A method and system for chlorophyll concentration inversion based on hierarchical feature embedding

Abstract

This invention relates to the fields of oceanography, remote sensing, and artificial intelligence, and provides a method and system for chlorophyll concentration inversion based on hierarchical feature embedding. The method includes: establishing a hierarchical feature embedding module for mapping two different levels of spectral features, one of which serves as prior optical knowledge, characterizing the inherent optical properties of the aquatic environment; the other replaces common spectral combination features, aiming to fully explore the potential relationships between multiple spectral bands while maintaining high computational efficiency. Based on this, a Bayesian inference framework is introduced into the embedding module to quantify the uncertainty of chlorophyll a estimation results. Furthermore, an attention mechanism is applied to allow the hierarchically mapped spectral features to dynamically influence each other, thereby ensuring that the method adaptively selects appropriate optical relationships for chlorophyll a concentration estimation in different aquatic environments.

Description

Technical Field

[0001] This invention relates to the fields of oceanography, remote sensing, and artificial intelligence, and provides a method and system for chlorophyll concentration inversion based on hierarchical feature embedding. Background Technology

[0002] With the rapid development of human production activities, excessive and unbalanced nutrient input has altered the nutrient level and structure of the marine environment, leading to a series of environmental problems such as harmful algal blooms, expansion of marine hypoxia areas, and ocean acidification. These problems seriously affect marine ecosystems, fisheries, and coastal construction and management. To address these issues, we propose a Bayesian deep learning-based chlorophyll a concentration inversion method, namely VB-OCformer, which overcomes the limitations of traditional ocean color methods through deep learning algorithms and effectively solves the aforementioned problems. Summary of the Invention

[0003] To address the problems existing in the prior art, the present invention aims to provide a Bayesian deep learning-based method for chlorophyll a concentration inversion based on hierarchical feature embedding. This method overcomes the limitations of traditional ocean color methods by using deep learning algorithms and replaces traditional optical theories with prior optical knowledge established by the model, thereby more effectively dealing with the complex inherent optical characteristics of different water bodies. At the same time, it provides global ocean chlorophyll a concentration estimation results using limited spectral bands and quantifies the uncertainty of the estimation results.

[0004] To address the above problems, the present invention adopts the following technical solution: a chlorophyll concentration inversion method based on hierarchical feature embedding, comprising:

[0005] Acquire spectral remote sensing reflectance data from ocean color satellites;

[0006] Acquire measured chlorophyll a concentration data within a preset ocean range, and use the measured chlorophyll a concentration data as target reference data for training a Bayesian deep learning model;

[0007] The spectral remote sensing reflectance data and the measured chlorophyll a concentration data are matched in time and space and integrated to construct a dataset for training a Bayesian deep learning model.

[0008] The dataset is standardized to eliminate the influence of different units of measurement, resulting in a standardized dataset.

[0009] The standardization process is as follows: ;

[0010] in, i This represents the i-th piece of raw data that has not undergone standardization. It is the mean of the dataset; It is the standard deviation of the dataset; This represents a standardized dataset;

[0011] The standardized dataset is shuffled to form a matching dataset, which is then divided into three subsets: a training set, a validation set, and a test set. These subsets are used for training, optimizing, and evaluating the Bayesian deep learning model, respectively, to obtain a Bayesian deep learning model with hierarchical feature embedding. This Bayesian deep learning model is used to obtain real-time chlorophyll a concentration data in the ocean area based on real-time spectral remote sensing reflectance data acquired by ocean color satellites.

[0012] Preferably, the step of acquiring measured chlorophyll a concentration data within a preset ocean area and using the measured chlorophyll a concentration data as target reference data for training a Bayesian deep learning model includes:

[0013] Collect measured chlorophyll a concentration data within a pre-defined marine area;

[0014] The collected measured chlorophyll a concentration data were cleaned and purified.

[0015] High performance liquid chromatography was selected to measure chlorophyll a concentration data;

[0016] Logarithmic transformation was performed on the processed measured chlorophyll a concentration data to obtain target reference data.

