Remote sensing estimation method and system for total algae content based on vertical pre-classification and deep learning

Abstract

The present invention relates to the technical field of remote sensing monitoring of water ecological environment, and discloses a remote sensing estimation method and system for the total amount of algae based on vertical pre-classification and deep learning, wherein the method comprises: obtaining the inversion result of chlorophyll a concentration of the surface of the water body; obtaining the chlorophyll a profile distribution type based on the inversion result of chlorophyll a concentration of the surface of the water body and water environment information; performing vertical multi-layer continuous chlorophyll a concentration and total amount of algae inversion based on the chlorophyll a profile distribution type and the inversion result of surface chlorophyll a concentration, and obtaining the estimation result of the total amount of algae. The method combines the advantages of phytoplankton profile monitoring and deep learning algorithms, and can perform stable and high-precision estimation of the total amount of algae in water bodies with large land-based influences, complex vertical distribution of phytoplankton and complex optical characteristics. The present invention will provide more effective technical support for the research on remote sensing estimation of the total amount of algae, timely discover and warn the occurrence of harmful algal blooms, and reduce the impact of algal bloom disasters.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing monitoring of aquatic ecological environment, and in particular to a method and system for remote sensing estimation of total algae content based on vertical pre-classification and deep learning. Background Art

[0002] Traditional methods for estimating the total amount of algae mainly rely on sampling analysis and remote sensing monitoring of the optical characteristics of phytoplankton. Sampling and analysis of the total amount of algae has the advantages of being direct, fast, and accurate, but its temporal and spatial representativeness is limited and it is difficult to reflect its dynamic changes. Satellite remote sensing has a unique perspective and large-scale synchronous observation capabilities, and can quickly, efficiently, and periodically observe the temporal and spatial changes of phytoplankton. However, most traditional remote sensing estimation models for the total amount of algae are based on the assumption that algae are vertically uniform, ignoring the actual situation of non-uniform distribution of algae, which brings large errors to the estimation of algal biomass. Class II water bodies such as estuaries, nearshore areas, lakes, and reservoirs are greatly affected by terrestrial sources, and the vertical distribution and optical characteristics of phytoplankton are complex. Existing remote sensing estimation methods for the total amount of algae are difficult to achieve stable and high accuracy in long-term, large-scale remote sensing monitoring. Summary of the invention

[0003] In order to address the deficiencies of the prior art, the present invention provides a method and system for remote sensing estimation of the total amount of algae based on vertical pre-classification and deep learning. Based on multi-source remote sensing data, monitoring data of phytoplankton-related ecological indicators (surface and vertical multi-layer chlorophyll a concentration, biomass, etc.) and deep learning algorithms, a remote sensing inversion model of the total amount of algae that takes into account vertical non-uniform distribution is developed. This method combines the advantages of phytoplankton profile monitoring and deep learning algorithms, and can perform stable and high-precision estimation of the total amount of algae in Class II water bodies that are greatly affected by terrestrial sources and have complex vertical distribution and optical characteristics of phytoplankton. The application of this method and system will provide more effective technical support for the research on remote sensing estimation of the total amount of algae, timely discover and warn the occurrence of harmful algal blooms, and reduce the impact of algal bloom disasters.

[0004] On the one hand, a remote sensing estimation method for the total amount of algae based on vertical pre-classification and deep learning is provided, including: obtaining the chlorophyll a concentration inversion results of the water body surface; obtaining the chlorophyll a profile distribution type based on the chlorophyll a concentration inversion results of the water body surface and water environment information; performing vertical multi-layer continuous chlorophyll a concentration and total amount of algae inversion based on the chlorophyll a profile distribution type and the chlorophyll a concentration inversion results to obtain the estimation result of the total amount of algae.

[0005] On the other hand, a remote sensing estimation system for the total amount of algae based on vertical pre-classification and deep learning is provided, including: a first acquisition module, which is configured to obtain the chlorophyll a concentration inversion result of the water body surface; a second acquisition module, which is configured to obtain the chlorophyll a profile distribution type based on the chlorophyll a concentration inversion result of the water body surface and water environment information; an estimation module, which is configured to perform vertical multi-layer continuous chlorophyll a concentration and total amount of algae inversion based on the chlorophyll a profile distribution type and the chlorophyll a concentration inversion result, to obtain the estimation result of the total amount of algae.

[0006] The above technical solution has the following advantages or beneficial effects: In order to obtain accurate vertical distribution of phytoplankton, the present invention simulates the vertical distribution characteristics of phytoplankton based on field test data and HydroLight numerical simulation. This model fully considers the actual situation of non-uniform vertical distribution of algae and takes multiple factors into account, thereby improving the simulation accuracy.

[0007] Based on field satellite-ground synchronous experiments and the HydroLight model, the present invention simulates and functionally characterizes the main vertical distribution of phytoplankton and its optical absorption and scattering characteristics, and utilizes the advantages of the Transformer deep learning algorithm in nonlinear fitting to improve the accuracy and applicability of the model.

[0008] Efficient and accurate data processing flow: Accurate and comprehensive vertical pre-classification can filter data and decompose it into smaller, simpler multi-level subsets, reducing the amount of data that the model needs to process at each calculation, thereby shortening processing time and improving computational efficiency.

[0009] Ability to capture spatiotemporal correlations: The spatiotemporal attention mechanism helps the model capture the dependencies between different time points and spatial locations. It can effectively analyze the spatiotemporal changes in algae concentrations in water bodies, track changing trends, identify periodic patterns and make trend predictions.

[0010] Comprehensive analysis capability of multivariate data: The inter-variable attention mechanism enables the model to understand the interaction between different input variables. This is especially important for processing multispectral remote sensing data or multi-sensor fusion data, because different variables may jointly affect the estimation results of the target variable. Through this mechanism, the model can consider various influencing factors more comprehensively and improve the accuracy of the estimation.

[0011] Accuracy and robustness: By simultaneously modeling complex relationships in spatial, temporal, and variable dimensions, the model's prediction accuracy can be greatly improved, especially in tasks with complex spatiotemporal changes, such as remote sensing estimation of algae abundance, which can more accurately estimate algae concentrations at different times and locations. At the same time, the model can handle complex spatiotemporal dependencies and multivariate interactions, and is more stable in different environments and data sets, with strong generalization capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0013] Figure 1 This is a flow chart of the method of embodiment 1.

[0014] Figure 2 This is a flow chart of obtaining the chlorophyll a concentration inversion result of the water surface in Example 1.

[0015] Figure 3 Schematic diagram of the internal structure of the Transformer model of Example 1.

