavelength of 1240nm.

[0027] Using the occurrence layer in the JRC Global Surface Water Dataset (GSW1_0), areas with an annual occurrence frequency of surface water ≥90% were selected as baseline water body masks, ensuring that the analysis area is limited to stable water bodies. Multispectral indices include turbidity index, cloud and flare mask, shadow, and aquatic vegetation suppression index.

[0028] Among them, the turbidity index After removing pixels from the water body that are turbid due to high suspended sediment concentration, the result is represented as: ; in, This represents the reflectivity value in the red light band, with a center wavelength of 645nm. This represents the reflectance value in the shortwave infrared band 1, with a center wavelength of 1240 nm. The reflectance of turbid water in the red light band increases significantly due to backscattering by suspended particles, sometimes exceeding its reflectance in the shortwave infrared band. Clean water should meet the following requirements. Therefore, turbid water pixels with a turbidity index less than 0.09 are removed.

[0029] Clouds and solar flares exhibit high reflectivity across all visible light wavelengths. By using cloud and flare masks to remove pixels affected by thick clouds and solar flares, pixels that simultaneously meet the following conditions are identified as either clouds or flares and removed: , , ,in, This indicates the reflectivity in the green light band, with a center wavelength of 555nm. This indicates the reflectivity in the blue light band, with a center wavelength of 469nm.

[0030] By removing dark pixels such as mountain shadows and cloud shadows, as well as areas with low reflectivity and no signal, pixels that simultaneously meet the following conditions are identified as shadows and removed: , , .

[0031] The inhibition index of aquatic vegetation The disturbance caused by the suppression of aquatic vegetation in wetlands and shallow water areas is represented as: ; This index constructs a baseline by introducing shortwave infrared bands, enhancing the green peak signal of green vegetation. The vegetation coverage area is determined by CMI>0.02, and interfering pixels are removed.

[0032] It should be noted that by constructing a dynamic and high-precision water body mask, this invention can effectively remove non-water body pixels and interfering water body pixels with complex optical characteristics from remote sensing images, providing a clean and reliable water body pixel dataset for subsequent extraction of phytoplankton phenological parameters.

[0033] Preferably, in step S1, multi-scale aggregation of remote sensing image data with different spatial resolutions is performed to obtain the time series of algae coverage, including: The original remote sensing image data with different spatial resolutions are divided into high-resolution and low-resolution image data. After extracting the algae coverage features, the high-resolution original algae coverage time series and the low-resolution algae coverage time series are obtained. In this process, low-resolution image data is divided into several grid cells of the same size. Based on the proportion of effective pixels in the grid cells, the daily algae coverage value is determined, and a low-resolution algae coverage time series is obtained. The low-resolution algae coverage time series was reconstructed by sequence difference to obtain the complete low-resolution algae coverage time series.

[0034] In this embodiment, the high-resolution original algae coverage time series is the original FAC time series of 250-meter pixels; the low-resolution refers to the collected 5km image data. The lake is divided into multiple units using a grid with a side length of 5km. If the percentage of effective pixels in each unit is ≥20% on a certain day, the average FAC of these pixels is taken as the FAC value of that unit on that day; otherwise, the FAC value is marked as missing.

[0035] Specifically, the boundaries of the four lakes were divided into several 5*5km grid units using a fishing net with a side length of 5km. Due to the irregular shape of the lakes, some areas were less than 5km but were still considered as independent units. Assuming that a unit contains a total of n pixels in the image on a certain day, of which m are non-masked pixels, when m≥n / 5, the average FAC of the non-masked pixels is calculated as the FAC value of that unit on that day; if m<n / 5, then the FAC of that unit on that day is missing. The FAC time series of each independent unit is established according to the above method.

[0036] Furthermore, 5km sequence interpolation reconstruction is performed, and feature selection is conducted, including meteorological factors, historical FAC dynamic features, and annual DOY (Dollar-of-Year) data. Meteorological factors can include daily average temperature, daily maximum temperature, daily / monthly maximum precipitation, wind speed in the u / v direction, daily average atmospheric pressure, and solar radiation. Historical FAC dynamic features can be the average and maximum FAC values ​​within a 51-day sliding window centered on the current day. The XGBoost algorithm is used to divide the effective FAC observation data and corresponding features of each 5km unit into training and testing sets at a 4:1 ratio, and hyperparameter optimization is performed using GridSearchCV. For sparse data units, a spatial proximity strategy is adopted, merging the data of four adjacent units for joint training. The trained model is then used to process the FAC time series interpolation of the target unit, ultimately obtaining a complete daily FAC sequence.

[0037] Specifically, the raw daily FAC observation data of the 5km image data and its corresponding multi-dimensional feature data, such as water optical parameters and meteorological conditions, are first read and aligned by date to generate a complete daily timeline within the collection period, ensuring that the subsequent interpolation results are continuous and uninterrupted in time.

[0038] Furthermore, key temporal features are created, including sliding window mean features and sliding window peak features. The sliding window mean feature calculates the average of all valid high FAC observations within a 25-day window before and after each date to capture the short-term background state of algal blooms. The sliding window peak feature calculates the average of the top three FAC values ​​within a 25-day window before and after each date to characterize the potential intensity of local algal blooms. Missing values ​​in the calculated temporal features are filled using linear interpolation to ensure the integrity of the feature matrix.

[0039] For a specific 5km grid cell with insufficient effective observation data from FAC, a spatial proximity strategy is used to process the data. This strategy merges the data of the target cell with its four directly adjacent cells to the east, south, west, and north, generating an expanded and richer training dataset.

[0040] Using the merged dataset, an Extreme Gradient Boosting Tree (XGBoost) regression model is trained. This trained model is then used to interpolate all feature data (including fully interpolated temporal features) of the target unit, yielding the daily FAC interpolation result for that unit. RandomizedSearchCV is used to perform a random search within a predefined parameter space, combined with K-fold cross-validation (K=4) to evaluate parameter performance. Key hyperparameters optimized include the number of trees, maximum tree depth, learning rate, subsampling ratio, feature sampling ratio, and regularization parameters. Multiple rounds of training and validation are conducted, and the optimal hyperparameter combination performing best on the cross-validation set is selected to construct the final model.