[0017] Preferably, the spectral remote sensing reflectance data and the measured chlorophyll a concentration data are matched temporally and spatially to construct a dataset for training a Bayesian deep learning model, including:

[0018] The spectral remote sensing reflectance data corresponding to the measured chlorophyll a concentration data were extracted using a 3 × 3 pixel bounding box.

[0019] Two different time windows, "daily window" and "8-day window", are used to match the data in time and space. A corresponding maximum coefficient of variation (CV) threshold is set for each time window. If the CV of the matched spectral remote sensing reflectance data exceeds the specified threshold, it is excluded.

[0020] Preferably, the dataset is standardized to eliminate the influence of different units of measurement, resulting in a standardized dataset, including:

[0021] The standardization process is as follows: ;

[0022] in, i represents the i-th unstandardized raw data item; μ is the mean of the dataset; σ is the standard deviation of the dataset; This represents a standardized dataset; that is, the standardized data = (original data - mean) ÷ standard deviation.

[0023] Preferably, the step of shuffling the standardized dataset to form a matching dataset, and dividing the matching dataset into three subsets: a training set, a validation set, and a test set, for use in training, optimization, and evaluation of the Bayesian deep learning model, includes:

[0024] To eliminate potential biases in the data, such as order, distribution, and local features, the standardized dataset is shuffled to form the matching dataset.

[0025] The matching dataset is divided into three subsets: training set, validation set, and test set. The matching dataset is sequentially entered into these three subsets, and the training, optimization, and evaluation of the Bayesian deep learning model are performed sequentially.

[0026] During the training phase, the Bayesian deep learning model has an inference phase.

[0027] The reasoning stage is processed as follows: ;

[0028] in, The model parameters of the Bayesian deep learning model during the inference phase are represented. and This represents a variational approximation of the posterior distribution of the parameters; This represents the identity matrix sampled from the standard normal distribution;

[0029] The Bayesian deep learning model training uses mean squared error (MSE) as the main loss function; mean squared error (MSE) is used as the regularization penalty term in the loss function during variational Bayesian learning.

[0030] The total loss function is: ;

[0031] in, These are regularization weight coefficients. The number of Monte Carlo samplings is indicated; Mean Squared Error (MSE): reflects the overall difference between the predicted and measured values; Variational Lower Bound (ELBO): serves as a "penalty term"; that is, Total Loss = MSE + Penalty Term (ELBO × coefficient);

[0032] When the verification set is used, the verification phase is processed as follows: ;

[0033] The model parameters represent the Bayesian deep learning model during the validation phase; It is the approximate mean of the variational posterior distribution of the parameters;

[0034] When using the test set, the performance of all methods in chlorophyll a concentration inversion was comprehensively evaluated, and their overall performance and performance under different nutrient levels were analyzed. Evaluation indicators included fitting slope, fitting bias, correlation coefficient, mean absolute error, median absolute error, and root mean square error.

[0035] Preferably, the inference operation of the Bayesian deep learning model includes:

[0036] During the model inference phase, a standardized processing method consistent with the model training is adopted, and the spectral remote sensing reflectance data to be estimated is preprocessed using the constructed matching dataset;

[0037] The following weight sampling method is used for model inference:

[0038] ;

[0039] in, These represent the model parameters during the inference process. and This represents a variational approximation of the posterior distribution of the parameters. This represents the identity matrix sampled from the standard normal distribution;

[0040] Based on the Bayesian inference framework, the Monte Carlo sampling method is used to repeatedly sample the parameters, thereby obtaining the estimated chlorophyll a concentration and its uncertainty quantification results, mainly including the standard deviation (SD) and CV, which are calculated as follows:

[0041] ;

[0042] ;

[0043] ;

[0044] in, , and These represent the mean, SD, and CV of the estimated chlorophyll a concentration, respectively. Furthermore, The sample variance of the estimated results is represented by the number of Monte Carlo samplings and the confidence level. and This indicates that the values ​​are set to 10 and 0.95 respectively. It is based on the specified confidence level and degrees of freedom The obtained critical value of the t-distribution;

[0045] The final estimated values ​​of chlorophyll a concentration from different ocean color satellites were compared with other publicly available chlorophyll a concentration data products.