[0016] Figure 4 This is a flow chart for obtaining the chlorophyll a profile distribution type based on the chlorophyll a concentration inversion result and water environment information of the water body surface in Example 1. DETAILED DESCRIPTION

[0017] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0018] With the continuous advancement of research, machine learning has become an important way to invert the total amount of algae with its powerful nonlinear fitting ability. Machine learning algorithms can help to extract effective information from massive remote sensing data to conduct remote sensing estimation of phytoplankton and its related ecological indicators, and realize long-term and large-scale monitoring of the total amount of algae. By continuously optimizing the algorithm, the generalization ability and robustness of the model can be improved, so that it can realize the monitoring of the total amount of algae on a larger scale and with higher spatiotemporal resolution, providing more scientific support for water resources management and aquatic environment protection.

[0019] Embodiment 1, as Figure 1 As shown, this embodiment provides a remote sensing estimation method for the total amount of algae based on vertical pre-classification and deep learning, including: S101: obtaining the chlorophyll a concentration inversion result of the water body surface; S102: based on the chlorophyll a concentration inversion result of the water body surface and water environment information, obtaining the chlorophyll a profile distribution type; S103: based on the chlorophyll a profile distribution type and the chlorophyll a concentration inversion result, performing vertical multi-layer continuous chlorophyll a concentration and total amount of algae inversion to obtain an estimation result of the total amount of algae.

[0020] Furthermore, if Figure 2As shown, the S101: obtaining the inversion result of the chlorophyll a concentration on the surface of the water body, specifically includes: (1-1): constructing a training set and a validation set, wherein the training set and the validation set are both the optimal characteristic band combination of the known chlorophyll a concentration on the surface of the water body; (1-2): using the optimal characteristic band combination as the input value of the Transformer model, using the known chlorophyll a concentration on the surface of the water body as the output value of the Transformer model, training the Transformer model, and obtaining a trained Transformer model; verifying the trained Transformer model, if the verification passes, it indicates that the current Transformer model is the final trained Transformer model; (1-3): inputting the optimal characteristic band combination of the water environment for estimating the total amount of algae into the final trained Transformer model, and obtaining the inversion result of the chlorophyll a concentration on the surface of the water body.

[0021] Furthermore, the (1-1): constructing a training set and a validation set specifically includes: (1-1-1): obtaining surface water sample data from different lakes and reservoirs; the surface water sample data from different lakes and reservoirs include: optical properties and water quality parameters; the optical properties include: absorption coefficient and scattering coefficient; the water quality parameters include: chlorophyll a concentration, suspended solids and colored soluble organic matter (CDOM); (1-1-2): performing preprocessing operations on the acquired data, and the preprocessing operations include: removing missing values ​​and outliers; (1-1-3): performing feature screening on the data after the preprocessing operation to obtain the optimal feature band combination, and obtaining the training set and the validation set according to the optimal feature band combination and the corresponding chlorophyll a concentration.

[0022] Furthermore, the (1-1-3) performs feature screening on the data after the preprocessing operation to obtain the optimal feature band combination, specifically including: (1-1-31) using continuous wavelet transform to extract the first feature band; (1-1-32) using correlation analysis to extract the second feature band; (1-1-33) concatenating the first feature band and the second feature band to obtain the optimal feature band combination.

[0023] Furthermore, (1-1-31) a continuous wavelet transform is used to extract the first characteristic band, specifically including: extracting data with a band of 400-900 nm from the data after the preprocessing operation, and using a continuous wavelet transform (CWT) to extract intrinsic and apparent optical features of different scales; selecting a band with a signal-to-noise ratio (SNR) greater than 5 dB, a mean square error (MSE) less than 0.1, and a p-value less than 0.05 with chlorophyll a concentration as the first characteristic band.

[0024] As a linear transformation method, wavelet analysis can obtain the time information and frequency characteristics of the signal by translating and scaling the mother wavelet basis of the original signal.

[0025] The original reflectance spectrum signal and the wavelet basis function are multiplied and accumulated in the entire interval, and finally a two-dimensional matrix wavelet coefficient is obtained, which contains scale and band information.

[0026] The wavelet coefficients obtained by transforming different wavelet basis functions and the measured chlorophyll concentration data were subjected to correlation analysis, and the wavelet coefficients with the highest correlation were extracted for machine learning modeling.

[0027] There are many components in water bodies and their optical properties are complex, but the optically active substances determine the reflectivity off water, which mainly include four types of substances: pure water, phytoplankton, non-pigmented suspended matter, and colored soluble organic matter. These substances affect the water body's emissivity off water through absorption and scattering. Among them, the optical information that does not change with the incident light source and angle is called the inherent optical properties (IOPs) of the water body; on the contrary, the optical quantities affected by the incident light source, incident angle, diffuse attenuation coefficient, etc. are called apparent optical properties (AOPs), which are the reflectivity off water. The most direct representation of .

[0028] , where It is the reflectivity above the water surface, usually less than 10%. is the sky reflectivity, usually calculated by the Fresnel formula.

[0029] Furthermore, (1-1-32) correlation analysis was used to extract the second characteristic band, specifically including: calculating the Pearson product-moment correlation coefficient Pearson for the reflectance of each band and the surface chlorophyll a concentration to evaluate the influence of the band on the concentration, and selecting the band with a correlation coefficient higher than 0.6 as the second characteristic band.

[0030] Furthermore, if Figure 3 As shown, the trained Transformer model includes: an input layer, a multi-head attention mechanism layer, a first layer normalization module (Layer Normalization), a feedforward neural network, a second layer normalization module and a decoder connected in sequence; the input end of the first layer normalization module is also connected to the input layer; the input end of the second layer normalization module is also connected to the input end.

[0031] It should be understood that in order to provide the necessary initial conditions and basic data for profile simulation, and help understand the optical properties and biological distribution of water bodies. The present invention analyzes the biological optical properties of different lakes and reservoirs based on surface in-situ observation data, and constructs a remote sensing inversion model of chlorophyll a concentration in the surface layer of water bodies based on the Transformer deep learning algorithm: based on Pearson correlation and continuous wavelet coefficients, the optimal band combination of Sentinel-2 MSI (MultiSpectral Instrument, Level-2A) remote sensing image inversion of surface chlorophyll a concentration is selected; the optimal band combination and surface chlorophyll a concentration are used as the input and output of Transformer deep learning, respectively, the Transformer model is trained, and hyperparameters such as the number of attention heads and dimensions are calibrated to obtain a remote sensing inversion model of surface chlorophyll a concentration.

[0032] Data collection: Collect surface water sample data from different lakes and reservoirs, including optical properties (such as absorption and scattering coefficients) and water quality parameters (such as chlorophyll a concentration, suspended matter, CDOM, etc.).

[0033] Data processing: Remove missing values ​​and outliers, ensure data quality, and interpolate and standardize the data.

