A method and system for high-spectral recognition of algal community structure in a bloom period

Abstract

The application discloses a kind of algal bloom period algal community structure hyperspectral identification method and system, and the application constructs ecological mechanism constraint type loss function.Significantly improve the consistency, reliability and explainability of algal species classification and chlorophyll-a concentration inversion result, effectively overcome the result contradiction problem caused by pure data-driven model;By constructing a joint identification network, introduce ecological environment suitability index chart and algal species classification and chlorophyll-a priori chart, greatly improve the generalization adaptation ability to complex and changeable water environment conditions;Through the collection of multi-source data, construct information-rich multi-source data cube, clearly distinguish between "promote factor" and "inhibitory factor", clearly "ecological suitable promotion penalty term" and "ecological inhibition constraint penalty term", fully excavate the synergistic constraint relationship of multi-source environmental factors on algal growth, realize the deep integration of meteorology, water quality, remote sensing data and mechanism synergy.

Description

Technical Field

[0001] This invention relates to the field of water environment remote sensing monitoring and experimental technology, specifically to a hyperspectral identification method and system for algal community structure during algal blooms. Background Technology

[0002] The composition of algae and the concentration of chlorophyll a in water bodies are important indicators reflecting the material composition of aquatic ecosystems, the level of eutrophication, and phytoplankton biomass. Among these, algal species type reflects changes in the composition of dominant phytoplankton groups in the water, while chlorophyll a concentration is often used to characterize algal biomass and the nutrient status of the water. Traditional algal species identification and chlorophyll a measurement usually rely on field sampling, laboratory microscopic observation, chromatographic analysis, or spectrophotometry. Although the results are relatively reliable, they suffer from insufficient timeliness, limited spatial coverage, and difficulty in achieving large-scale continuous monitoring.

[0003] Remote sensing technology can rapidly retrieve chlorophyll a concentration and spatial distribution of algae in water bodies based on the optical reflection, absorption, and scattering characteristics of water materials. With the development of multispectral and hyperspectral remote sensing data and deep learning technology, using features such as remote sensing reflectance, red-edge band, near-infrared band, phytoplankton index, and normalized differential chlorophyll index for algae identification and chlorophyll a concentration retrieval has become an important technological direction.

[0004] However, existing methods still have the following shortcomings: First, some methods mainly rely on empirical thresholds or statistical regression models, which are difficult to adapt to spectral differences under different water bodies, seasons, and algal compositions. Second, although deep learning methods can improve inversion accuracy, model training usually only aims to minimize classification or concentration errors, lacking explicit constraints on algal spectral response patterns, environmental suitability, and chlorophyll a change mechanisms, which may lead to contradictions between algal species classification results and chlorophyll a concentration inversion results. Third, existing multi-source fusion methods usually use meteorological and water quality information as ordinary input features, and have not fully utilized the constraints of factors such as temperature, nutrients, wind speed, hydrodynamics, and transparency on algal growth and chlorophyll a changes.

[0005] Therefore, there is an urgent need to propose a new method for remote sensing determination of algal species classification and chlorophyll a concentration, which integrates remote sensing spectral physical characteristics, water quality chemical properties, and ecological environment mechanism knowledge into the model training process. By constructing an ecological mechanism constraint loss function, the consistency, reliability, and interpretability of the algal species classification and chlorophyll a concentration inversion results can be improved. Summary of the Invention

[0006] The purpose of this invention is to provide a hyperspectral identification method and system for algal community structure during algal blooms, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a hyperspectral identification method for algal community structure during algal blooms, comprising: Step 1: Multi-source data acquisition and preprocessing Collect multi-source data of the target water area and perform preprocessing. Then, stack the preprocessed data according to the channel dimension to construct a multi-source data cube X. Step 2: Constructing a map of aquatic ecological environment suitability indicators Driving factors were screened and normalized, and promoting and inhibiting factors were distinguished. Forward and reverse mapping processes were applied respectively, and a weighted fusion algorithm was used to generate an ecological environment suitability index map. ; Step 3: Construct a priori map of algal species classification and chlorophyll a Construct a growth mechanism ecological rule base, input the current environmental conditions to calculate the rule matching degree, and weight and aggregate the prior graph. ; Step 4: Construct a joint identification network X is input into a joint recognition network constructed from an encoder, an algal species classification head, and a chlorophyll a concentration regression head. The recognition network outputs the algal species classification probability. Continuous values ​​of chlorophyll a concentration ; Step 5: Construct an ecological mechanism-constrained loss function Loss by algal species classification Regression loss of chlorophyll a concentration A composite loss function is constructed by superimposing four types of ecological mechanism constraint penalty terms on the basic loss term, thereby affecting the algal species classification results, chlorophyll a concentration measurement results, and ecological environment constraint map. and rule prior graph Maintain consistency between them; Step Six: Model Training and Measurement Results Output The joint recognition network is trained using training samples. After training, multi-source data is input into the model, and the model outputs an algal species classification map, a chlorophyll a concentration distribution map, and measurement results.

[0008] Preferably, in step one, the multi-source data includes multispectral remote sensing image data or hyperspectral remote sensing reflectance data, meteorological data, and water quality data from ground monitoring stations.

[0009] Preferably, meteorological data includes water temperature or sea surface temperature, wind speed, wind direction, and precipitation; ground monitoring station water quality data includes one or more of chlorophyll a, total nitrogen, total phosphorus, nitrogen-to-phosphorus ratio, transparency, conductivity, and suspended solids concentration. Remote sensing image data preprocessing: Radiometric calibration, atmospheric correction, and geometric correction are performed sequentially on the original multispectral remote sensing images to eliminate sensor errors and atmospheric scattering interference; invalid areas, including land and clouds, are removed using water body mask extraction technology and cloud shadow removal technology to accurately preserve water image data; and the water remote sensing reflectance is obtained based on the preprocessed water image. Meteorological data preprocessing: The meteorological data from discrete stations are projected onto a spatial grid that is completely consistent with the remote sensing image using Kriging interpolation or bilinear interpolation to achieve spatial resolution matching between the meteorological data and the remote sensing data; the wind speed data is further decomposed into horizontal U-components and vertical V-components to improve the hydrodynamic meteorological characteristic parameters. Water quality data preprocessing: Core water quality indicators were selected, including chlorophyll a, total nitrogen, total phosphorus, nitrogen-to-phosphorus ratio, transparency, conductivity, and suspended solids concentration. The interquartile range method was used to remove outliers in the water quality data to avoid interference from monitoring errors. Based on the imaging date of the remote sensing image, water quality monitoring data with matching time were screened, and missing data were filled in using linear interpolation to achieve accurate matching of the time dimension between water quality data and remote sensing data.