[0046] Preferably, the evaluation of the Bayesian deep learning model includes:

[0047] It was compared with several chlorophyll a concentration inversion methods, including three widely accepted empirical algorithms and five machine learning models;

[0048] During the evaluation process, the empirical algorithm refits the matching dataset using the least squares method; the training process of the machine learning model is basically the same as that of this method, and the respective optimization strategies of these models are used.

[0049] A comprehensive evaluation was conducted on the chlorophyll a concentration inversion performance of all methods, and their overall performance and performance under different nutrient levels were analyzed. Beneficial effects

[0050] Compared with the prior art, the advantages of this invention are:

[0051] A hierarchical embedding module was established to map two different levels of spectral features. One level serves as prior optical knowledge to characterize the inherent optical properties of the aquatic environment. The other level replaces common spectral combination features, aiming to fully explore the potential relationships between multiple spectral bands while maintaining high computational efficiency.

[0052] A Bayesian inference framework is introduced into the embedding module to quantify the uncertainty of chlorophyll a estimation results;

[0053] An attention mechanism is applied to enable the spectral features of the hierarchical mapping to dynamically influence each other, thereby ensuring that the method of the present invention can adaptively select appropriate optical relationships for estimating chlorophyll a concentration in different water environments. Attached Figure Description

[0054] Figure 1 This is a flowchart of a chlorophyll concentration inversion method based on hierarchical feature embedding proposed in this invention;

[0055] Figure 2 The diagram shows the module structure of a chlorophyll concentration inversion system based on hierarchical feature embedding proposed in this invention. Detailed Implementation

[0056] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0057] Please see Figure 1 A method for chlorophyll concentration inversion based on hierarchical feature embedding includes:

[0058] S100. When acquiring global spectral remote sensing reflectance data products from five satellites, select five satellites: MERIS, MODIS-Aqua, MODIS-Terra, SeaWiFS, and SNPP-VIIRS. Extract data according to their respective spatial resolution (4km or 9km) and time range (September 1997 to December 2023), covering both daily and 8-day time resolutions. For each satellite data, use the coefficient of variation (CV) threshold for the corresponding time window (e.g., CV≤1.0 for the daily window and CV≤0.5 for the 8-day window of MERIS) to filter out data with excessive fluctuations. Collect the filtered valid data to form a spectral remote sensing reflectance dataset covering global sea areas and spanning a long time series, providing multi-source and highly adaptable input features for subsequent matching with measured chlorophyll a concentration data and model training; please refer to Table 1.

[0059] Table 1. Statistical summary of detailed information from various ocean color satellites

[0060]

[0061] S200. Next, extract the spectral data corresponding to the measured chlorophyll a concentration data using a 3×3 pixel bounding box, requiring at least 5 effective pixels to reduce noise. Then, filter out data with excessive fluctuations according to the coefficient of variation (CV) thresholds for the daily and 8-day windows corresponding to each satellite (e.g., CV≤1.0 for the daily window of MERIS). Prioritize the use of the daily window, and switch to the 8-day window if the daily window is not met. Finally, collect the processed spectral remote sensing reflectance data as input features for the model to adapt to different satellite characteristics and cover global ocean dynamics, providing a reliable "spectral fingerprint" basis for subsequent inversion.

[0062] To provide a "standard answer" for the Bayesian deep learning model, pre-defined measured chlorophyll a concentration data were first collected, allowing the Bayesian deep learning model to learn the correspondence between spectrum and concentration. Next, data with duplicate times and locations were removed to avoid redundant information interfering with the Bayesian deep learning model's learning and to reduce resource waste. Then, data with concentrations higher than 30 mg·m³ were excluded. - For samples with a depth of 3 or more than 10 m, extreme concentration data are not universal and excessive water depth will reduce the matching degree between satellite observation and measured chlorophyll a concentration data. To avoid the model from overgeneralizing, only data measured by high-reliability high-performance liquid chromatography is retained to ensure the accuracy of the "standard answer". Finally, to adapt to the data distribution characteristics, improve the model fitting effect, and provide a high-quality target reference for subsequent training, the measured chlorophyll a concentration data is logarithmically transformed.