[0034] Feature band screening: Connect the wavelet transform features and sensitive bands through feature splicing to form a complete feature matrix; perform linear transformation on the combined feature matrix to map the features to the dimensional space required by the model; add position coding to each feature sample.

[0035] Model input: Convert the preprocessed remote sensing image data into a three-dimensional tensor, which contains the optimal spectral feature (spectral), spatial (spatial), and temporal (temporal) information. Add position coding and add it to the feature matrix. The optimal spectral feature is the optimal feature band combination.

[0036] The overall architectural design of Transformer's STV Block: First, the multi-head self-attention mechanism is used to extract and fuse the input remote sensing data in spatial, temporal and spectral dimensions; then, a feedforward neural network is applied to process the feature representation of each position to further extract and optimize the features; finally, layer normalization and residual connections are used to stabilize the training process and improve model performance.

[0037] In order to make the model focus on different parts of the data in different subspaces at the same time, a multi-head attention mechanism is designed: Transformer's STV Block transforms the input data linearly. , ( is the biased output vector, is the weight matrix, is the input vector, is the bias vector) to calculate the Query (Q), Key (K) and Value (V) matrices respectively.

[0038] Split the resulting query, key, and value vectors into heads, each head has a dimension of the original , each head processes a subspace of the input, learns different feature subspaces, and calculates attention weights in spatial, temporal, and spectral dimensions.

[0039] Transformer uses the dot product attention mechanism: (1) Calculation and , and use these weights from Extract relevant information from: ,in, is the dimension of the Key vector. (2) The outputs of all heads are concatenated and then integrated into a final feature representation through linear transformation: ,in, is the weight matrix of the linear transformation; (3) add the output of the multi-head attention mechanism to the original input through the residual connection: ; (4) Perform layer normalization after the attention layer: .

[0040] In order to enhance the expressive power of the model, a feedforward network is designed to perform nonlinear transformation and mapping on the features. (1) The output obtained from the multi-head attention mechanism is used as input, and the input features are linearly transformed and mapped to a higher-dimensional space through a fully connected layer and an activation function is used. Introduce nonlinear features to obtain intermediate feature representation; (2) map the intermediate feature representation back to the original embedding dimension through a second fully connected layer.

[0041] (3) Add a residual connection between the input and output of the feedforward neural network: ; (4) Use layer normalization to normalize the results after residual connection.

[0042] (5) The output layer uses a fully connected layer to map the features extracted by the Transformer's STV Block to the final output dimension and convert the feature extraction results into the total amount of algae. The ReLU activation function is applied after the output layer to ensure that the output prediction value is non-negative.

[0043] Through the above steps, the model can handle complex spatiotemporal dependencies and multivariate interactions, and thus performs more stably in different environments and data sets, has stronger generalization capabilities, and effectively improves the accuracy and robustness of the predictions.

[0044] Model training: Use the feature matrix and algae monitoring data to perform multiple rounds of iterative training, and continuously adjust the hyperparameters in the model to the optimal level: Data preparation: De-noise and standardize the acquired algae monitoring data, construct a three-dimensional tensor based on the data's time sequence, image size, and number of spectral bands, and divide the data set into training, validation, and test sets.

[0045] Loss function: In model training, it is necessary to define a suitable loss function to measure the difference between the model prediction result and the true label in order to optimize the model parameters. In this regression task, the mean square error (MSE) is used as the loss function. Assume that the predicted value output by the model is , the actual value is , the sample size is , then the cross entropy loss function The expression is: .

[0046] Training: Input the training data into the model and perform forward propagation to calculate the predicted value. Calculate the error between the predicted value and the actual value based on the loss function. Calculate the gradient through the back propagation algorithm and update the model parameters. Repeat the above steps until the training converges or reaches the set number of iterations.

[0047] Hyperparameter tuning: Bayesian optimization is used to tune hyperparameters, model the relationship between hyperparameters and validation set performance, and gradually optimize the validation set performance of the model.

[0048] (1) Conduct preliminary experiments using common hyperparameter combinations (eight-head attention, attention dimension d_k, embedding dimension 256, number of layers 6, dropout probability 0.1, learning rate 0.0001, batch size 64, gradient clipping 0.1), and record their validation set performance.

[0049] (2) Use preliminary experimental results to train a Gaussian process model and learn the relationship between hyperparameters and validation set performance.

[0050] (3) Based on the Gaussian process model, calculate the prediction performance and uncertainty of all hyperparameter combinations.

[0051] (4) Select the hyperparameter combination that maximizes the expected improvement or probability improvement for the experiment, evaluate its performance on the validation set, and record the objective function value.

[0052] (5) Add the new hyperparameter combination and its corresponding performance results to the training data and update the Gaussian process model.

[0053] (6) Repeat steps (3)-(5), continue to select the next hyperparameter combination and conduct experiments until the preset maximum number of iterations or target performance is reached.

[0054] Model evaluation: The K-Fold Cross-Validation method is used to comprehensively evaluate the robustness and generalization ability of the model. The specific approach is to divide the data set into K subsets, select one of the subsets as the validation set in turn, and use the remaining subsets for training, repeat K times, and finally calculate the average performance index of each validation. This method can reduce the accidental impact caused by data division and ensure the reliability of the evaluation results. According to the evaluation results, the model can be optimized and improved to improve the performance of total algae estimation.

[0055] Furthermore, the method further comprises: using appropriate evaluation indicators (such as root mean square error, mean absolute error, The performance of the model in remote sensing quantitative monitoring and algae identification tasks is evaluated by using the following indicators: Root mean square error (RMSE): ; Mean absolute error (MAE): ; Coefficient of determination ( ): ;Relative square error (RSE): ;in, Represents the predicted value, represents the real value, is the sample size.

[0056] It should be understood that CWT is a time-frequency analysis method used to decompose signals and extract features from them. CWT can not only analyze the frequency components of the signal, but also retain the time (or space) information, so it is particularly suitable for analyzing signals with non-stationary characteristics such as spectral data.

[0057] The core of CWT is the wavelet function , which is a localized wave function with zero mean and finite energy: , The shape of determines the characteristics of CWT, which can be a Mexican hat wavelet, Morlet wavelet, or Haar wavelet, etc.

[0058] CWT by scaling and translating the wavelet function To analyze the signal: Scaling : Controls the width of the wavelet function, reflecting the frequency information. Smaller scale Corresponding to high frequency details, larger scales correspond to low frequency trends. : Controls the position of the wavelet function, reflecting the time or space position of the signal.

[0059] Through these transformations, the wavelet function is defined as: .

[0060] CWT is defined as: .

[0061] in, is the signal to be analyzed. is in scale and location The CWT coefficients under .