[0041] It should be noted that this step constructed a model for chlorophyll a concentration inversion that was not universally applicable to inland waters, replacing the FAC (algal bloom coverage) model. This provided more reliable time series data for subsequent phenological analysis. Furthermore, the time series reconstruction and noise filtering process was optimized to take into account the rapid changes in phytoplankton and the characteristics of MODIS data, effectively improving the quality and usability of the time series data.

[0042] Preferably, in step S2, obtaining the phenological parameters of the plant during its growth cycle includes: Based on the algae coverage index, the growth cycle of the low-resolution algae coverage time series is segmented to obtain multiple independent growth cycles. Based on independent growth cycles, curve fitting was performed on the time series of low-resolution algal cover for independent growth cycles to obtain several optimal fitting parameters. Construct a parameter mapping model, train the model using the optimal fitting parameters, and establish a mapping relationship between low-resolution and high-resolution data.

[0043] In this embodiment, for the continuous time series interpolated from each 5km grid cell, a sliding window detection method is used with a window size of 60 days and a step size of 1 day to identify the lowest FAC value trough between adjacent years, dividing the long-term series into multiple independent growth cycles. For example, 22 years of data are divided into 22 cycles, each containing a complete ecological process from initiation, development to decline. For each growth cycle, an asymmetric Gaussian function is used to fit the dynamic changes of FAC within each growth cycle, expressed as: ; Where t is the t-th day in the growth cycle; It is the daily algal activity characterization index of FAC after reconstructing the time series; It is the maximum value of FAC during the growth cycle, i.e., the peak amplitude; Indicates the peak position; It is the baseline value for the dormant season preceding the growing season; It is the baseline of the dormant season after the growing season. and Control the steepness of the peak on the left and right sides respectively.

[0044] To ensure the convergence of the fitting process and the robustness of the results, the model parameters were initialized using a phenological pattern-based strategy, with peak amplitude... Initializing the values ​​to the mean of the high percentile range (e.g., the top 10%) of all valid FAC observations within the growth cycle effectively approximates the true peak biomass while avoiding initialization bias caused by isolated extreme outliers; peak position Initialized to the date sequence number (DOS) corresponding to the maximum FAC observation value within this growth cycle, this serves as a benchmark to capture the core time node of algal blooms; left and right baselines , Initialized to statistics (such as median or mean) of all valid FAC observations within a complete phenological window (e.g., 1 to 2 months) before and after the start and end of the growth cycle, respectively, it can robustly estimate the baseline activity level of algae during the inactive growth season; width parameter , The initial value is set as an empirical proportion (such as 1 / 10 to 1 / 5 of the total duration of the growth cycle) that is associated with the total duration of the growth cycle, ensuring that the initial curve shape can adapt to the actual time scale of the phenological process and provide a reasonable starting point for subsequent optimization.

[0045] Furthermore, parameter optimization was performed using the Levenberg-Marquardt algorithm for nonlinear least-squares fitting, minimizing the residual between the fitted curve and the observed values ​​to obtain the optimal set of fitting parameters for each growth cycle at a resolution of 5 km. .

[0046] Preferably, in step S2, constructing the parameter mapping model includes: Based on the high-resolution original algae coverage time series of the target area within each growth cycle, multiple statistical features of each pixel in the series are calculated. A parameter mapping model is constructed using a machine learning model. Statistical features are used as model inputs, and a set of fitted parameters from a low-resolution sequence is used as model outputs. The parameter mapping model is trained to establish a mapping relationship between multiple statistical features and the set of fitted parameters in order to obtain the predicted fitted parameters.

[0047] It should be noted that phytoplankton biomass dynamics exhibit significant transient characteristics, with aggregation-dissipation processes on a small scale often completed within hours. However, due to limitations in the observation capabilities of the MODIS satellite—namely, the daily revisit cycle and limited water penetration depth—observational data at 250 m resolution struggles to accurately capture these rapid changes. Specifically, most pixels only record the low biomass background value after the algal bloom dissipates, while a few pixels occasionally capture algal aggregation peaks, resulting in significant high-value outliers in the data. This "all or nothing" observation mode produces obvious signal discontinuities at the pixel scale, leading to significant noise in the FAC time series constructed based on the raw observational data.

[0048] In this embodiment, step S2, constructing the parameter mapping model, is as follows: Raw FAC observation data for each growth cycle First calculate the weighted center point. As the initial value of the peak parameter, The data is divided into growth stages as boundaries. and recession That is, in The period before that was the growth period, in After that comes the decline. according to Divide into four equal parts in ascending order. according to Divide into four equal parts in descending order to obtain 8 subsets. arrive For each subset, calculate its FAC mean and corresponding... The weighted average yielded 16 statistical characteristics.

[0049] First, a validated set of fitted parameters within a 5km grid cell is extracted as a reliable phenological baseline. Simultaneously, statistical characteristics of the original FAC time series of all 250m pixels within the corresponding grid cell are calculated, including mean, variance, skewness, kurtosis, and quantiles. By establishing a machine learning mapping relationship between the above statistical characteristics and the six fitted parameters, effective downscaling of phenological parameters from 5km to 250m scale is achieved, enabling direct estimation of pixel-level phenological parameters based on the original 250m observation data.

[0050] Through experiments, 16 statistical feature values ​​were finally extracted from the growth cycle, forming 6 feature groups, and 6 parameter mapping models were trained for each group. It is assumed that there is raw data within a certain growth cycle of a pixel. ,in, These are the original FAC observations. This corresponds to the observation time, which is the number of days in the growth cycle.