[0063] S300. When matching spectral remote sensing reflectance data with measured chlorophyll a concentration data in time and space to construct a training dataset, firstly, extract the corresponding spectral data using a 3×3 pixel bounding box, ensuring at least 5 valid pixels to reduce the impact of noise from individual pixels; then, prioritize matching using a daily time window, switching to an 8-day window if this is not met, to improve the matching success rate while also considering timeliness; next, filter data with excessive fluctuations according to the CV threshold of different ocean color satellites to avoid the Bayesian deep learning model learning incorrect patterns; finally, integrate them into a matching dataset, mapping "spectral clues" to "concentration answers," providing effective samples for the training of the Bayesian deep learning model, enabling it to grasp the correlation between the two.

[0064] S400. When standardizing the matching dataset, according to the formula... (Where D is the original data, and μ and σ are the mean and standard deviation of the dataset, respectively), that is, the standardized data = (original data - mean) ÷ standard deviation; the spectral remote sensing reflectance data and the measured chlorophyll a concentration data are converted into a standardized dataset of a uniform scale; this is done to eliminate the difference in units between different data, avoid interference from the Bayesian deep learning model's learning of the patterns of measured chlorophyll a concentration data due to differences in the units and numerical ranges of the original data, and ensure that subsequent loss calculation and other steps are performed on a consistent scale, providing a stable data foundation for the training of the Bayesian deep learning model.

[0065] S500. The standardized dataset is randomly rearranged to eliminate potential biases. After forming a matching dataset, it is divided into a training set, a validation set, and a test set at a ratio of 80%, 10%, and 10%, respectively, for training, optimization, and evaluation of the Bayesian deep learning model.

[0066] After inputting spectral remote sensing reflectance data into the Bayesian deep learning model, the model extracts two types of features—prior optical knowledge and potential band relationships—through a hierarchical feature embedding module. These features are then dynamically interacted via an attention mechanism to output the predicted chlorophyll a concentration. Simultaneously, based on the total loss function (…),… The mean squared error and variational lower bound are calculated using Monte Carlo sampling, and the model parameters (such as hierarchical embedding weights) are iteratively adjusted through backpropagation. and To minimize the total loss and complete the training;

[0067] The validation set is 10% of the matching dataset. During training, it is input into the model after a preset number of rounds, and the validation loss and evaluation metric are calculated accordingly. The regularization coefficient is then adjusted based on these results. Hyperparameters are used to trigger early stopping when the validation loss does not decrease significantly over multiple rounds, saving the current parameters to avoid overfitting. The input is given after the Bayesian deep learning model has been trained; it does not participate in parameter adjustment, only calculating indicators such as correlation coefficient (CC), mean absolute error (MAE), and root mean square error (RMSE). These are compared with various chlorophyll a concentration-based inversion methods, including three widely accepted empirical algorithms and five machine learning models.

[0068] During the evaluation process, the empirical algorithm refits the matching dataset using the least squares method; the training process of the machine learning model is basically the same as that of this method, and the respective optimization strategies of these models are used.

[0069] A comprehensive evaluation of the chlorophyll a concentration inversion performance of all methods was conducted, and the overall performance and performance under different nutrient levels were analyzed.

[0070] S201. Collect measured chlorophyll a concentration data within a preset marine area. The collected measured chlorophyll a concentration data mainly comes from multiple public data sources. For the chlorophyll a concentration data measured at the same time and space from each data source, outliers are first screened out, and then the average value of the remaining normal data is calculated. The average value of the measured chlorophyll a concentration data is used as the final reference data.

[0071] S202. To obtain a high-quality chlorophyll a concentration dataset, the collected data underwent rigorous processing. First, the collected measured chlorophyll a concentration data were cleaned, removing data with duplicate times and locations. Next, data with chlorophyll a concentrations higher than 30 mg·m³ were excluded. - ³ or data samples with a water depth exceeding 10 m were used to ultimately obtain a higher quality chlorophyll a concentration dataset.