[0062] It should be understood that Transformer is a deep learning model architecture that excels in processing long sequences of data by parallel processing and capturing global dependencies. In the development of remote sensing estimation models for algae, STVTransformer Block can be used. STV Transformer Block is a specific Transformer module design that is often used to process complex data. Its multi-head self-attention mechanism can simultaneously consider the relationship between spatial position, temporal changes, and spectral variables. This can improve the model's ability to capture spatiotemporal correlations, analyze the spatiotemporal changes in algae concentrations in water bodies, and make trend predictions. Multi-head attention mechanism: The input data is transformed through multiple different linear transformations to obtain the 'Q' (query), 'K' (key), and 'V' (value) matrices. Each attention head independently calculates the attention weight and generates an output, and then the outputs of all heads are concatenated together and the final output is generated through linear transformation.

[0063] Furthermore, if Figure 4 As shown, the S102: based on the chlorophyll a concentration inversion result of the water body surface and the water environment information, the chlorophyll a profile distribution type is obtained, including: (2-1): based on the chlorophyll a concentration inversion result of the water body surface and the water environment information, a three-dimensional bio-optical database is constructed; (2-2): the remote sensing data is optimally spectrally matched with the three-dimensional bio-optical database; (2-3): the water column profile distribution function is determined, a decision tree model is established, and the chlorophyll a profile distribution type is determined through the decision tree model and the water column profile distribution function.

[0064] Furthermore, (2-1): based on the inversion results of chlorophyll a concentration in the water surface and water environment information, a three-dimensional bio-optical database is constructed, including: according to the water body bio-optical parameters, ecological environment parameters and the inversion results of chlorophyll a concentration in the water surface, the HydroLight radiation transfer model is used to simulate and construct a three-dimensional bio-optical database for multiple situations in different lakes and reservoirs.

[0065] The HydroLight radiative transfer model is a radiative transfer model used to simulate light transmission in water bodies, and is mainly used to study the propagation and reflection characteristics of light in water bodies. The model takes into account various factors in water bodies, such as the optical properties of water, the effects of suspended particles and dissolved substances on light, so that it can accurately simulate the distribution of light under different water depths and conditions. The HydroLight radiative transfer model is widely used in remote sensing and ecological research, helping scientists understand the optical properties of water bodies and their effects on organisms and the environment.

[0066] Furthermore, (2-1): based on the chlorophyll a concentration inversion results and water environment information of the water body surface, a three-dimensional bio-optical database is constructed, which specifically includes: (2-1-1): determining the water environment for estimating the total amount of algae, collecting data of different hydrological, meteorological and ecological environmental conditions, and the collected data include: water temperature, transparency and nutrient concentration; (2-1-2): determining the inherent optical quantity; the inherent optical quantity includes: particle backscattering probability, particle and non-pigment particle absorption coefficient; the inherent optical quantity is taken from several water environment lake biomass, respectively. The average value of the bio-optical survey; the HydroLight radiation transfer model is used to simulate the radiation transfer under different situations; the water characteristics such as bottom sediment type, colored soluble organic matter (CDOM) and inorganic suspended matter are respectively the average values ​​of the bio-optical survey of several water environments, and the surface chlorophyll concentration uses the inversion result of chlorophyll a concentration in the water surface; the simulated water depth is twice the maximum measured value; (2-1-3): Based on the collected data, the inherent optical quantity and the inversion result of chlorophyll a concentration in the water surface, a three-dimensional bio-optical database is established.

[0067] It should be understood that the three-dimensional bio-optical database includes the three elements of water color (chlorophyll a concentration, CDOM concentration and suspended matter concentration) of multi-layer water bodies, as well as the inherent optical characteristics of water bodies under optical conditions (absorption and scattering coefficients) of the three elements of water color.

[0068] Furthermore, the (2-2): optimally spectrally matching the remote sensing data with the three-dimensional bio-optical database includes: optimally matching the remote sensing reflectance of the Sentinel MSI (Multispectral Instrument) remote sensing image with the bio-optical database.

[0069] Furthermore, the (2-2): optimally spectrally matching the remote sensing data with the three-dimensional bio-optical database specifically includes: (2-2-1) acquiring remote sensing data of Sentinel MSI from the remote sensing data platform Copernicus Data Hub, wherein the acquired remote sensing data has the same time period and location as the bio-optical database.

[0070] It should be understood that the Copernicus Data Hub is the official data access platform that aims to provide high-quality Earth observation data, mainly from the Copernicus satellite series (Sentinel-1, Sentinel-2, Sentinel-3, etc.). The platform allows users to access and download data from these satellites.

[0071] (2-2-2) Perform radiometric calibration on the acquired remote sensing data and convert each pixel value Convert to reflectance: ; ;in, is the incident radiation; is the remote sensing reflectivity; is the image value obtained after radiometric calibration, is the gain factor, which is used to convert Convert to physical quantity. Represents each pixel value Convert to a ratio of physical quantities. is the bias term used to correct for systematic errors or background noise. Indicates the reference value adjustment when converting DN to physical quantity.

[0072] The reflectivity, water vapor content, aerosol concentration and solar radiation are input into the MODTRAN model for atmospheric correction. The atmospheric effect is deducted from the observed reflectivity to obtain the true remote sensing reflectivity of the water body.

[0073] It should be understood that MODTRAN, which stands for Moderate Resolution Atmospheric Transmission, is a computer model used to simulate the transmission of electromagnetic radiation in the Earth's atmosphere. The model is mainly used to calculate the effects of the atmosphere on the absorption, scattering and emission of radiation transmission processes. It combines a variety of atmospheric conditions, including different profiles of temperature, pressure, humidity and aerosol content, to predict the interaction of light of different wavelengths with the atmosphere.

[0074] It should be understood that atmospheric influence refers to the various interferences caused by the atmosphere to the reflectivity data observed by the sensor, including various gases absorbing light of specific wavelengths, resulting in a decrease in the intensity of radiation reflected from the water body; aerosols and other particulate matter in the atmosphere scatter light, changing the radiation path reflected from the water body to the sensor, affecting the observed light intensity and color; the atmosphere itself emits radiation at certain wavelengths, and these background radiations mix with the radiation reflected by the water body, affecting the final observation results.

[0075] (2-2-3) Use the least squares method to calculate the similarity between the real remote sensing reflectance of the water body and the spectral data in the three-dimensional bio-optical database; find the best matching band combination based on the maximum similarity. Based on the best matching band combination, expand the surface inversion results to the vertical direction.

[0076] It should be understood that the least squares method is a mathematical optimization technique for data fitting, which aims to find the best function model by minimizing the sum of squares of errors. (i=1,2,...,n), the model we want to fit is , then the goal of the least squares method is to minimize the following loss function: ;in, is the sum of square errors.