[0010] Preferably, the promoting factors include temperature factor T, nitrogen factor N, phosphorus factor P, and algal sensitivity index A; wherein, algal sensitivity index A is obtained by normalizing and weighting at least two of the normalized differential chlorophyll index NDCI, phytoplankton index FAI, chlorophyll index MCI, and red edge ratio index REI, and the higher the normalized value of the promoting factor, the stronger its supporting effect on algal growth or chlorophyll a concentration increase; The inhibition factors include wind speed factor W and hydrodynamic / transparency composite factor H; where H is obtained by normalizing at least one of flow velocity, mixing intensity and transparency, and the higher the normalized value of the inhibition factor, the stronger the water disturbance or light inhibition effect.

[0011] Preferably, in step three, the ecological rule base is represented in the form of "preconditions-conclusions"; before rule discrimination, temperature T, total phosphorus TP, wind speed W, transparency SD, nutrient comprehensive factor NUT, and hydrodynamic factor HD are normalized to the [0,1] interval, and discrimination is performed according to the following numerical standards: Defined as high temperature, Defined as medium temperature, Defined as high total phosphorus, Defined as low wind speed, Defined as high transparency, Defined as low transparency, Defined as high in nutrients, Defined as low in nutritional value Defined as weak hydrodynamics; Specifically, there are three rules: Rule 1: When , and When the spatial location is determined to meet the prior condition for cyanobacterial growth dominance; Rule 2: When , and When the spatial location satisfies the prior condition of diatom or green algae dominance, it is determined that the location satisfies the prior condition of diatom or green algae dominance. Rule 3: When , and When the spatial location satisfies the prior condition of increased chlorophyll a concentration, it is determined that the location satisfies this condition.

[0012] Preferably, in step four, the shared encoder is used to extract spatial features, spectral features and environmental features from the multi-source data cube X. The encoder adopts a convolutional Swin-Transformer hybrid architecture: the local spatial detail features of the multi-source data are extracted through the convolutional layer, and the global feature association relationship is modeled through the Transformer module to output a unified multi-source shared feature representation. The algae classification head receives the feature map output from the shared encoder and outputs a probability vector for each pixel or image patch belonging to a different algae species. ; The chlorophyll a concentration regression head receives shared features and outputs a continuous value graph of chlorophyll a concentration. .

[0013] Preferably, in step five, the four types of ecological mechanism constraint penalty terms are as follows: Ecological suitability promotes punishment items : ,in Used to constrain the ecological suitability of aquatic waters When the model output is lower than The situation; Ecological inhibition constraint penalty items : Used to constrain the ecological suitability of aquatic waters At that time, the predicted concentration of normalized chlorophyll a was higher than The situation; Rule consistency penalty : Used to constrain the comprehensive output of the model With the rule prior graph The difference between them shall not exceed the allowable deviation. ; Category—Concentration Consistency Penalty ,in This represents the algal biomass level obtained from the algal species classification probability mapping. This represents the concentration level obtained by mapping the normalized concentration of chlorophyll a, used to constrain the biomass level corresponding to the classification output, ensuring that the deviation from the level corresponding to the chlorophyll concentration does not exceed the allowable deviation. ; , , , The tolerance coefficient is determined using a validation set within the range of 0.01–0.10. This indicates the total number of samples. For the first Index of each sample.

[0014] Preferably, in step six, the model is trained using training samples with algal species labels and measured chlorophyll a concentration values. In each round of training, the ecological environment constraint map is first invoked. Algal species classification / chlorophyll a priori diagram Then the network outputs the algal species classification results. Results of chlorophyll a concentration The loss value is calculated based on the total loss function, and backpropagation and parameter updates are performed using the Adam or SGD optimizer.

[0015] A hyperspectral identification system for algal community structure during algal blooms, used to implement a hyperspectral identification method for algal community structure during algal blooms, comprising: I. Data Acquisition and Preprocessing Module The data acquisition and preprocessing module is the basic data processing unit at the front end of the system. It integrates the functions of raw data acquisition and access and standardized preprocessing of multi-source data, and communicates and connects with the back-end modules of the system. This module is used to collect and access multi-source basic data of the target monitoring area, specifically including multispectral remote sensing image data, ERA5 reanalysis meteorological data, and ground monitoring station water quality data; at the same time, it performs full-process standardized preprocessing on the acquired multi-source raw data, which is used for radiometric calibration, atmospheric correction, and water body mask extraction of remote sensing images, and outlier removal, interpolation, time matching and normalization of meteorological and water quality data. II. Data Processing and Intelligent Recognition Module The data processing and intelligent recognition module is the core computing unit of the system, integrating the ecological suitability construction module, the prior map construction module, the joint recognition module, and the ecological mechanism loss constraint module; The ecological environment constraint map construction module is used to construct an ecological environment constraint map based on temperature, nitrogen, phosphorus, algae sensitivity index, wind speed, and hydrodynamic / transparency factors. ; The prior graph construction module is used to construct graphs based on the ecological rule base and current environmental conditions. ; The joint identification module is used to output algal species classification results and continuous chlorophyll a concentration measurement results using a deep learning network; The loss constraint module is used to construct the ecological mechanism constraint loss function and participate in parameter optimization during the model training phase; III. Result Output Module The results output module is used to output algal species classification diagrams, chlorophyll a concentration distribution diagrams, and corresponding statistical analysis results.