[0051] Specifically, first calculate the parameters. initial value As the weighted center point of the growth cycle, the life cycle is divided into left and right parts, with the initial value... Represented as: ; in, This represents the i-th observation in the original sequence. The original data is divided into two parts. The original dataset is denoted as the growth period. , The original data set is denoted as According to the former Sort the data from smallest to largest, divide it into four equal parts, and denote them as follows: - The latter according to Sort from largest to smallest, and divide it into four equal parts, denoted as . - Calculate separately middle mean and corresponding The weighted mean of the features, respectively , The calculation formula is as follows: ; ; in, Let represent the k-th set within the growth and decline phases, partitioned according to the quartiles of the FAC size. This represents the number of original observations in the k-th set. This represents the mean of the FAC in the k-th set. This represents the time (DOY) mean of the FAC observations in the k-th set, with the values ​​used as weights.

[0052] In addition, the parameters in the asymmetric Gaussian function and Without specific phenological significance, establishing a mapping relationship is quite difficult; therefore, two dynamic threshold parameters are used. and As an alternative, among which , It satisfies the following formula, expressed as: ; ; in, , This represents the asymmetric Gaussian function values ​​used to fit the FAC interannual curve by substituting the two parameters. and and and The relationship is represented as: ; ; These two parameters represent the t values ​​at which the function F(t) reaches its maximum amplitude of 20% on both sides. 20% is a trade-off between balancing sensitivity and stability, and they respectively replace the parameters. and Control the growth / decay rate.

[0053] Finally, model training is performed. The 16 statistical features mentioned above are combined into 6 groups, which are used as independent variables to train a machine learning model to predict 6 fitting parameters. The machine learning model can be XGBoost, random forest, support vector regression, etc. To improve the interpretability and stability of the model, a dynamic threshold parameter is used. and , that is, the dates when the function value reaches 20% of the maximum amplitude, respectively, and six independent parameter mapping models are constructed.

[0054] In this study, XGBoost, Random Forest, and Support Vector Regression were selected as machine learning algorithms to construct parameter mapping models. Different feature combinations were used as inputs, and prediction model structures for corresponding parameters were established by setting parameters such as the number of decision trees, their depth, and the kernel function. R0 was then calculated. 2 The performance of the model was evaluated using indicators such as mean absolute error, root mean square error, and relative error. The optimal model was selected as the mapping model, and the optimal prediction model with six parameters was constructed. The specific phenological parameters are shown in Table 1.

[0055] Table 1: Parameter Table A FAC peak The mean of FAC in the m-th set, where m = 3, 4, 5, 6 p Peak time The FAC observations in the m-th set are used as the time mean corresponding to the weights, where m = 3, 4, 5, 6. <![CDATA[b L ]]> Left baseline value The mean of FAC in the m-th set, where m = 1, 2, 3, 4 <![CDATA[b R ]]> Right baseline value The mean of FAC in the m-th set, where m = 5, 6, 7, 8 <![CDATA[Q L ]]> Average growth rate The FAC observations in the m-th set are used as the time mean corresponding to the weights, where m = 1, 2, 3, 4. <![CDATA[Q R ]]> Average rate of decline The FAC observations in the m-th set are used as the time mean corresponding to the weights, where m = 5, 6, 7, 8.

[0056] Preferably, in step S3, identifying and extracting key phenological parameters of the growth cycle includes: By using a sliding window, the phenological segmentation index of the high-resolution original algal cover time series is calculated to divide the growth cycle. Based on each growth cycle segmented by each high-resolution pixel, statistical features are extracted and input into the parameter mapping model to predict the fitting parameter set for each growth cycle of that pixel, thus obtaining the fitting time series of that cycle. Based on the maximum amplitude of each growth cycle, the optimal phenological threshold is dynamically calculated, and the time data of plant growth are obtained from the fitted time series.

[0057] Preferably, the phenological segmentation index of the high-resolution original algal cover time series is calculated to divide the growth cycle, including: The high-resolution original time series of algal bloom coverage for each year is aligned by the day-in-year sequence. Using a preset sliding window, the mean, standard deviation, and maximum value of algae coverage at each location on each day of the year are calculated and weighted to obtain the phenological segmentation index. The annual day corresponding to the minimum value of the phenological segmentation index is selected as the segmentation point of the annual growth cycle of the pixel. That is, the annual growth cycle of the pixel starts from the annual day of that year and ends on the annual day of the following year.

[0058] Specifically, for each 250-meter pixel in the original 22-year FAC sequence, a segmentation algorithm suitable for sparse and high-noise sequences is proposed. The data for each year are aligned by the day of year (DOY), where DOY = 1, 2, ..., 366. A 21-day sliding window is used to calculate the mean, standard deviation, and maximum value of the FAC at each DOY position. These three indicators are then normalized and weighted to construct a phenological segmentation index. , represented as: ; in, This represents the normalized mean of FAC. This represents the normalized standard deviation of FAC. This represents the maximum value of FAC after normalization.

[0059] The DOY corresponding to the minimum PSI value is selected as the annual growth cycle dividing point for that pixel. That is, the annual growth cycle of that pixel begins on the DOY day of that year, marking day 1 of the corresponding growth cycle, and ends on the DOY day of the following year. For Hulun Lake, which experiences an ice age, two dividing points DOY are set. S(Start) and DOY E (End), and are respectively limited to the DOY range of [120,150] and [300,330].

[0060] Furthermore, parameter transfer and fitting are performed. For each growth cycle segmented from each pixel, 16 statistical features are extracted and input into the six-parameter mapping model trained in step S2 to predict the fitting function parameters for that pixel and that growth cycle. This allows us to obtain the fitted time series of the cycle. Using these parameters and the fitting function, we can obtain the smooth fitting curve of the cell for the cycle. By determining the relationship between A and the optimal threshold in the fitting function parameters, we can determine the optimal threshold of the growth cycle. We can extract the number of days SOS and EOS at the beginning and end of the growth cycle and convert these two data into the day of the year SOY and EOY.