[0077] Application steps: (1) Choose a suitable function form , such as the linear model Or polynomial model. Construct loss function according to the selected model ; (2) Calculate the partial derivative of the loss function and set the result to zero to find the optimal solution for the model parameters (such as slope and intercept); (3) Solve the obtained equation to obtain the parameter value of the model. Use new data points to verify the fitting effect of the model and calculate indicators such as R² and RMSE for evaluation.

[0078] Furthermore, the (2-3): determining the water column profile distribution function, establishing a decision tree model, and determining the chlorophyll a profile distribution type through the decision tree model and the water column profile distribution function, specifically includes: establishing a decision tree model to identify the chlorophyll a profile distribution type, thereby determining the water column chlorophyll a profile distribution function to characterize the chlorophyll a profile distribution type and distribution function.

[0079] Combined with the vertical distribution data of algae and underwater light field data under different hydrological and meteorological environmental conditions obtained from field experiments, the main vertical distribution types of algae in the study area were analyzed, the vertical distribution function was calibrated, and the parametric expression of each vertical distribution type was realized.

[0080] Furthermore, the (2-3): determining the water column profile distribution function specifically includes: (2-3-1) if the change amplitude of the algae concentration in the depth direction is less than the set threshold, the algae distribution type in the current area is uniform distribution, and (2-3-2) If the algae are concentrated in a certain layer of the water body, and the rate of decrease of the algae concentration with increasing depth is greater than the set threshold, then the algae distribution type in this area is surface enrichment distribution, which is represented by an exponential model; (2-3-3) If the algae concentration in the bottom layer of the water body is higher than the set threshold, and the algae concentration in the surface layer of the water body is lower than the set threshold, then the algae distribution type in this area is bottom enrichment distribution, which is represented by an inverse exponential distribution model; (2-3-4) If the algae concentration in the middle layer of the water body is higher than the set threshold, and the algae concentrations in both the surface and bottom layers of the water body are lower than the set threshold, then the algae distribution type in this area is bottom enrichment distribution, which is represented by an inverse exponential distribution model. If the concentration of algae in the water body is lower than the set threshold, the algae distribution type in the area is a mid-layer enrichment type distribution, which can be represented by a Gaussian distribution model. (2-3-5) If the floating algae form multiple enrichment layers in the water body, and the algae concentrations in the surface layer, middle layer and bottom layer of the water body are all higher than the set threshold, the algae distribution type in the area is a multi-layer distribution, which can be represented by the superposition of multiple Gaussian distribution functions. (2-3-6) If the algae concentration in the local patch in the middle of the water body is higher than the set threshold, and the other patches are lower than the set threshold, the algae distribution type in the area is a patchy distribution, which can be represented by a segmented Gaussian model.

[0081] The vertical distribution function of phytoplankton is used to describe the distribution of phytoplankton in the water body as it changes with depth. The vertical distribution function is very important for understanding the concentration changes of phytoplankton at different water depths and their response to light conditions, nutrient distribution and water physical properties. Common vertical distribution functions are as follows: Exponential function distribution: ;in, Is Depth The chlorophyll concentration at is the surface concentration, is the attenuation coefficient. This function indicates that the concentration of phytoplankton decays exponentially with increasing depth. is a constant.

[0082] Inverse exponential distribution model: ;in, Is Depth The chlorophyll concentration at is the algae concentration in the bottom layer, is the coefficient of variation.

[0083] Gaussian distribution: ;in, Is Depth The chlorophyll concentration at is more than The chlorophyll concentration, is the standard deviation, which is a parameter related to the maximum chlorophyll concentration in the layer. is the depth of the chlorophyll maximum.

[0084] Superposition of multiple Gaussian distribution functions: ;in, Is Depth The chlorophyll concentration at It is The maximum concentration of the enriched layer, is the depth of the enriched layer, is the width of the layer, is depth, Represents a constant.

[0085] Piecewise Gaussian model: .

[0086] in, Is Depth The chlorophyll concentration at They represent the different depth limits of algae concentration changes. Indicates the depth range arrive The rate of change of concentration between Indicates the depth range arrive The rate of change of concentration between .

[0087] Furthermore, the (2-3): establishing a decision tree model, determining the chlorophyll a profile distribution type through the decision tree model and the water column profile distribution function, specifically including: constructing a decision tree to identify six algae distribution types, first setting the root node threshold to 0.5 based on the depth to concentration ratio (DMC), and for samples with a depth to concentration ratio DMC greater than 0.5, further judging them as surface enriched or mid-layer enriched based on the reflectance variation coefficient (RCC) threshold of 1.2.

[0088] For samples with a depth-to-concentration ratio DMC less than or equal to 0.5, they are judged as uniformly distributed or bottom-enriched based on the local concentration standard deviation (LCSD) threshold of 1.0; if the standard deviation is greater than 1.0, it is further identified as bottom-enriched or patchy distribution through the bottom concentration to surface concentration ratio (BSCR) threshold of 2.0; if the DMC is less than or equal to 0.5 and the standard deviation is less than or equal to 1.0, it is classified as uniformly distributed.

[0089] By judging the nodes layer by layer, the decision tree can efficiently classify the samples into various distribution types. The results are as follows: (1) When the depth to concentration ratio DMC ≤ 0.5 and the local concentration difference ≤ 1.0, it is determined to be a uniform distribution.

[0090] (2) When the depth-to-concentration ratio DMC>0.5 and the reflectivity variation coefficient RCC>1.2, it is determined to be a surface-enriched distribution.

[0091] (3) When the depth-to-concentration ratio DMC>0.5, the reflectivity variation coefficient RCC≤1.2 and the local concentration variation rate ≤20%, it is determined to be a mid-layer enrichment type.

[0092] (4) When the depth-to-concentration ratio DMC is ≤ 0.5, the local concentration standard deviation LCSD is greater than 1.0, and the bottom-to-surface concentration ratio BSCR is greater than 2.0, it is determined to be a bottom-enriched distribution.

[0093] (5) When the depth-to-concentration ratio DMC>0.5, the reflectivity variation coefficient RCC≤1.2, and the local concentration variation rate>20%, it is determined to be a multi-level distribution.

[0094] (6) When the depth-to-concentration ratio DMC ≤ 0.5, the local concentration standard deviation LCSD > 1.0, and the bottom-to-surface concentration ratio BSCR ≤ 2.0, it is determined to be a patchy distribution.

[0095] The root node is the depth to concentration ratio (DMC), and the threshold is 0.5; if DMC>0.5, classification is continued based on the reflectance variation coefficient (RCC), with a threshold of 1.2; if RCC>1.2, it is determined to be a surface enriched distribution, and if RCC≤1.2, identification is continued based on the local concentration change rate, with a threshold of 20%; if the local concentration change rate>20%, it can be determined as a multi-level distribution, and if the local concentration change rate≤20%, it is determined to be a mid-layer enriched type.