[0016] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention constructs an ecological mechanism-constrained loss function. Based on traditional classification and regression losses, this loss function adds four types of mechanism-constrained penalty terms: ecological suitability promotion penalty term, ecological inhibition constraint penalty term, rule consistency penalty term, and classification-concentration consistency penalty term. This significantly improves the consistency, reliability, and interpretability of algal species classification and chlorophyll a concentration inversion results, effectively overcoming the inconsistencies in results caused by purely data-driven models.

[0017] 2. This invention significantly improves the generalization adaptability to complex and variable aquatic environmental conditions by constructing a joint identification network and introducing an ecological environment suitability index map and an algal species classification and chlorophyll a prior map. Furthermore, it enables the model to provide effective training and supervision signals based on ecological mechanism knowledge even in scenarios lacking sufficient measured label data, significantly reducing the dependence on large-scale manually labeled samples, and enhancing the model's migration and generalization capabilities under different water body types and seasonal conditions.

[0018] 3. This invention collects multi-source data and constructs an information-rich multi-source data cube, clearly distinguishing between "promoting factors" and "inhibiting factors," and identifying "ecological suitability promotion penalty items" and "ecological inhibition constraint penalty items." It fully explores the synergistic constraint relationship between multi-source environmental factors on algal growth, and achieves deep integration and mechanistic synergy of meteorological, water quality, and remote sensing data. Attached Figure Description

[0019] Figure 1 This is an overall flowchart of a hyperspectral identification method for algal community structure during algal blooms according to the present invention; Figure 2 This is an overall flowchart of the algal species classification and chlorophyll a concentration determination method of the hyperspectral identification method of algal community structure during algal blooms according to the present invention. Figure 3 This is a schematic diagram illustrating the construction of an ecological environment suitability index map for a hyperspectral identification method for algal community structure during algal blooms according to the present invention. Figure 4This is a schematic diagram of the process of algal species classification and chlorophyll a priori map construction and matching of a hyperspectral identification method for algal community structure during algal blooms according to the present invention. Figure 5 This is a schematic diagram illustrating the ecological mechanism constraint loss function of the hyperspectral identification method for algal community structure during algal blooms according to the present invention. Figure 6 This is a schematic diagram of the joint recognition network training process of the hyperspectral recognition method for algal community structure during algal blooms according to the present invention. Figure 7 This is a schematic diagram of the chlorophyll a inversion results of a hyperspectral identification method for algal community structure during algal blooms according to the present invention. Figure 8 This is a spatial representation of the chlorophyll a inversion results of the hyperspectral identification method for algal community structure during algal blooms according to the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see Figure 1-8 This invention proposes a hyperspectral identification method for algal community structure during algal blooms, which is implemented through the hyperspectral identification system for algal community structure during algal blooms constructed below: Step 1: Multi-source data acquisition and preprocessing The main purpose of this step is to select target lakes, reservoirs, estuaries, or nearshore sea areas as the study area, collect multi-source environmental monitoring data within the target water area at the same time window, and perform standardized preprocessing to provide spatiotemporally aligned input features for subsequent model construction.

[0022] Multi-source data includes remote sensing data, meteorological data, water quality data, and auxiliary data (optional).

[0023] Remote sensing data: multispectral remote sensing image data or hyperspectral remote sensing reflectance data; Meteorological data: water temperature or sea surface temperature, wind speed, wind direction, precipitation; Water quality data: one or more of the following: chlorophyll a, total nitrogen, total phosphorus, nitrogen-to-phosphorus ratio, transparency, conductivity, and suspended solids concentration; Supporting data: hydrodynamic data, transparency data, or historical monitoring data.

[0024] Data preprocessing: The above data undergoes radiometric calibration, atmospheric correction, water mask extraction, outlier removal, missing value imputation, spatial interpolation, temporal matching, and normalization to obtain spatiotemporally aligned input features. The specific preprocessing methods are as follows: 1.1 Remote Sensing Image Data Preprocessing: Radiometric calibration, atmospheric correction, geometric correction, water mask extraction, and cloud shadow removal are performed sequentially. Specifically: First, based on the gain and offset coefficients in the sensor calibration file, the original digital quantization value DN is converted into the top-of-atmosphere radiance. ,Right now: ; In the formula, This is the linear amplification factor by which the sensor converts the received incident radiation energy into a digital signal (DN value). This is a digital quantization value (grayscale value). This represents the sensor offset (bias coefficient) at the corresponding wavelength λ. This formula is the core linear formula for absolute radiometric calibration of satellite remote sensing data, used to convert the physically meaningless digital grayscale values ​​(DN) originally recorded by satellite sensors into apparent radiance values ​​at the top of the atmosphere with clear physical meaning. ; Combining solar irradiance, solar zenith angle, and Earth-Sun distance, the radiance is converted into atmospheric top reflectivity. Then, using atmospheric correction algorithms from ACOLITE, 6S, L2gen, or the sensor itself, the effects of Rayleigh scattering, aerosol scattering, and solar flares are subtracted to obtain the water body's radiance away from the water surface. and downward irradiance above the water surface , and according to Calculate the remote sensing reflectance of water bodies For sensors that have already provided secondary water color products, the Rrs product can be read directly after quality control and water masking; finally, the Rrs of each band are reprojected and resampled to a unified spatial grid as input for subsequent spectral index calculations and models.

[0025] 1.2 Meteorological Data Preprocessing: Select core meteorological indicators such as temperature, wind speed, wind direction, and precipitation for the target time period. Use Kriging interpolation or bilinear interpolation to project the discrete station meteorological data onto a spatial grid that perfectly matches the remote sensing imagery. The specific process includes: ① Convert meteorological data and remote sensing images to the same coordinate reference system; ② Define the target grid using the coordinates of the upper left corner, spatial resolution, number of rows and columns, and coverage area of ​​the preprocessed remote sensing image; ③ Calculate the instantaneous, daily average, or previous day's cumulative values ​​of corresponding meteorological variables according to the time of remote sensing image transit; ④ For discrete station data, use ordinary kriging or inverse distance weighting to estimate variable values ​​at the center coordinates of the target grid; for regular grid data such as ERA5, use bilinear interpolation to resample to the target grid; ⑤ Use water masking to remove land areas and check the interpolation results for boundaries and outliers. Further decompose wind speed into an eastward U component and a northward V component. According to the meteorological definition of wind direction, wind direction... Indicates the direction of the wind and measures clockwise from due north. ,in For wind speed; if the data source uses the wind direction definition, then remove the negative sign accordingly. The decomposed U and V components, along with other meteorological variables, serve as model inputs.