[0061] Preferably, the optimal phenological threshold is dynamically calculated, and the time data of plant growth are obtained from the fitted time series, including: By fitting the maximum amplitude determined from the parameter set, a functional relationship between the maximum amplitude and the optimal active amplitude threshold is constructed. The optimal function form and function parameters are determined by the least squares method, and the optimal active amplitude threshold is calculated. Based on the optimal activity amplitude threshold, the date on which the algal bloom coverage value first rises above the threshold from the left baseline during the growth period is selected as the growth start date, and the date on which the algal bloom coverage value last falls to the threshold during the decline period is selected as the growth end date.

[0062] Specifically, a functional relationship is constructed between the maximum amplitude A and the optimal activity amplitude threshold (threshold) to achieve threshold adaptation. A reference phenological period for a batch of growth cycles is determined through visual interpretation, specifically the 15-day interval containing SOS and EOS in the i-th growth cycle of a 5km unit. The corresponding threshold is then calculated and fitted to A. This embodiment proposes three candidate function forms, including: The exponential decay function is expressed as: ; in, , , It is a constant.

[0063] The modified power function is expressed as: ; in, , , It is a constant.

[0064] The Sigmoid decay function is expressed as: ;

[0065] For the three functional forms above, the maximum amplitude A after fitting the function for each growth cycle is the independent variable, and the threshold is the dependent variable. Substituting these values ​​into the function, the parameters of the three fitting functions are determined using the least squares method, and the coefficient of determination R is used. 2 The functional relationship between the final A and the optimal threshold is determined, and the optimal activity amplitude threshold is calculated. Small amplitude corresponds to a higher threshold to suppress noise, and large amplitude corresponds to a lower threshold to capture decay signals.

[0066] For each growth cycle of each pixel, an adaptive threshold is calculated based on its maximum amplitude A using a selected optimal functional relationship. Then, on the fitted curve, the first time the FAC value during the growth cycle rises from the left baseline is found. The date on which the value rises above the threshold is the growth start date (SOS), and the date on which the FAC value last drops to the threshold during the decline period is the growth end date (EOS).

[0067] It should be noted that this invention significantly improves the accuracy and robustness of phenological parameter extraction under high-noise and sparse data. Through the innovative strategy of "low-resolution fitting and high-resolution transfer," it effectively avoids the problem of easy failure or huge deviation when directly fitting 250-meter pixel-level data with high noise and high missing data. Taking advantage of the relatively continuous and high signal-to-noise ratio of 5-kilometer scale data, a robust phenological curve fitting model is first established at this scale. Subsequently, the complex mapping relationship between the original time-series statistical features and the fitting function parameters is learned through a machine learning model, and this relationship is transferred to the pixel level. This makes pixel-level phenological extraction not only dependent on fragile data cleaning and direct fitting, but also inherits the reliable ecological process laws established on a large scale. Thus, it can still produce stable and reliable results when facing severe data missingness and noise interference.

[0068] Meanwhile, this invention also achieves adaptive dynamic optimization of the phenological extraction threshold, improving the model's applicability and accuracy across different lakes and years. It overcomes the shortcomings of traditional fixed threshold methods in adapting poorly to inland water bodies with strong spatiotemporal heterogeneity, and can automatically adapt to various scenarios, from oligotrophic to eutrophic, and from normal to anomalous years. By constructing a quantitative functional relationship between the maximum amplitude and the optimal activity amplitude threshold, the threshold can be dynamically adjusted according to the algal bloom intensity. For weaker algal blooms, a higher threshold is used to effectively suppress misjudgments due to background noise; for strong algal blooms, a lower threshold is used to accurately capture the initiation and dissipation signals of the bloom. This makes the identification of phenological parameters more consistent with actual ecological processes and significantly reduces identification bias introduced by inappropriate threshold selection.

[0069] This invention also possesses good universality and scalability, with a wide range of applications. It can be applied to multiple typical lakes in different climate zones and with different trophic levels, and its effectiveness in identifying different phytoplankton characterization indices (such as FAC and Chl-a) has been verified. Without relying on specific sensors or indices, the invention's concept of "sequence reconstruction-parameter mapping-threshold adaptation" can be easily transferred to the phenological extraction of other medium-to-high resolution remote sensing data (such as OLCI and VIIRS) or other water body optical parameters (such as NDVI and PC), providing a universal technical solution for large-scale, long-term lake aquatic ecological environment monitoring.

[0070] Example 2, step S2, constructing the growth curve benchmark model may include: Step 1: Obtain meteorological data and historical FAC data from low-resolution image data cell grids, and extract multi-peak data of plant growth through wavelet transform; Specifically, environmental data and historical FAC data are acquired from a 5km cell grid. The environmental data includes water surface temperature, solar radiation, precipitation, wind speed, etc. The environmental data and FAC data are aligned in time and then standardized.

[0071] Furthermore, multi-peak data of plant growth were extracted using wavelet transform. The number and location of potential peaks were identified using continuous wavelet transform (CWT) on the 5km FAC time series. Local maxima on the scale-time plane were analyzed to extract an initial peak candidate set. K is the upper limit of the number of candidate peaks, set to 3-5.

[0072] Step 2: Construct a multi-peak Gaussian mixture model. The parameters of each Gaussian component are controlled by environmental factors (environmental data). Use variational Bayesian method to optimize the model parameters and the number of peaks. Determine significant peaks through posterior probability, and determine the actual number and location of peaks.

[0073] Specifically, the multi-peak Gaussian mixture model is represented as: ; in, Let t be the FAC time series at time t, and b be the baseline value, representing the background FAC level during the absence of algal blooms; Let be the amplitude of the k-th Gaussian component, i.e., the peak intensity; This represents the peak position of the k-th Gaussian component; is the standard deviation of the k-th Gaussian component, used to control the peak width.

[0074] Among them, amplitude parameter Dynamic regulation by environmental factors, expressed as: ; in, The fundamental amplitude coefficient of the k-th Gaussian component. Let be the weighting coefficient of the k-th Gaussian component with respect to the j-th environmental feature transformation function. Let j be the transformation function for the j-th environmental feature, which can be a quadratic polynomial, spline basis function, etc. Let m be the normalized values ​​of n environmental variables at time t, where the environmental variables are water surface temperature, solar radiation, precipitation, and wind speed, and m is the number of environmental feature transformation functions.