[0072] S203. By analyzing the usage scenarios and comparing with other measurement methods, it was found that high performance liquid chromatography (HPLC) has better reliability than other measurement methods. Therefore, HPLC-based data on chlorophyll a concentration was used to measure the actual concentration data.

[0073] S204. To adapt to the data distribution characteristics and improve the model's fitting effect on data with different concentration ranges, especially when the chlorophyll a concentration span is large, logarithmic transformation can compress the scale of high concentration values, making the data distribution more uniform and facilitating the model's learning of the correlation between spectrum and concentration. Each effective chlorophyll a concentration data point after screening is transformed using a logarithmic function with a base of 10. calculate , and obtain the converted value.

[0074] S301. Match the above-mentioned spectral remote sensing reflectance data with the measured chlorophyll a concentration data in time and space to construct the dataset for subsequent training of the Bayesian deep learning model. Use 3 × 3 pixel boxes to extract the spectral remote sensing reflectance data corresponding to the measured chlorophyll a concentration data. For each matched data pixel box, it is required to contain at least five valid spectral remote sensing reflectance pixel values ​​greater than 0, that is, ensure at least 5 valid pixels to reduce the influence of noise in a single pixel.

[0075] S302. Two different time windows (daily and 8-day) are used for temporal and spatial matching of data. A corresponding maximum coefficient of variation (CV) threshold is set for each time window. If the CV of the matched data exceeds the specified threshold, it is excluded. The CV threshold is set for each ocean color satellite spectral remote sensing reflectance data product. During the data matching process, the daily time window is used first; if this window cannot meet the matching requirements, the process switches to the 8-day time window to construct the matching dataset.

[0076] First, try the "daily window": find satellite data from the same day as the measured chlorophyll a concentration data to match; if the measured chlorophyll a concentration data from the same day is not qualified, then use the "8-day window" (find satellite data within 8 days before and after the measurement date).

[0077] When pairing, the coefficient of variation (CV) needs to be calculated: if the data fluctuation within this window is too large (CV exceeds the threshold corresponding to each satellite in Table 1), then it is not necessary.

[0078] S401. Standardize the data in the matching dataset to eliminate the influence of different units of measurement. The standardization formula is as follows:

[0079] ;

[0080] in, i represents the i-th piece of raw data that has not undergone standardization; It is the mean of the dataset; It is the standard deviation of the dataset; This represents a standardized dataset; that is, the standardized data = (original data - mean) ÷ standard deviation.

[0081] When standardizing the spectral remote sensing reflectance data and measured chlorophyll a concentration data in the matching dataset, first iterate through each original data item in the dataset and calculate the overall mean μ and standard deviation σ of the dataset; then, for the i-th original data item... i, according to the formula The data is transformed to obtain standardized data. This process unifies the original data with different dimensions (such as the decimal range of spectral data and the numerical range of concentration data) to a similar scale, eliminating the interference of data dimension differences on the learning of the Bayesian deep learning model, ensuring that subsequent loss calculation and other steps are performed on a consistent scale, providing a stable data foundation for the training of the Bayesian deep learning model, and enabling the Bayesian deep learning model to focus more on the patterns of the data itself rather than the differences in numerical magnitude.

[0082] S501. In order to eliminate potential biases in the data such as order, distribution, and local features, the standardized dataset is shuffled to form the matching dataset. This avoids the training effect being affected by the order of the data or other factors. The matching dataset is divided into three subsets: training set, validation set, and test set, which are used for the training, optimization, and evaluation of the Bayesian deep learning model, respectively.

[0083] The matching dataset is divided into three subsets: a training set, a validation set, and a test set. The matching dataset is sequentially entered into these three subsets, and the training, optimization, and evaluation of the Bayesian deep learning model are performed sequentially, including:

[0084] S5011. During the training phase, there is an inference phase for the Bayesian deep learning model.