[0026] 1.3 Water quality data preprocessing: Select chlorophyll a ( Core water quality indicators such as total nitrogen, total phosphorus, nitrogen-to-phosphorus ratio, transparency, conductivity, and suspended solids concentration are analyzed. The interquartile range method is used to remove outliers in the water quality data to avoid interference from monitoring errors. Based on the imaging date of the remote sensing image, water quality monitoring data with matching time are selected, and missing data are filled in using linear interpolation to achieve accurate matching of the time dimension between water quality data and remote sensing data.

[0027] 1.4. Resample all preprocessed spatial variables into the same H×W grid. Each remote sensing band, spectral index, meteorological variable, or water quality auxiliary variable is treated as a feature channel, and these channels are stacked along the channel dimension, i.e., stitched together along the Channels dimension on a unified H×W grid to construct a single sample. .

[0028] For example, using 6 remote sensing reflectance bands + NDCI and FAI spectral indices + 4 meteorological channels (sea surface temperature SST, wind speed U / V component, and precipitation) + 4 water quality channels (total nitrogen, total phosphorus, nitrogen-phosphorus ratio, and transparency), with Channels=C=16, each channel corresponds to the same spatial location at the same pixel position. The key principle for training data is that the measured chlorophyll a value of the current period can only be used as the target for the model to learn and predict, and cannot be used as input value into subsequent models. If you want the model to refer to past chlorophyll a conditions, you can only use data from the previous day or earlier, not the data of the current day.

[0029] After combining Batch_size samples, the dimensions of the multi-source data cube X are [Batch_size, Channels, Height, Width], representing the number of batch samples, the number of feature channels, the number of image grid rows, and the number of image grid columns, respectively. Height and Width are both represented by the number of pixels, and this dimensional order is consistent with the NCHW input format commonly used in convolutional neural networks.

[0030] Step Two: Construct a map of the suitability index for the aquatic ecosystem (e.g., ... Figure 3 (shown) 2.1 Screening of Driving Factors: Core driving factors affecting algal growth were screened, including temperature factor (T), nitrogen factor (N), phosphorus factor (P), algal sensitivity index (A), wind speed factor (W), and hydrodynamic / transparency comprehensive factor (H). Specifically, T is obtained from monitored or remotely sensed values ​​of water temperature and sea surface temperature; N is obtained by normalizing one or more of total nitrogen and inorganic nitrogen; P is obtained by normalizing one or more of total phosphorus and phosphate; A is not a directly measured single indicator, but a derived spectral characteristic calculated based on remotely sensed water reflectance, specifically obtained by weighting at least two of the normalized differential chlorophyll index (NDCI), phytoplankton index (FAI), maximum chlorophyll index (MCI), and red edge ratio index (REI); W is obtained from measured or reanalyzed wind speed, or can be calculated from U and V components; H is a derived comprehensive factor, calculated by normalizing one or more of flow velocity, mixing intensity, and transparency.

[0031] The specific construction method of the algae sensitivity index A is as follows:

[0032] In the formula, to For the index portfolio weights, the default value is... Alternatively, it can be determined according to the absolute value of the Spearman correlation coefficient between each index and the measured values ​​of chlorophyll a in the training set, i.e. .

[0033] Calculation of raw values ​​for Normalized Differential Chlorophyll Index (NDCI), Phytoplankton Index (FAI), Maximum Chlorophyll Index (MCI), and Red Edge Ratio Index (REI): ; ; ; The raw index values ​​of the Normalized Differential Chlorophyll Index (NDCI), Phytoplankton Index (FAI), Maximum Chlorophyll Index (MCI), and Red Edge Ratio Index (REI) are obtained through the above calculation formulas. These raw values ​​are then normalized to obtain... , , , ; In the formula, Represents remote sensing reflectance. Represents the red-edge band. Represents the red light band. Representing the near-infrared band, In the shortwave infrared band, 709, 681, and 753 are specific center wavelength values ​​for the band, in nanometers. The meaning of the Normalized Differential Chlorophyll Index (NDCI) calculation formula is: it is calculated by the ratio of the reflectance difference between two red-edge / near-infrared characteristic bands to the sum of the values. ε is a very small positive number used to avoid the denominator being 0. The meaning of the Red Edge Ratio Index (REI) calculation formula is as follows: It belongs to the ratio type index and is calculated by the ratio of the remote sensing reflectance of the red edge characteristic band to the reflectance of the reference band. ε is added to the denominator to avoid division by zero. The meaning of the formula for calculating the phytoplankton index (FAI) is as follows: It is a baseline-corrected algae index, which uses the reflectance of the red light band and the short-wave infrared band to construct a baseline, and calculates the deviation of the near-infrared band reflectance from the baseline. The maximum chlorophyll index (MCI) calculation formula means that it belongs to the baseline correction index. The baseline is constructed using the reflectance of the two bands of 681nm and 753nm, and the difference between the reflectance of the 709nm characteristic band and the baseline is calculated.

[0034] When the sensor does not have the corresponding center wavelength, the closest red light, red edge, near-infrared or short-wave infrared bands are used as substitutes.

[0035] 2.2 Constructing an ecological environment suitability index map: For promoting factors (temperature T, nitrogen N, phosphorus P, and algae sensitivity index A), forward normalization was used; for inhibiting factors (wind speed W and hydrodynamic / transparency combined factor H), reverse normalization was used. The forward and reverse processing procedures are as follows: For the promoting factor x, ; For the inhibitory factor x, ; In the formula, and Determined by training samples or a pre-defined reasonable range. To prevent extremely small positive numbers with a denominator of zero. For example, if the reasonable temperature range is 15-30℃ and the temperature of a certain pixel is 27℃, then the forward normalization value is (27-15) / (30-15)=0.80; if the reasonable wind speed range is 0-8m / s and the wind speed of a certain pixel is 2m / s, then the reverse normalization value is (8-2) / (8-0)=0.75.