[0075] The variational Bayesian method is used to optimize the model parameters and the number of peaks. Given the parameters, the joint probability of the observed data's FAC and the existence of latent variables with peaks is used to determine significant peaks through posterior probability, with a threshold set at 0.75, ultimately determining the actual number of peaks. and location .

[0076] Set constraints, perform parameter optimization and model validation. Constraints include: b is greater than or equal to 0. Greater than or equal to 0; It should be located within the corresponding seasonal window; for example, the spring peak should be from March to June. Meeting ecological constraints, such as steep growing seasons, This indicates that the decline phase is more gradual than the growth phase. The model is also evaluated and validated using the root mean square error.

[0077] Step 3: Based on the above steps, input the low-resolution algal bloom coverage time series to obtain multiple growth peak parameter sets for each low-resolution unit, i.e., for each 5km unit. Phenological parameter set of each growth peak and the environmental weighting coefficient corresponding to the growth peak. .

[0078] Furthermore, a parameter mapping model is constructed using a spatiotemporal awareness graph attention transfer network, including: A spatiotemporal graph structure is constructed, representing the lake region as a graph G=(V,E), where V represents the set of nodes, with each high-resolution pixel being a node. The node features in the set include the original FAC sequence and its statistical characteristics. E represents the set of edges, constructed through hydrological connectivity and geographic distance, and is expressed as follows: ; in, Let i be the Euclidean distance between pixels i and j. This is the distance decay coefficient, which represents the degree to which geographical distance affects edge weights; its value is 1km. The hydrological connectivity coefficient can be determined based on historical flow fields or FAC correlation.

[0079] A spatiotemporal feature encoder is constructed, including a temporal encoding module, a spatial attention mechanism module, and a spatial aggregation module. The temporal encoding module constructs a multi-scale LSTM encoder for each pixel's FAC sequence and its 16 statistical features (the extraction of statistical features is the same as in Example 1) to capture short-term and long-term dependencies. The spatial attention mechanism module calculates the attention coefficients between nodes and measures the intensity of spatial influence through operations such as splicing attention vectors. The spatial aggregation module aggregates neighbor information through attention weighting, that is, it aggregates the neighbor set of a node and the attention coefficients between nodes through a nonlinear activation function.

[0080] The multi-task parameter prediction layer outputs the phenological parameter set corresponding to the pixel prediction, and a multi-head prediction branch is designed, with each branch corresponding to a phenological parameter set. ,in, This represents the baseline parameters that migrated at a 5km scale. For parameter mapping functions, i.e., multilayer perceptron, This represents the parameter values ​​in the predicted phenological parameter set, which in this embodiment is on a 5km scale. The final representation of node i is the feature obtained by fusing spatiotemporal information through a spatiotemporal feature encoder, representing the baseline parameters for migration from low-resolution pixels; simultaneously, a physical consistency constraint loss is introduced, including peak temporal increment. And the growth period is steeper than the decline period. The weighted summation is then used to calculate the constraint loss value.

[0081] In step S3, similar to Example 1, the high-resolution original algae coverage time series for each year is aligned according to the accumulated days of the year. Using a preset sliding window, the mean, standard deviation, and maximum value of algae coverage at each accumulated day are calculated and weighted to obtain the phenological segmentation index. The accumulated day corresponding to the minimum value of the phenological segmentation index is selected as the segmentation point for the annual growth cycle of that pixel. Simultaneously, based on step S2 in this example, 16 statistical features and environmental data features of the high-resolution 250m FAC time series are obtained within the life cycle. The 16 FAC statistical features and 4 environmental features for each pixel in each growth cycle are output, forming a 20-dimensional feature vector. This vector serves as the input to the parameter mapping model. Based on the parameter values ​​of the 5km pixel, the parameter mapping model outputs the predicted wind quantity and phenological parameter set for 250m. Then, based on the predicted parameters, the fitted FAC curve for each pixel in each growth cycle at 250m is reconstructed.

[0082] Furthermore, in this embodiment, based on the trained parameter mapping model, key phenological parameters of the growth cycle are identified and extracted through a preset dynamic threshold, as follows: Step 1: Calculate the first threshold by setting a baseline threshold and combining it with an environmental threshold and a signal quality threshold, i.e., by dynamically adjusting it using environmental indices and signal quality indicators. Specifically, the baseline threshold can be set to 0.2. The environmental threshold and signal quality threshold are calculated separately, and then combined to obtain the first threshold. First, based on water temperature, solar radiation, and precipitation, temperature and solar radiation are calculated by using real-time environmental data and suitable environmental data for phytoplankton through a Gaussian response function, indicating the existence of an optimal value. At the same time, solar radiation is also calculated using an exponential function to determine whether the actual solar radiation value is greater than the minimum solar radiation required for growth, thus obtaining the water temperature threshold and solar radiation threshold. Precipitation is calculated through a linear negative effect, indicating that precipitation inhibits phytoplankton growth. By weighting and combining the thresholds of the three factors, an environmental index is obtained. The environmental index is then smoothed using a Sigmoid function to obtain the environmental threshold, ensuring that the threshold increases when the environment is extremely unsuitable and decreases when the environment is suitable.

[0083] Environmental indices are expressed as: ; in, , , These are the weighting coefficients for temperature, radiation, and precipitation, respectively. The water meter temperature for that day of that year. For the optimal water surface temperature for phytoplankton, this embodiment uses 22°C. This is the width parameter for the temperature response curve, with a value of 8. This refers to the solar radiation on that day. The minimum solar radiation required for phytoplankton growth. The width parameter of the radiation response curve is set to 100. This is an indicator function; it returns 1 when the condition is true and 0 otherwise. This represents the normalized precipitation for that day and year.

[0084] Environmental threshold Represented as: ; in, This is the steepness coefficient of the Sigmoid function, controlling the transition speed, and its value is 3. The midpoint of the Sigmoid function corresponds to a moderately suitable condition, with a value of 0.5.