[0096] If DMC≤0.5, classification will continue based on the local concentration standard deviation (LCSD) with a threshold of 1.0; if LCSD≤1.0, it will be determined as a uniform distribution; if LCSD>1.0, identification will continue based on the bottom concentration to surface concentration ratio (BSCR) with a threshold of 2.0; if BSCR>2.0, it can be determined as a bottom-enriched distribution; and if BSCR≤2.0, it can be determined as a patchy distribution.

[0097] It should be understood that DMC (Depth-to-Concentration Ratio) refers to the ratio of algae concentration at different depths in the water column to the concentration at a reference depth. The calculation method is: collect water samples at different depths, measure the algae concentration, select a reference depth, and calculate the ratio of the concentration at each depth to the concentration at that depth.

[0098] RCC (Reflectance Change Coefficient) refers to the degree of change in remote sensing reflectance measured under different conditions (such as time, depth, and light). It is used to evaluate how the reflectance of water bodies or surface features changes with environmental conditions. The calculation method is: obtain remote sensing reflectance data at different times or conditions, and use a formula to calculate the degree of change in reflectance, usually expressed as a standard deviation or coefficient of variation.

[0099] LCSD (Local Concentration Standard Deviation) is a statistical indicator used to describe the degree of concentration distribution variation of a substance (such as chlorophyll, suspended solids, etc.) in a specific area. It reflects the degree of dispersion of concentration within a specific spatial range. The calculation method is: collect samples in a specific water area, obtain the concentration value of the corresponding substance, and use the formula Calculate the standard deviation of the local concentrations.

[0100] BSCR (Bottom to Surface Concentration Ratio) is an indicator used to evaluate the distribution characteristics of a substance (such as chlorophyll, suspended solids, etc.) in water bodies at different depths. This ratio can help understand the vertical distribution of algae or other organisms in water bodies and reflect the relative concentration difference between the bottom layer and the surface layer. The calculation method is: collect samples at different water depths (such as the surface and bottom layers), record the concentration of each layer, and calculate the ratio of the bottom layer to the surface layer concentration.

[0101] Further, the S103: based on the chlorophyll a profile distribution type and the surface chlorophyll a concentration inversion result, vertical multi-layer continuous chlorophyll a concentration and total algae inversion are performed to obtain the estimated result of the total algae, including: (3-1) if the chlorophyll a in the water column is vertically uniformly distributed, and the variation range of the chlorophyll a concentration in the vertical direction is less than the set threshold, the total algae in the water column is calculated by multiplying the chlorophyll a concentration inversion result of the water body surface by the water depth; ;in, The total amount of algae in the water column, Indicates the chlorophyll a concentration in the surface layer of the water body, Indicates the depth of the water body; (3-2) If the chlorophyll a in the water column is unevenly distributed, the remote sensing reflectance of each band of Sentinel MSI is first matched with the Hydrolight simulated bio-optical library for optimal spectrum matching, and then the decision tree corresponding to the chlorophyll a profile type, DMC, RCC, LCSD, and BSCR is used for identification to determine the chlorophyll a profile of the water column, and the water depth is integrated to obtain the total amount of algae.

[0102] Furthermore, the (3-2) includes: (3-2-1) acquiring remote sensing reflectance data of each band from Sentinel MSI.

[0103] Table 1 Remote sensing reflectivity data of each band

[0104] .

[0105] (3-2-2) The remote sensing reflectance data were compared with the Hydrolight simulated bio-optical database. The best matching spectral band was found by minimizing the squared difference between the predicted value and the observed value through the least squares method. The chlorophyll a concentration on the water column surface was determined through the three-dimensional bio-optical database.

[0106] It should be understood that HydroLight is a computational model based on radiation transfer theory, which is mainly used to simulate the optical properties and bio-optical behavior in water. Its bio-optical library contains a large number of simulation results involving spectral reflectance, absorption and scattering properties under different water environmental conditions, as well as the relationship between these properties and chlorophyll a.

[0107] The simulated reflectance data under the corresponding water body conditions were obtained from the three-dimensional database. For each band, the square difference between the remote sensing reflectance (actual value) and the HydroLight simulated reflectance (predicted value) was calculated. The square difference was minimized to obtain the band with the strongest correlation with the chlorophyll a concentration in the water body. Using the obtained optimal band, the surface chlorophyll a concentration corresponding to the band under the corresponding water body conditions was searched in the bio-optical library.

[0108] For each band, the square difference between the remote sensing reflectance (actual value) and the HydroLight simulated reflectance (predicted value) is calculated. The band corresponding to the combination with the smallest square difference is the best matching spectral band.

[0109] After finding the best matching spectral band, the chlorophyll concentration under the corresponding water conditions in the three-dimensional database is the chlorophyll a concentration on the water column surface.

[0110] (3-2-3) Input the optical characteristics of the water environment to be estimated into the established decision tree, determine the chlorophyll a profile type at different depth levels, thereby determining the water column profile distribution function, and perform water depth integration to obtain the total amount of algae. The water depth integration means: based on the known surface chlorophyll a concentration, the water column profile distribution function is integrated and summed from the water depth of 0 to the maximum water depth. Based on the surface chlorophyll a concentration, the profile function is used to perform integration and summation.

[0111] The optical properties of the water environment to be estimated include: all nodes involved in the decision tree, DMC, RCC, LCSD, and BSCR.

[0112] Data collection: (1) Download cloud-free, high-quality Sentinel MSI remote sensing images, and obtain water-leaving radiation and remote sensing reflectance through radiometric calibration and atmospheric correction. (2) Collect water samples at different water depths and corresponding radiometric and algae absorption coefficients through stratified water samplers, measure environmental factors such as transparency, dissolved oxygen, salinity, temperature, wind speed, and wind direction, and obtain vertical distribution of phytoplankton and underwater light field data under different environmental conditions. (3) Collect water body spectra synchronized / quasi-synchronized with satellite images and build a remote sensing reflectance spectrum library.

[0113] Development of surface chlorophyll concentration inversion model based on Transformer deep learning: Based on deep learning algorithms such as Transformer, we used diagnostic spectral features and algae total monitoring data to develop an algae total remote sensing estimation model; and used R 2 , MSE, RMSE and other evaluation indicators were used to adjust the degree of model fitting, optimize key parameters, and converge the model when the number of iterations or the residual error was minimum. Finally, a high-precision and robust remote sensing estimation model for the total algae was developed. Subsequently, the estimation model was extrapolated to the remote sensing image based on the spectral response function of the Sentinel MSI image, and the remote sensing distribution dataset of the total algae was inverted.