[0036] Finally, a weighted fusion method is used to generate an ecological environment suitability index map with values ​​ranging from [0,1]. .

[0037]

[0038] In the formula, Represents the normalized mapping function. to This indicates the corresponding weighting coefficient. The weighting coefficients can use the default equal-weighting setting. Alternatively, a data-driven approach can be used to determine the factor, which involves calculating the Spearman correlation coefficient between each factor and the measured concentration of chlorophyll a in the training set or the algal species label. and order When there are insufficient training samples or the correlation coefficient cannot be stably calculated, the default equal weighting method is used.

[0039] By using the above weighted combination, an ecological environment constraint index map with values ​​ranging from [0,1] is generated. .

[0040] The specific functions of the ecological environment suitability index map are as follows: (1) Through weighted fusion ( to The final output is [0,1], which directly reflects the overall support of the environmental conditions at that location for the overall growth of algae: 1 represents the optimal environment and 0 represents the worst environment.

[0041] (2) Ecological environment suitability index map In the ecological mechanism constraint loss function proposed below, this index map is used to construct two types of penalty terms, laying the foundation for the construction of subsequent penalty terms.

[0042] (3) The ecological environment suitability index map and the prior map proposed in step three form complementary constraints. The suitability index map is responsible for “total amount is correct” (the concentration cannot deviate from the environmental carrying capacity), while the prior map is responsible for “species is correct” (the dominant algal species should conform to ecological rules). Together, they constitute a multi-dimensional ecological compliance verification of the model output results.

[0043] Step 3: Construct a priori map of algal species classification and chlorophyll a (e.g.) Figure 4 (See the construction flowchart) An ecological rule base tailored to the growth mechanisms of different algae was constructed, and the rule base was expressed in the form of "preconditions-conclusions". Before rule discrimination, temperature T, total phosphorus TP, wind speed W, transparency SD, nutrient comprehensive factor NUT, and hydrodynamic factor HD were normalized to the [0,1] interval, and discrimination was performed according to the following numerical standards: Defined as high temperature, Defined as medium temperature, Defined as high total phosphorus, Defined as low wind speed, Defined as high transparency, Defined as low transparency, Defined as high in nutrients, Defined as low in nutritional value Defined as weak hydrodynamics; Specifically, there are three rules: Rule 1 (Cyanobacterial Advantage): When , and When the location is determined to meet the a priori conditions for cyanobacterial growth dominance, in the preferred embodiment, the corresponding water temperature or sea surface temperature T≥25℃, total phosphorus TP≥0.10mg / L and wind speed W≤3m / s.

[0044] Rule 2 (Diatom or Green Algae Advantage): When , and When the spatial location satisfies the prior conditions for diatom or green algae dominance, in the preferred embodiment, the conditions are 15℃≤T<25℃, transparency SD≥1.0m and nutrient salt comprehensive factor NUT<0.40.

[0045] Rule 3 (Chlorophyll a increases): When , and When the spatial location is determined to meet the prior condition of increased chlorophyll a concentration, in the preferred embodiment, the corresponding nutrient comprehensive factor NUT≥0.70, transparency SD≤0.5m and flow velocity v≤0.10m / s or normalized hydrodynamic intensity HD≤0.30.

[0046] The current environmental conditions are input and matched against each rule in the rule base, and the matching degree of each rule is calculated: ; In the formula, Let represent the number of preconditions that the j-th rule satisfies at position i. The matching degree represents the total number of prerequisites for this rule, i.e., the matching degree = satisfying prerequisites / total number of prerequisites. The value range is [0,1], used to implement soft activation of rules. Soft activation of rules is achieved by calculating the matching degree, avoiding information loss caused by hard thresholds. An example is shown below: For rule 1, if a cell satisfies and But not satisfied If it satisfies two of the three prerequisites, then the matching degree is... =2 / 3=0.67; If all three preconditions are met, then =1; if none of these conditions are met, then =0.

[0047] The above problems are solved by calculating the matching degree, achieving partial matching and outputting a continuous value in the range [0,1], reflecting the degree of conformity between environmental conditions and rules. For example, if a pixel meets two of the three prerequisites, the matching degree is 0.67, indicating that there is a moderate tendency for cyanobacteria to dominate at that location. Soft activation based on matching degree is more refined and reasonable than hard threshold (either 1 or 0).

[0048] Further weighted aggregation of the matching degrees of each rule yields a priori graph. The weighted aggregation formula is as follows: ; In the formula, This represents the weight coefficient of the j-th rule. The weight coefficients are assigned equal weights by default, meaning that when the rule base contains 3 rules... Alternatively, the accuracy of the training set can be used to determine this, i.e. ,in The proportion of rule conclusions for the j-th rule on activated samples in the training set that are consistent with the actual algal species label or chlorophyll a concentration level.

[0049] By weighted aggregation and fusion of multiple rules, a composite prior is generated. Environmental conditions at the same spatial location may partially match multiple rules simultaneously, for example: Rule 1 (Cyanobacterial dominance) matching degree = 0.7; Rule 2 (diatom / green algae dominance) matching degree = 0.4; Rule 3 (Chlorophyll a increases) matching degree = 0.6.

[0050] At this point, a single rule cannot simply represent the state of that location; rather, the contributions of all rules need to be considered. Weighted aggregation achieves this goal: Multiply the matching degree of each rule by its weight coefficient. (Reflecting the prior credibility or importance of the rule), and then summed.

[0051] The output can be multi-dimensional (e.g., outputting the prior probability of cyanobacteria, the prior of chlorophyll a level, etc.), or it can be compressed into a single comprehensive index.