[0085] The reasoning stage is processed as follows: ;

[0086] in, The model parameters of the Bayesian deep learning model during the inference phase are represented. and This represents a variational approximation of the posterior distribution of the parameters; This represents the identity matrix sampled from the standard normal distribution;

[0087] During the inference phase of the Bayesian deep learning model on the training set, the variational approximation of the posterior distribution of the parameters obtained during training is first performed. (Approximate mean of parameters) and (Distribution dispersion), generating the identity matrix by randomly sampling from the standard normal distribution. Next, to ensure the standard deviation of the parameter is non-negative, it is calculated and then multiplied element-wise (Hadamard product) to obtain the parameter fluctuation term; finally, this fluctuation term is added to the sum. To obtain the model parameters during the inference phase Through this process, the model parameters can be... Nearby The range of control is flexible and fluctuates, adapting to the complex optical characteristics of different water bodies, making the inference results more consistent with the diversity of the actual marine environment;

[0088] Based on the Bayesian inference framework, the Monte Carlo sampling method is used to repeatedly sample the parameters, thereby obtaining the estimated chlorophyll a concentration and its uncertainty quantification results, mainly including the standard deviation (SD) and CV, which are calculated as follows:

[0089] ;

[0090] ;

[0091] ;

[0092] in, , and These represent the mean, SD, and CV of the estimated chlorophyll a concentration, respectively. Furthermore, The sample variance of the estimated results is represented by the number of Monte Carlo samplings and the confidence level. and This indicates that the values ​​are set to 10 and 0.95 respectively. It is based on the specified confidence level and degrees of freedom The obtained critical value of the t-distribution;

[0093] The estimated chlorophyll a concentrations from different ocean color satellites were obtained. Compared with other publicly available chlorophyll a concentration data products, the proposed Bayesian deep learning model significantly reduced typical systematic biases, especially under extreme environmental conditions, where the obtained chlorophyll a concentration estimates showed a clear advantage. This indicates that the proposed Bayesian deep learning model can more accurately obtain chlorophyll a concentration results in complex marine environments, especially in high-latitude or extreme water quality environments, exhibiting more reliable adaptability and estimation capabilities. This proposed Bayesian deep learning model is ultimately termed a Bayesian deep learning model with hierarchical feature embedding.

[0094] In the validation set phase, S5012 uses the approximate mean of the variational posterior distribution of the parameters of the Bayesian deep learning model as the parameters of the Bayesian deep learning model during the validation phase. Input the validation set data into the model to obtain the prediction results. The Bayesian deep learning model’s current performance is evaluated by calculating the error between the predicted value and the measured value (such as MSE), and hyperparameters such as the regularization weight coefficient λ are adjusted accordingly. If the error does not improve for several consecutive rounds, early stopping is triggered to optimize the model’s generalization ability.

[0095] The Bayesian deep learning model training uses mean squared error (MSE) as the main loss function; mean squared error (MSE) is used as the regularization penalty term in the loss function during variational Bayesian learning.

[0096] The total loss function is: ;

[0097] in, These are regularization weight coefficients. The number of Monte Carlo samplings is indicated; Mean Squared Error (MSE): reflects the overall difference between the predicted and measured values; Variational Lower Bound (ELBO): serves as a "penalty term"; that is, Total Loss = MSE + Penalty Term (ELBO × coefficient);

[0098] In the training of the Bayesian deep learning model, the mean squared error (MSE) is first calculated. This is the average of the squared differences between the predicted values ​​of the Bayesian deep learning model and the measured chlorophyll a concentration values ​​on the training set, quantifying the overall gap between the two. Simultaneously, during variational Bayesian learning, the variational lower bound (ELBO) is calculated through Monte Carlo sampling (m times) to measure the closeness between the approximate parameter distribution and the true posterior distribution, and this is used as a regularization penalty term. Then, based on preset regularization weight coefficients... Sum of ELBO and Multiplying yields a penalty term, and finally, the MSE is added to this penalty term to obtain the total loss function. This process balances prediction accuracy with the rationality of parameter distribution, avoiding overfitting of the Bayesian deep learning model to noise in the training data, thus enabling the Bayesian deep learning model to maintain stable generalization ability in complex marine environments.