[0085] Furthermore, the FAC signal quality of each pixel at each time point is evaluated. The signal-to-noise ratio (SNR) is calculated by comparing the fitted value of the reconstructed FAC curve with the background noise level. A lower threshold is used for high-quality signal areas, and a higher threshold is used for low-quality signal areas to avoid noise interference. The background noise level can be the moving median of the past 30 days. The SNR is also smoothed using the Sigmoid function, while a minimum adjustment coefficient is set to ensure that the threshold of 0.3 is not too low, thus obtaining the signal quality threshold.

[0086] Furthermore, the product of the baseline threshold, the environmental threshold, and the signal quality threshold is used as the first threshold.

[0087] Step two: Assign a corresponding threshold to each identified peak, and calculate the second threshold based on the first threshold and the set peak weight.

[0088] Specifically, weights are assigned based on the amplitude proportion of each peak, that is, the proportion of the predicted amplitude of the current peak to the predicted amplitude of the total number of peaks. Strong peaks receive higher weights. Then, based on the first threshold and the peak weights, a second threshold is calculated. The threshold is lowest near the peak position and gradually increases towards both sides, as shown below: ; in, This represents the threshold value for the k-th peak at position (x, y) on day d of the year. The first threshold, The peak weight is d, where d is the d-th day of the year. The predicted location of the k-th peak is the number of days in the growth cycle. is the standard deviation of the Gaussian function, with a value of 20.

[0089] Step 3: Predicting the location based on the peak value To determine the peak time range, the Gaussian distribution characteristics are used, and ±2 standard deviations are taken. The range is used to obtain the expected start time and expected end time of the kth peak.

[0090] Step four: Considering that phytoplankton growth is usually faster than decline, formulate a growth-decline asymmetric threshold adjustment strategy. The threshold during the growth period gradually increases from the bottom, and the threshold during the decline period gradually increases from the peak position, with a gentler slope during the decline period.

[0091] Specifically, the asymmetric threshold function is defined as follows: ; ; in, Let be the threshold value for the k-th peak value during the growth phase. Let be the threshold for the k-th peak during the decline period. To adjust the coefficient and control the intensity of the threshold change over time, the value ranges from 0 to 1, and in this embodiment, it is set to 0.4. Let k be the expected start time of the kth peak. This represents the expected end time of the k-th peak.

[0092] It should be noted that this step is based on a peak threshold and is adjusted asymmetrically according to the growth stage. For example, at the beginning of the growth period, the threshold is [value missing]. At the peak position The threshold decreased by 40% ( =0.4), which ensures accurate capture of SOS (start of growth date) and EOS (end of growth date).

[0093] Step 5: Determine the final threshold based on the asymmetric threshold function, and extract phenological parameters from the fitted curve, including SOS and EOS for each peak, as well as phenological parameters for the overall growth cycle.

[0094] Specifically, for the extraction of unimodal phenological parameters, for each peak value k, the date on which the FAC first exceeds the growth period threshold is identified as the peak value. During the recession, find the date on which the FAC last exceeded the recession threshold as... If there is no d that satisfies the condition, then and Set it to NaN.

[0095] The phenological parameters of multiple peaks were integrated into overall growing season parameters, and the dominant peak with the largest biomass was identified. The start date of the entire growing season was the date with the smallest start date among the single peaks, and the end date was the date with the largest end date among the single peaks. The total growing duration was calculated by subtracting the start date from the end date of the entire growing season. The peak corresponding to the maximum growth duration was identified as the dominant peak by multiplying the amplitude parameter corresponding to each peak by the maximum value. Finally, the number of days in the growing cycle was calculated from the dividing point and converted back to annual days to obtain the key phenological parameters for the final 250m pixel.

[0096] It should be noted that this invention solves the problem of unstable high-resolution data quality by establishing a multi-scale parameter migration from a 5km reference cell to a 250m pixel. Considering environmental factors, the multi-peak Gaussian process model can accurately identify multi-peak events, establish a dynamic functional relationship between environmental variables and growth parameters, and ensure the asymmetry of the growth-decay process through parameter constraints, thus more realistically reflecting the phytoplankton life cycle.

[0097] The above is an illustrative scheme of a method for extracting phenological parameters of lake phytoplankton based on multi-scale parameter migration and dynamic thresholding according to this embodiment. It should be noted that the technical solution of this system for extracting phenological parameters of lake phytoplankton based on multi-scale parameter migration and dynamic thresholding is based on the same concept as the aforementioned method for extracting phenological parameters of lake phytoplankton based on multi-scale parameter migration and dynamic thresholding. Details not described in detail in this embodiment can be found in the description of the aforementioned method for extracting phenological parameters of lake phytoplankton based on multi-scale parameter migration and dynamic thresholding.

[0098] Example 3: This example provides a system for extracting phenological parameters of lake phytoplankton based on multi-scale parameter migration and dynamic thresholding, including: The acquisition and processing module is used to acquire remote sensing image data of the target area, preprocess the remote sensing image data, and construct multi-scale time series image data. The mapping module is used to divide the plant growth cycle based on multi-scale time series image data, so as to identify the growth cycle, construct the growth curve benchmark model and parameter mapping model, and obtain the phenological morphological parameters of the plant during the growth cycle. The output module is used to perform pixel-level growth cycle segmentation on the original pixel sequence of image data, and based on the trained parameter mapping model, identify and extract key phenological parameters of the growth cycle through a preset dynamic threshold.

[0099] This embodiment also provides a computer device suitable for extracting lake phytoplankton phenological parameters based on multi-scale parameter migration and dynamic thresholds, including: The system includes a memory and a processor. The memory stores computer-executable instructions, and the processor executes these instructions to implement a method for extracting phenological parameters of lake phytoplankton based on multi-scale parameter migration and dynamic thresholds, as proposed in the above embodiments.

[0100] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements a method for extracting lake phytoplankton phenological parameters based on multi-scale parameter migration and dynamic thresholds as proposed in the above embodiment.