[0052] Using the above method, a priori diagram of algal species classification and chlorophyll a is generated. This is used for rule consistency constraints in subsequent model training.

[0053] Step 4: Construct a joint identification network (e.g., Figure 6 (shown) A joint recognition network for algal species classification and chlorophyll a concentration was constructed. This joint recognition network serves as the core carrier of the model, and the network outputs the algal species classification probability. Continuous values ​​of chlorophyll a concentration The identification network comprises three core structures: an encoder, an algal species classification head, and a chlorophyll a concentration regression head. 1. The shared encoder is used to extract spatial features, spectral features and environmental features from a multi-source data cube X. The encoder adopts a convolutional Swin-Transformer hybrid architecture: the convolutional layer extracts local spatial detail features of the multi-source data, and the Transformer module models the global feature association relationship to output a unified multi-source shared feature representation. 2. The algae classification head receives the feature map output from the shared encoder and outputs a probability vector for each pixel or image patch belonging to a different algae species. ; 3. The chlorophyll a concentration regression head receives shared features and outputs a continuous chlorophyll a concentration graph. .

[0054] Step 5: Construct an ecological mechanism-constrained loss function ( Figure 5 (shown) In addition to the basic classification loss and concentration regression loss, an ecological mechanism constraint penalty term is introduced to affect the algal species classification results, chlorophyll a concentration measurement results, and ecological environment constraint map. and rule prior graph Maintain consistency between them.

[0055] Loss by algal species classification Regression loss of chlorophyll a concentration Using the basic loss term and combining four types of ecological mechanism constraint penalty terms, a composite loss function integrating data-driven and mechanism constraint is constructed to achieve ecological mechanism compliance constraints in model training.

[0056] 1. Construction of basic loss terms: The specific construction process of basic loss terms is as follows: Loss function for algal species classification task: The algae classification head first outputs the probability that the i-th sample belongs to the k-th algae species using the Softmax function. The true label for each algal species uses an independent code. For K-type algae, multi-class cross-entropy loss is used: ; In the formula, The total number of samples within the batch. For category weights, For real labels, when the first The sample belongs to the first The value is 1 if the condition is met, otherwise it is 0.

[0057] When the number of samples from different algal species is unbalanced, FocalLoss is used instead of multi-class cross-entropy loss, i.e., multiplying the cross-entropy term by a factor. To reduce the influence of easily classified samples, the calculation formula is as follows: , where γ is the focusing parameter (usually taken as 2), used to control the suppression strength for easily classified samples.

[0058] Loss function for chlorophyll a concentration regression task: Chlorophyll a concentration regression head output Predicted concentration values ​​for each sample The corresponding on-site measured concentration is Mean square error When there are multiple outliers in the measured data, Huber loss is used instead of mean square error loss, and quadratic error and linear error are calculated in segments with the error threshold δ.

[0059] Before training, only training set statistics are used to normalize the chlorophyll a labels. After inference, a denormalization operation is performed to convert the predicted values ​​back to their original concentration units. The basic loss is written as... ,in The dimensions and gradient size used to balance classification and concentration regression tasks.

[0060] 2. Construction of Ecological Mechanism Constraint Penalty Terms: Four types of mechanism constraint penalty terms are added to correct the mechanism violation problem caused by the pure data-driven model. To avoid using ambiguous expressions, this embodiment defines the penalty terms with clear thresholds, allowable deviations, and calculable formulas.

[0061] (1) Ecological suitability promotion penalty item When the ecological suitability of the water area When, if the model's overall response value Below Imposing a penalty, calculated as follows: ,in , Let be the predicted probability that the i-th spatial location belongs to the k-th algal species. Predicted concentration of chlorophyll a after normalization; (2) Ecological inhibition constraint penalty items When the aquatic ecosystem is suitable When, if the chlorophyll a normalized predicted concentration Higher than If so, a penalty will be imposed, calculated as follows: ; (3) Rule consistency penalty When the model's overall response value With the rule prior graph The difference between them exceeds the allowable deviation When a penalty is imposed, the calculation formula is: ; (4) Classification-Concentration Consistency Penalty When algae classification results Results with chlorophyll a concentration The difference in the level mapping exceeds the allowable deviation. When a penalty is imposed, the calculation formula is: ,in This represents the algal biomass level obtained from the algal species classification probability mapping. This represents the concentration level obtained by mapping the normalized concentration of chlorophyll a.

[0062] , , , The tolerance deviation coefficient is determined through a validation set within the range of 0.01-0.10, with a preferred value of 0.05. This indicates the total number of samples. For the first Index of each sample.

[0063] 3. Construction of the total loss function: Integrating the basic loss term and various penalty terms to form the final total loss function: ; Where α is the weighting coefficient of the regression loss for chlorophyll a concentration, and λ1, λ2, λ3, and λ4 are the weighting coefficients of the penalty terms for the four types of mechanisms. α is preferably selected from {0.1, 0.5, 1, 2, 5}, and λ1, λ2, λ3, and λ4 are preferably selected from {0.01, 0.05, 0.1, 0.5, 1}, respectively. Before training, grid search, random search, or Bayesian optimization is performed using the validation set, and a comprehensive evaluation function is used. Determine the optimal weight combination. Where F1 is the F1 score for algal species classification. β1 and β2 are the normalized root mean square error of chlorophyll a concentration, Vio is the mechanism violation rate, and β1 and β2 are the balance coefficients of the evaluation function, preferably both set to 1. Once determined, the weights are fixed for model training, and the joint identification network parameters are updated through backpropagation of the gradient of the total loss function.

[0064] The logical relationship for defining the loss function in this step is: build the network → define the loss → train. The core function of this loss function is to integrate ecological mechanism knowledge (environmental suitability, ecological rules, algae-concentration consistency) into the training process of deep learning models in a differentiable mathematical form, forcing the model to learn mapping relationships that conform to natural laws, thereby improving the ecological rationality, internal consistency and interpretability of the inversion results.