[0114] Characterization of water column profile distribution function: Based on the results of a large number of lake and reservoir water column bio-optical surveys and surface model inversion results, a three-dimensional bio-optical database for multiple situations in different lakes and reservoirs was constructed based on HydroLight radiation transmission simulation, and then the optimal spectrum matching between Sentinel MSI remote sensing reflectance and the bio-optical database was performed. Then a decision tree model was established to identify the chlorophyll a profile distribution type, thereby determining the water column Chl-a profile distribution function to characterize the chlorophyll a profile distribution type and distribution function.

[0115] Three-dimensional estimation of total algae: Based on the inversion results of chlorophyll a concentration in surface modeling and the chlorophyll a profile distribution type identified in the profile simulation, vertical multi-layer continuous chlorophyll a concentration and total algae (multi-layer integration) inversion is performed. For the estimation of the total amount of water column algae: (1) If the chlorophyll a in the water column is vertically uniformly distributed, the total amount of water column algae is calculated by multiplying the surface chlorophyll a concentration by the water depth; (2) If the chlorophyll a in the water column is non-uniformly distributed, the remote sensing reflectance of each band of Sentinel MSI is first matched with the Hydrolight simulated bio-optical library of the corresponding chlorophyll a profile type, DMC, RCC, LCSD, BSCR and inorganic suspended matter concentration to determine the water column chlorophyll a profile, and the total amount of algae is obtained by water depth integration.

[0116] Embodiment 2: This embodiment provides a remote sensing estimation system for the total amount of algae based on vertical pre-classification and deep learning, including: a first acquisition module, which is configured to: obtain the chlorophyll a concentration inversion result of the surface layer of the water body; a second acquisition module, which is configured to: obtain the chlorophyll a profile distribution type based on the chlorophyll a concentration inversion result of the surface layer of the water body and water environment information;

[0117] The estimation module is configured to: perform vertical multi-layer continuous chlorophyll a concentration and total algae inversion based on the chlorophyll a profile distribution type and the surface chlorophyll a concentration inversion results to obtain the estimation result of the total algae amount.

[0118] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A remote sensing estimation method for total algae abundance based on vertical pre-classification and deep learning, characterized by: include: Obtain the chlorophyll a concentration inversion results of the water surface, including: Constructing a training set and a validation set, wherein both the training set and the validation set are the optimal characteristic band combinations of the known chlorophyll a concentration in the surface layer of the water body; The optimal characteristic band combination is used as the input value of the Transformer model, and the known chlorophyll a concentration of the surface layer of the water body is used as the output value of the Transformer model, and the Transformer model is trained to obtain a trained Transformer model; the trained Transformer model is verified, and if the verification passes, it means that the current Transformer model is the final trained Transformer model; The optimal characteristic band combination of the water environment to be estimated for the total amount of algae is input into the final trained Transformer model to obtain the chlorophyll a concentration inversion result of the water surface; Based on the chlorophyll a concentration inversion results of the water surface and water environment information, the chlorophyll a profile distribution type is obtained, including: Based on the inversion results of chlorophyll a concentration in the water surface, a three-dimensional bio-optical database was constructed; Optimal spectral matching of remote sensing data with the three-dimensional bio-optical database; Determine the water column profile distribution function, establish a decision tree model, and determine the chlorophyll a profile distribution type through the decision tree model and the water column profile distribution function; Determine the water column profile distribution function, including: If the change in algae concentration in the depth direction is less than the set threshold, the algae distribution type in the current area is uniform distribution, and express; If the algae are concentrated in a certain layer of the water body, and the rate of decrease of algae concentration with increasing depth is greater than the set threshold, the algae distribution type in this area is surface enrichment distribution, which is represented by an exponential model; If the algae concentration in the bottom layer of the water body is higher than the set threshold, and the algae concentration in the surface layer of the water body is lower than the set threshold, the algae distribution type in this area is bottom-enriched distribution, which is represented by the inverse exponential distribution model; If the algae concentration in the middle layer of the water body is higher than the set threshold, and the algae concentrations in the surface and bottom layers of the water body are lower than the set threshold, the algae distribution type in this area is a middle layer enrichment type distribution, which is represented by a Gaussian distribution model; If phytoplankton forms multiple enrichment layers in the water body, and the algae concentrations in the surface layer, middle layer and bottom layer of the water body are all higher than the set threshold, then the algae distribution type in this area is multi-level distribution, which is represented by the superposition of multiple Gaussian distribution functions; If the algae concentration in the local patch in the middle of the water body is higher than the set threshold, and the other patches are lower than the set threshold, the algae distribution type in this area is patchy distribution, which can be represented by a piecewise Gaussian model; Based on the chlorophyll a profile distribution type and the surface chlorophyll a concentration inversion results, vertical multi-layer continuous chlorophyll a concentration and total algae inversion were performed to obtain the estimated results of the total algae, including: If the chlorophyll a in the water column is evenly distributed vertically, and the vertical variation of the chlorophyll a concentration is less than the set threshold, the total amount of algae in the water column is calculated by multiplying the chlorophyll a concentration inversion result of the water body surface by the water depth; ； in, The total amount of algae in the water column, Indicates the chlorophyll a concentration in the surface layer of the water body, Indicates the depth of a body of water; If the chlorophyll a in the water column is unevenly distributed, the remote sensing reflectance of each band of Sentinel MSI is first matched with the Hydrolight simulated bio-optical library for optimal spectrum matching, and then a decision tree is used for identification to determine the chlorophyll a profile of the water column, and the total amount of algae is obtained by water depth integration.

2. The method for remote sensing estimation of total algae amount based on vertical pre-classification and deep learning as claimed in claim 1, characterized in that: Construct training sets and validation sets, including: Obtaining surface water sample data of different lakes and reservoirs; the surface water sample data of different lakes and reservoirs include: optical properties and water quality parameters; the optical properties include: absorption coefficient and scattering coefficient; the water quality parameters include: chlorophyll a concentration, suspended matter and colored soluble organic matter; Performing a preprocessing operation on the acquired data, wherein the preprocessing operation includes: removing missing values ​​and outliers; The data after preprocessing operation is subjected to feature screening to obtain the optimal feature band combination, and the training set and the validation set are obtained according to the optimal feature band combination and the corresponding chlorophyll a concentration.

3. The method for remote sensing estimation of total algae amount based on vertical pre-classification and deep learning as claimed in claim 2, characterized in that: The data after preprocessing is subjected to feature screening to obtain the optimal feature band combination, including: Continuous wavelet transform is used to extract the first characteristic band; Correlation analysis is used to extract the second characteristic band; The first characteristic band and the second characteristic band are connected in series to obtain an optimal characteristic band combination; The first characteristic band is extracted by using continuous wavelet transform, specifically including: extracting data with a band of 400-900 nm from the data after preprocessing operation, and extracting intrinsic and apparent optical features of different scales by using continuous wavelet transform; selecting a band with a signal-to-noise ratio greater than 5 dB, a mean square error less than 0.1, and a p-value less than 0.05 with chlorophyll a concentration as the first characteristic band; Correlation analysis was used to extract the second characteristic band, specifically including: calculating the Pearson product-moment correlation coefficient between the reflectance of each band and the surface chlorophyll a concentration to evaluate the influence of the band on the concentration, and selecting the band with a correlation coefficient higher than 0.6 as the second characteristic band.