[0101] The storage medium proposed in this embodiment belongs to the same inventive concept as the method for extracting lake phytoplankton phenological parameters based on multi-scale parameter migration and dynamic threshold proposed in the above embodiments. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0102] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0103] Example 4, refer to Table 2 and Figure 3 As an embodiment of the present invention, a method for extracting phenological parameters of lake phytoplankton based on multi-scale parameter migration and dynamic threshold is provided. In order to verify its beneficial effects, it is scientifically demonstrated through economic benefit calculation and simulation experiments.

[0104] To better evaluate the performance improvement of this invention, four comparison methods C1-C4 are proposed for comparison, as shown in Table 2. The data used are FAC time series extracted from MODIS-Aqua resolution images, and three amplitude thresholds of 10%, 20%, and 50% are used to determine SOS and EOS, as detailed below: Comparison Method 1 uses the fitting method proposed in this invention to smooth the annual time, and only replaces the active amplitude threshold with three fixed amplitude threshold extraction parameters; Comparison Method 2 calculates the FAC average within the time window in 10-day units. If all values ​​within a window are missing, the average value of that window represents the missing state. The time series is compressed to 1 / 10 of its original length, and linear interpolation is used to fill the window values. Then, harmonic filtering is used to smooth the series, and finally, the time series is restored to its original length. Since the filtering method cannot smoothly smooth the annual time series into a strictly monotonically increasing and then monotonically decreasing form, the maximum value of each year after filtering is taken as the maximum amplitude A, and the minimum value to its left is taken as the left baseline value b. L The minimum value to its right is taken as the right baseline value b. R The time corresponding to the first time the amplitude threshold is reached on the left is SOS, and the time corresponding to the last time the amplitude threshold is reached on the right is EOS. Comparing method 3 with method 2, the operation process is the same, except that the harmonic filtering is replaced with Gaussian filtering to smooth the time series; In comparison method 4, an improved Bayesian land surface phenology model is used. First, the FAI time series is fitted using a double logistic function based on the Bayesian hierarchical model. Then, SOS and EOS are extracted using three fixed amplitude thresholds.

[0105] Table 2: Summary Table of Comparison Methods Data and Representation Index 250m FAC time series 250m FAC time series 250m FAC time series 250m FAC time series Reconstructing time series Method of the present invention HANTS Gaussian filter BLSP fitting Extraction phenotypes Fixed amplitude threshold Fixed amplitude threshold Fixed amplitude threshold Fixed amplitude threshold The algorithm's performance is evaluated by comprehensively determining the start and end times of different years in the sampled pixels as true values ​​and comparing them with the model's extracted parameters. The evaluation is based on R², absolute percentage difference (APD), unbiased absolute percentage difference (UAPD), bias, root mean square error (RMSE), and unbiased relative percentage root mean square error (URMSE).

[0106] Experimental results are as follows Figure 3 As shown, this invention uses 50 selected sampling points covering different years to form 330 sets of phenological true values ​​to verify the extraction accuracy. Here, N is the effective extracted value of the phenological parameter, representing the total number of corresponding phenological parameters that can be successfully extracted after reconstructing the time series. The optimal fixed thresholds for comparison methods 1-4 are 15%, 50%, 50%, and 15%, respectively, and the scatter plot presents the optimal fixed threshold results for the comparison methods. The scatter plot shows that the method of this invention performs best among all comparison methods. The coefficients of determination (R²) of its extracted start date (SOS), end date (EOS), and growing season duration (LOS) compared to the measured true values ​​reach 0.90, 0.64, and 0.78, respectively, significantly better than other comparison methods.

[0107] The extraction accuracy of methods 1 and 4 is the second lowest, with R² ranging from 0.40 to 0.60. The optimal fixed threshold for both methods is 15%. Among them, method 1, which uses the time series reconstruction method proposed in this invention, performs slightly better than method 4, which uses the LSP method. Furthermore, the effective extraction value (N) for each parameter of method 1 is higher than that of method 4, demonstrating the higher robustness of the time series reconstruction method proposed in this invention. On the other hand, method 2 based on harmonic decomposition and method 3 based on Gaussian filtering have relatively low accuracy, with R² for each phenological parameter below 0.5 and effective extraction values ​​(n) lower than the other two comparison methods.

[0108] The main reason for the lower accuracy of Methods 2 and 3 lies in the specific climatic conditions of the study area (Taihu Lake and Chaohu Lake). The plum rain season lasts for about 1-2 months each year, resulting in large-scale systematic data gaps in remote sensing observations. Methods such as harmonic analysis and Gaussian filtering are highly dependent on the continuity of time series and are extremely sensitive to such long-term data gaps, making it difficult to accurately reconstruct complete phenological curves, thus introducing significant biases in the extraction of phenological parameters.

[0109] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for extracting phenological parameters of lake phytoplankton based on multi-scale parameter migration and dynamic thresholding, characterized in that, include: Acquire remote sensing image data of the target area, and preprocess the remote sensing image data to construct multi-scale time series image data; Based on the multi-scale time series image data, the plant growth cycle is divided to identify the growth cycle, and a growth curve benchmark model and parameter mapping model are constructed to obtain the phenological morphological parameters of the plant during the growth cycle. The original pixel sequence of the image data is segmented into pixel-level growth cycles, and based on the trained parameter mapping model, key phenological parameters of the growth cycle are identified and extracted through a preset dynamic threshold. Constructing a growth curve benchmark model includes: Meteorological data and historical FAC data of low-resolution image data cells are acquired, and multi-peak data of plant growth are extracted through wavelet transform. A multi-peak Gaussian mixture model is constructed, where the parameters of each Gaussian component are controlled by environmental factors. The variational Bayesian method is used to optimize the model parameters and the number of peaks. Significant peaks are determined through posterior probability, thus determining the actual number and location of peaks. The multi-peak Gaussian mixture model is expressed as follows: ； in, Let t be the FAC time series at time t, and b be the baseline value, representing the background FAC level during the absence of algal blooms; Let be the amplitude of the k-th Gaussian component, i.e., the peak intensity; This represents the peak position of the k-th Gaussian component; is the standard deviation of the k-th Gaussian component, used to control the peak width; Inputting a low-resolution time series of algal bloom coverage yields multiple growth peak parameter sets for each low-resolution cell, i.e., for each 5km cell. Phenological parameter set for each growth peak ; Constructing the parameter mapping model includes: Based on the high-resolution original algal bloom coverage time series of the target area within each growth cycle, multiple statistical features of each pixel in the series are calculated. A parameter mapping model is constructed using a machine learning model. The high-resolution statistical features are used as the model input, and the fitting parameter set of the low-resolution sequence is used as the model output. The parameter mapping model is trained to establish the mapping relationship between the multiple statistical features and the fitting parameter set, so as to obtain the prediction fitting parameters of the high-resolution original algal cover time series.