[0065] Step Six: Model Training and Measurement Results Output Algal species category labels and chlorophyll a concentration labels were set for the training samples. Algal species category labels can be obtained through manual microscopy identification, historical survey data, laboratory test results, or expert interpretation; chlorophyll a concentration labels can be obtained through on-site sampling and laboratory analysis.

[0066] The joint recognition network was trained using training samples labeled with algal species and measured chlorophyll a concentrations. In each training round, an ecological constraint map was first generated. Algal species classification / chlorophyll a priori diagram Then, the joint identification network outputs the algal species classification results. Results of chlorophyll a concentration The loss value is calculated based on the total loss function, and backpropagation and parameter updates are performed using the Adam or SGD optimizer.

[0067] After training, the multi-source data of the water area to be tested is input into the model, which outputs an algal species classification map, a chlorophyll a concentration distribution map, and the measurement / analysis results (shown in...). Figure 8 The system can further analyze the area percentage, average chlorophyll a concentration, maximum concentration, spatial distribution range, and location of high-concentration areas for each algal species, providing technical support for water composition analysis and water environment monitoring and management.

[0068] This embodiment proposes a hyperspectral identification system for algal community structure during algal blooms, which implements the hyperspectral identification method for algal community structure during algal blooms proposed above. It can be deployed on ecological environment monitoring platforms, remote sensing operational processing platforms, and aquatic ecological risk early warning platforms to achieve automated and operational monitoring and decision support for algal blooms in lakes, reservoirs, estuaries, and nearshore seas. Specifically, it includes the following functional modules: I. Data Acquisition and Preprocessing Module The data acquisition and preprocessing module is the basic data processing unit at the front end of the system. It integrates the functions of raw data acquisition and access and standardized preprocessing of multi-source data, and communicates and connects with the back-end modules of the system.

[0069] This module is used to collect and access multi-source basic data of the target monitoring area, specifically including multispectral remote sensing image data, ERA5 reanalysis meteorological data, and ground monitoring station water quality data; at the same time, it performs full-process standardized preprocessing on the acquired multi-source raw data, which is used for radiometric calibration, atmospheric correction, and water body mask extraction of remote sensing images, and outlier removal, interpolation, time matching and normalization of meteorological and water quality data. II. Data Processing and Intelligent Recognition Module The data processing and intelligent recognition module is the core computing unit of the system, integrating the ecological suitability construction module, the prior map construction module, the joint recognition module, and the ecological mechanism loss constraint module.

[0070] The ecological environment constraint map construction module is used to construct an ecological environment constraint map based on temperature, nitrogen, phosphorus, algae sensitivity index, wind speed, and hydrodynamic / transparency factors. ; The prior graph construction module is used to construct graphs based on the ecological rule base and current environmental conditions. ; The joint identification module is used to output algal species classification results and continuous chlorophyll a concentration measurement results using a deep learning network; The loss constraint module is used to construct the ecological mechanism constraint loss function and participate in parameter optimization during the model training phase; III. Result Output Module The results output module is used to output algal species classification diagrams, chlorophyll a concentration distribution diagrams, and corresponding statistical analysis results.

[0071] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A hyperspectral identification method for algal community structure during algal blooms, characterized in that, include: Step 1: Multi-source data acquisition and preprocessing Collect multi-source data of the target water area and perform preprocessing. Then, stack the preprocessed data according to the channel dimension to construct a multi-source data cube X. Step 2: Constructing a map of aquatic ecological environment suitability indicators Driving factors were screened and normalized, and promoting and inhibiting factors were distinguished. Forward and reverse mapping processes were applied respectively, and a weighted fusion algorithm was used to generate an ecological environment suitability index map. ; Step 3: Construct a priori map of algal species classification and chlorophyll a Construct a growth mechanism ecological rule base, input the current environmental conditions to calculate the rule matching degree, and weight and aggregate the prior graph. ; Step 4: Construct a joint identification network X is input into a joint recognition network constructed from an encoder, an algal species classification head, and a chlorophyll a concentration regression head. The recognition network outputs the algal species classification probability. Continuous values ​​of chlorophyll a concentration ; Step 5: Construct an ecological mechanism-constrained loss function Loss by algal species classification Regression loss of chlorophyll a concentration A composite loss function is constructed by superimposing four types of ecological mechanism constraint penalty terms on the basic loss term, thereby affecting the algal species classification results, chlorophyll a concentration measurement results, and ecological environment constraint map. and rule prior graph Maintain consistency between them; Step Six: Model Training and Measurement Results Output The joint recognition network is trained using training samples. After training, multi-source data is input into the model, and the model outputs an algal species classification map, a chlorophyll a concentration distribution map, and measurement results. In step one, the multi-source data includes multispectral remote sensing image data or hyperspectral remote sensing reflectance data, meteorological data, and water quality data from ground monitoring stations; In step five, the four types of ecological mechanism constraint penalty terms are as follows: Ecological suitability promotes punishment items : ,in Used to constrain the ecological suitability of aquatic waters When the model output is lower than In the case of, Let be the predicted probability that the i-th spatial location belongs to the k-th algal species. Predicted concentration of chlorophyll a after normalization; Ecological inhibition constraint penalty items : Used to constrain the ecological suitability of aquatic waters At that time, the predicted concentration of normalized chlorophyll a was higher than The situation; Rule consistency penalty : Used to constrain the comprehensive output of the model With the rule prior graph The difference between them shall not exceed the allowable deviation. ; Category—Concentration Consistency Penalty ,in This represents the algal biomass level obtained from the algal species classification probability mapping. This represents the concentration level obtained by mapping the normalized concentration of chlorophyll a, used to constrain the biomass level corresponding to the classification output, ensuring that the deviation from the level corresponding to the chlorophyll concentration does not exceed the allowable deviation. ; , , , The tolerance coefficient is determined by the validation set within the range of 0.01-0.10; N represents the total number of samples, and i is the index of the i-th sample.