4. The method for remote sensing estimation of total algae amount based on vertical pre-classification and deep learning as claimed in claim 1, characterized in that: The trained Transformer model includes: an input layer, a multi-head attention mechanism layer, a first-layer normalization module, a feedforward neural network, a second-layer normalization module and a decoder connected in sequence; the input end of the first-layer normalization module is also connected to the input layer; the input end of the second-layer normalization module is also connected to the input end.

5. The method for remote sensing estimation of total algae amount based on vertical pre-classification and deep learning as claimed in claim 1, characterized in that: Establishing a decision tree model, and determining the chlorophyll a profile distribution type through the decision tree model and the water column profile distribution function, specifically including: The depth to concentration ratio DMC ≤ 0.5 and the local concentration difference ≤ 1.0 are determined to be uniformly distributed; The depth-to-concentration ratio DMC>0.5 and the reflectivity variation coefficient RCC>1.2 were determined to be a surface-enriched distribution; The depth-to-concentration ratio DMC>0.5, the reflectivity variation coefficient RCC≤1.2 and the local concentration variation rate ≤20% are determined as the mid-layer enrichment type; The depth-to-concentration ratio DMC≤0.5, the local concentration standard deviation LCSD>1.0, and the bottom-to-surface concentration ratio BSCR>2.0 were determined to be a bottom-enriched distribution; The depth-to-concentration ratio DMC>0.5, the reflectivity variation coefficient RCC≤1.2, and the local concentration variation rate>20% are determined as multi-level distribution; A patchy distribution was determined when the depth-to-concentration ratio DMC ≤ 0.5, the local concentration standard deviation LCSD > 1.0, and the bottom-to-surface concentration ratio BSCR ≤ 2.

0.

6. The method for estimating the total amount of algae by remote sensing based on vertical pre-classification and deep learning as claimed in claim 1, characterized in that: If the chlorophyll a distribution in the water column is non-uniform, the remote sensing reflectance of each band of Sentinel MSI is first matched with the Hydrolight simulated bio-optical library for optimal spectrum matching, and then a decision tree is used for identification to determine the chlorophyll a profile of the water column, and the total amount of algae is obtained by water depth integration, including: Obtain remote sensing reflectance data of each band; The remote sensing reflectance data were compared with the Hydrolight simulated bio-optical database, and the best matching spectral band was found by minimizing the squared difference between the predicted and observed values ​​using the least squares method. The chlorophyll a concentration on the surface of the water column was determined using the three-dimensional bio-optical database. The optical characteristics of the water environment to be estimated are input into the established decision tree to determine the chlorophyll a profile type at different depth levels, thereby determining the water column profile distribution function, and performing water depth integration to obtain the total amount of algae.

7. A remote sensing estimation system for algae abundance based on vertical pre-classification and deep learning, characterized by: include: The first acquisition module is configured to: obtain the chlorophyll a concentration inversion result of the water surface, specifically including: Constructing a training set and a validation set, wherein both the training set and the validation set are the optimal characteristic band combinations of the known chlorophyll a concentration in the surface layer of the water body; The optimal characteristic band combination is used as the input value of the Transformer model, and the known chlorophyll a concentration of the surface layer of the water body is used as the output value of the Transformer model, and the Transformer model is trained to obtain a trained Transformer model; the trained Transformer model is verified, and if the verification passes, it means that the current Transformer model is the final trained Transformer model; The optimal characteristic band combination of the water environment to be estimated for the total amount of algae is input into the final trained Transformer model to obtain the chlorophyll a concentration inversion result of the water surface; The second acquisition module is configured to: acquire the chlorophyll a profile distribution type based on the chlorophyll a concentration inversion result of the water body surface and the water environment information, including: Based on the inversion results of chlorophyll a concentration in the water surface, a three-dimensional bio-optical database was constructed; Optimal spectral matching of remote sensing data with the three-dimensional bio-optical database; Determine the water column profile distribution function, establish a decision tree model, and determine the chlorophyll a profile distribution type through the decision tree model and the water column profile distribution function; Determine the water column profile distribution function, including: If the change in algae concentration in the depth direction is less than the set threshold, the algae distribution type in the current area is uniform distribution, and express; If the algae are concentrated in a certain layer of the water body, and the rate of decrease of algae concentration with increasing depth is greater than the set threshold, the algae distribution type in this area is surface enrichment distribution, which is represented by an exponential model; If the algae concentration in the bottom layer of the water body is higher than the set threshold, and the algae concentration in the surface layer of the water body is lower than the set threshold, the algae distribution type in this area is bottom-enriched distribution, which is represented by the inverse exponential distribution model; If the algae concentration in the middle layer of the water body is higher than the set threshold, and the algae concentrations in the surface and bottom layers of the water body are lower than the set threshold, the algae distribution type in this area is a middle layer enrichment type distribution, which is represented by a Gaussian distribution model; If phytoplankton forms multiple enrichment layers in the water body, and the algae concentrations in the surface layer, middle layer and bottom layer of the water body are all higher than the set threshold, then the algae distribution type in this area is multi-level distribution, which is represented by the superposition of multiple Gaussian distribution functions; If the algae concentration in the local patch in the middle of the water body is higher than the set threshold, and the other patches are lower than the set threshold, the algae distribution type in this area is patchy distribution, which can be represented by a piecewise Gaussian model; The estimation module is configured to: perform vertical multi-layer continuous chlorophyll a concentration and total algae inversion based on the chlorophyll a profile distribution type and the surface chlorophyll a concentration inversion result, and obtain the estimation result of the total algae, including: If the chlorophyll a in the water column is evenly distributed vertically, and the vertical variation of the chlorophyll a concentration is less than the set threshold, the total amount of algae in the water column is calculated by multiplying the chlorophyll a concentration inversion result of the water body surface by the water depth; ； in, The total amount of algae in the water column, Indicates the chlorophyll a concentration in the surface layer of the water body, Indicates the depth of a body of water; If the chlorophyll a in the water column is unevenly distributed, the remote sensing reflectance of each band of Sentinel MSI is first matched with the Hydrolight simulated bio-optical library for optimal spectrum matching, and then a decision tree is used for identification to determine the chlorophyll a profile of the water column, and the total amount of algae is obtained by water depth integration.

Images