2. The method for extracting lake phytoplankton phenological parameters based on multi-scale parameter migration and dynamic thresholding as described in claim 1, characterized in that, Acquire remote sensing image data of the target area and preprocess the remote sensing image data, including: Collect raw remote sensing image data of the target area at different spatial resolutions, and extract the algae coverage features from the raw remote sensing image data; Global surface water data is collected, and a region is selected from the global surface water data as a reference water body mask. A dynamic water body mask algorithm is constructed using multispectral index and reflectance threshold to remove non-water body pixels and interfering water body pixels from the remote sensing image data, thereby obtaining the effective water body pixels in the remote sensing image data. The daily algae coverage index is calculated for the effective water body pixels, and multi-scale aggregation is performed on remote sensing image data with different spatial resolutions to obtain the algae coverage time series.

3. The method for extracting lake phytoplankton phenological parameters based on multi-scale parameter migration and dynamic thresholding as described in claim 2, characterized in that, Multi-scale aggregation of remote sensing image data at different spatial resolutions was performed to obtain the time series of algae cover, including: The original remote sensing image data with different spatial resolutions are divided into high-resolution and low-resolution image data. After extracting the algae coverage features, high-resolution original algae coverage time series and low-resolution algae coverage time series are obtained. The low-resolution image data is divided into several grid cells of the same size. The daily algae coverage value is determined based on the proportion of effective pixels in the grid cells, thus obtaining the low-resolution algae coverage time series. The low-resolution algae coverage time series was reconstructed by sequence difference to obtain the complete low-resolution algae coverage time series.

4. The method for extracting lake phytoplankton phenological parameters based on multi-scale parameter migration and dynamic thresholding as described in claim 3, characterized in that, Obtaining phenological parameters of plants during their growth cycle includes: Based on the algae coverage index, the growth cycle of the low-resolution algae coverage time series is segmented to obtain multiple independent growth cycles. Based on the independent growth cycle, curve fitting is performed on the low-resolution algae coverage time series of the independent growth cycle to obtain the fitting parameter set for each growth cycle. A parameter mapping model is constructed, and the model is trained using the fitted parameter set to establish a mapping relationship between low-resolution data and high-resolution data.

5. The method for extracting lake phytoplankton phenological parameters based on multi-scale parameter migration and dynamic thresholding as described in claim 2, characterized in that, Identifying and extracting key phenological parameters for the growth cycle includes: By using a sliding window, the phenological segmentation index of the high-resolution original algal cover time series is calculated to divide the growth cycle. Based on each growth cycle segmented by each high-resolution pixel, statistical features are extracted and input into the parameter mapping model to predict the fitting parameter set for each growth cycle of the pixel, thereby obtaining the fitting time series of the cycle. Based on the maximum amplitude of each growth cycle, the optimal phenological threshold is dynamically calculated, and the time data of plant growth are obtained from the fitted time series.

6. The method for extracting lake phytoplankton phenological parameters based on multi-scale parameter migration and dynamic thresholding as described in claim 5, characterized in that, Calculate the phenological segmentation index of the high-resolution original algal cover time series to divide the growth cycle, including: The high-resolution original algal bloom coverage time series for each year are aligned by year-to-day. The mean, standard deviation, and maximum value of algae coverage at each location on each day of the year are calculated using a preset sliding window and then fused to obtain the phenological segmentation index. The day corresponding to the minimum value of the phenological segmentation index is selected as the annual growth cycle segmentation point of the pixel. That is, the annual growth cycle of the pixel starts from the day of the year and ends on the day of the year of the following year.

7. The method for extracting lake phytoplankton phenological parameters based on multi-scale parameter migration and dynamic thresholding as described in claim 6, characterized in that, Dynamically calculate the optimal phenological threshold and obtain plant growth time data from the fitted time series, including: The maximum amplitude determined by the set of fitting parameters is used to construct a functional relationship between the maximum amplitude and the optimal active amplitude threshold. The optimal function form and function parameters are determined by the least squares method, and the optimal active amplitude threshold is calculated. Based on the optimal activity amplitude threshold, from the fitted time series, the date on which the algal bloom coverage value first rises above the threshold from the left baseline during the growth period is selected as the growth start date, and the date on which the algal bloom coverage value last falls to the threshold during the decline period is selected as the growth end date.

8. A system for extracting phenological parameters of lake phytoplankton based on multi-scale parameter migration and dynamic thresholding, employing the method for extracting phenological parameters of lake phytoplankton based on multi-scale parameter migration and dynamic thresholding as described in any one of claims 1 to 7, characterized in that, include: The acquisition and processing module is used to acquire remote sensing image data of the target area and preprocess the remote sensing image data to construct multi-scale time series image data. The mapping module is used to divide the plant growth cycle based on the multi-scale time series image data, so as to identify the growth cycle, construct a growth curve benchmark model and a parameter mapping model, and obtain the phenological morphological parameters of the plant during the growth cycle. The output module is used to perform pixel-level growth cycle segmentation on the original pixel sequence of the image data, and based on the trained parameter mapping model, identify and extract key phenological parameters of the growth cycle through a preset dynamic threshold.

Images