2. The hyperspectral identification method for algal community structure during algal blooms according to claim 1, characterized in that: Meteorological data include water temperature or sea surface temperature, wind speed, wind direction, and precipitation; ground monitoring station water quality data include one or more of the following: chlorophyll a, total nitrogen, total phosphorus, nitrogen-to-phosphorus ratio, transparency, electrical conductivity, and suspended solids concentration. Remote sensing image data preprocessing: Radiometric calibration, atmospheric correction, and geometric correction are performed sequentially on the original multispectral remote sensing images to eliminate sensor errors and atmospheric scattering interference; invalid areas, including land and clouds, are removed using water body mask extraction technology and cloud shadow removal technology to accurately preserve water image data; and the water remote sensing reflectance is obtained based on the preprocessed water image. Meteorological data preprocessing: The meteorological data from discrete stations are projected onto a spatial grid that is completely consistent with the remote sensing image using Kriging interpolation or bilinear interpolation to achieve spatial resolution matching between the meteorological data and the remote sensing data; the wind speed data is further decomposed into horizontal U-components and vertical V-components to improve the hydrodynamic meteorological characteristic parameters. Water quality data preprocessing: Core water quality indicators were selected, including chlorophyll a, total nitrogen, total phosphorus, nitrogen-to-phosphorus ratio, transparency, conductivity, and suspended solids concentration. The interquartile range method was used to remove outliers in the water quality data to avoid interference from monitoring errors. Based on the imaging date of the remote sensing image, water quality monitoring data with matching time were screened, and missing data were filled in using linear interpolation to achieve accurate matching of the time dimension between water quality data and remote sensing data.

3. The hyperspectral identification method for algal community structure during algal blooms according to claim 1, characterized in that: The promoting factors include temperature factor T, nitrogen factor N, phosphorus factor P, and algal sensitivity index A. Among them, algal sensitivity index A is obtained by normalizing and weighting at least two of the following: normalized differential chlorophyll index NDCI, phytoplankton index FAI, chlorophyll index MCI, and red edge ratio index REI. The higher the normalized value of the promoting factor, the stronger its supporting effect on algal growth or chlorophyll a concentration. The inhibition factors include wind speed factor W and hydrodynamic / transparency composite factor H; where H is obtained by normalizing at least one of flow velocity, mixing intensity and transparency, and the higher the normalized value of the inhibition factor, the stronger the water disturbance or light inhibition effect.

4. The hyperspectral identification method for algal community structure during algal blooms according to claim 1, characterized in that: In step three, the ecological rule base is represented in the form of "preconditions-conclusions". Before rule discrimination, temperature T, total phosphorus TP, wind speed W, transparency SD, nutrient comprehensive factor NUT, and hydrodynamic factor HD are normalized to the [0,1] interval, and discrimination is performed according to the following numerical standards: Defined as high temperature, Defined as medium temperature, Defined as high total phosphorus, Defined as low wind speed, Defined as high transparency, Defined as low transparency, Defined as high in nutrients, Defined as low in nutritional value Defined as weak hydrodynamics; Specifically, there are three rules: Rule 1: When , and When the spatial location is determined to meet the prior condition for cyanobacterial growth dominance; Rule 2: When , and When the spatial location satisfies the prior condition of diatom or green algae dominance, it is determined that the location satisfies the prior condition of diatom or green algae dominance. Rule 3: When , and When the spatial location satisfies the prior condition of increased chlorophyll a concentration, it is determined that the location satisfies this condition.

5. The hyperspectral identification method for algal community structure during algal blooms according to claim 1, characterized in that: In step four, the encoder is a shared encoder, which is used to extract spatial features, spectral features and environmental features from the multi-source data cube X. The encoder adopts a convolutional Swin-Transformer hybrid architecture: the local spatial detail features of the multi-source data are extracted through the convolutional layer, and the global feature association relationship is modeled through the Transformer module to output a unified multi-source shared feature representation. The algae classification head receives the feature map output from the shared encoder and outputs a probability vector for each pixel or image patch belonging to a different algae species. ; The chlorophyll a concentration regression head receives shared features and outputs a continuous value graph of chlorophyll a concentration. .

6. The hyperspectral identification method for algal community structure during algal blooms according to claim 1, characterized in that: In step six, the model is trained using training samples labeled with algal species and measured chlorophyll a concentrations. In each round of training, the ecological environment suitability index map is first invoked. Algal species classification / chlorophyll a priori diagram Then the network outputs the algal species classification results. Results of chlorophyll a concentration The loss value is calculated based on the total loss function, and backpropagation and parameter updates are performed using the Adam or SGD optimizer.

7. A hyperspectral identification system for algal community structure during algal blooms, used to implement the hyperspectral identification method for algal community structure during algal blooms as described in any one of claims 1-6, characterized in that... include: Data acquisition and preprocessing modules The data acquisition and preprocessing module is the basic data processing unit at the front end of the system. It integrates the functions of raw data acquisition and access and standardized preprocessing of multi-source data, and communicates and connects with the back-end modules of the system. This module is used to collect and access multi-source basic data of the target monitoring area, specifically including multispectral remote sensing image data, ERA5 reanalysis meteorological data, and ground monitoring station water quality data; at the same time, it performs full-process standardized preprocessing on the acquired multi-source raw data, which is used for radiometric calibration, atmospheric correction, and water body mask extraction of remote sensing images, and outlier removal, interpolation, time matching and normalization of meteorological and water quality data. II. Data Processing and Intelligent Recognition Module The data processing and intelligent recognition module is the core computing unit of the system, integrating the ecological suitability construction module, the prior map construction module, the joint recognition module, and the ecological mechanism loss constraint module; The ecological environment constraint map construction module is used to construct an ecological environment constraint map based on temperature, nitrogen, phosphorus, algae sensitivity index, wind speed, and hydrodynamic / transparency factors. ; The prior graph construction module is used to construct graphs based on the ecological rule base and current environmental conditions. ; The joint identification module is used to output algal species classification results and continuous chlorophyll a concentration measurement results using a deep learning network; The loss constraint module is used to construct the ecological mechanism constraint loss function and participate in parameter optimization during the model training phase; III. Result Output Module The results output module is used to output algal species classification diagrams, chlorophyll a concentration distribution diagrams, and corresponding statistical analysis results.

Images