[0099] S5014. In the test set phase, after the Bayesian deep learning model has been trained, the test set data is input into the final model to obtain predicted values. To comprehensively evaluate the reliability of the method of this invention, it is compared with several benchmark chlorophyll a concentration inversion methods, including three widely recognized empirical algorithms: OCI-Hu, OCI-Wang, and CI-OC3M, and five machine learning models: Support Vector Machine (SVM), Random Forest (RF), Gradient Boosting Decision Tree (GBDT), eXtreme Gradient Boosting (XGBoost), and Bayesian Neural Network (BNN). During the evaluation process, the empirical algorithms are refitted on the matching dataset using the least squares method. The training process of the machine learning models is basically the same as that of this method, and the optimization strategies of these models are used.

[0100] S502. A comprehensive evaluation is conducted on the performance of all methods in retrieving the measured chlorophyll a concentration, analyzing its overall performance and performance under different nutrient levels. Evaluation indicators include the slope, bias, correlation coefficient (CC), mean absolute error (MAE), median absolute error (MdAE), mean absolute percentage error (MAPE), and root mean squared error (RMSE). Except for MAPE, all other indicators are calculated based on the logarithmically transformed measured chlorophyll a concentration values.

[0101] The evaluation of different methods in the measured chlorophyll a concentration retrieval task shows that, compared with other commonly used methods, the method of this invention performs exceptionally well on the MODIS-Aqua ocean color satellite dataset, especially demonstrating high reliability under different water trophic conditions. Although its reliability is slightly insufficient in mesotrophic waters, it performs particularly well in hypotrophic and hypertrophic waters, especially in extreme marine environments, where it has a significant advantage over other methods. Furthermore, although there are significant differences in sensor design, temporal and spatial distribution of measured chlorophyll a data samples, and sample proportions under different trophic conditions among different ocean color satellites, a Bayesian deep learning method for chlorophyll a concentration retrieval based on basic feature embedding is ultimately derived and named VB-OCformer.

[0102] Please see Figure 2 A chlorophyll concentration inversion system based on hierarchical feature embedding, used to process the aforementioned chlorophyll concentration inversion method based on hierarchical feature embedding, includes:

[0103] A data input module is set up to acquire spectral remote sensing reflectance data from ocean color satellites; and to acquire measured chlorophyll a concentration data within a preset ocean area, using the measured chlorophyll a concentration data as target reference data for training a Bayesian deep learning model.

[0104] A data processing module is set up to perform time and space matching between the spectral remote sensing reflectance data and the measured chlorophyll a concentration data, and integrate them to construct a dataset for training a Bayesian deep learning model.

[0105] The dataset is standardized to eliminate the influence of different units of measurement, resulting in a standardized dataset.

[0106] The standardization process is as follows: ;

[0107] in, i represents the i-th piece of raw data that has not undergone standardization; It is the mean of the dataset; It is the standard deviation of the dataset; This represents a standardized dataset;

[0108] The system includes a hierarchical feature embedding module, a loss calculation and optimization module, and an evaluation module. The standardized dataset is shuffled to form a matching dataset, which is then divided into three subsets: a training set, a validation set, and a test set, which are used for training, optimization, and evaluation of the Bayesian deep learning model, respectively.

[0109] The result output module is set up to obtain a Bayesian deep learning model with hierarchical feature embedding. The Bayesian deep learning model is used to obtain real-time chlorophyll a concentration data in the ocean area based on real-time spectral remote sensing reflectance data obtained by ocean color satellite.

[0110] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, computer program products, or electronic devices. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

Claims

1. A method for chlorophyll concentration inversion based on hierarchical feature embedding, characterized in that, include: Acquire spectral remote sensing reflectance data from ocean color satellites; Acquire measured chlorophyll a concentration data within a preset ocean range, and use the measured chlorophyll a concentration data as target reference data for training a Bayesian deep learning model. The spectral remote sensing reflectance data and the measured chlorophyll a concentration data are matched in time and space and integrated to form a dataset for training a Bayesian deep learning model. The dataset is standardized to eliminate the influence of different units of measurement, resulting in a standardized dataset. The standardization process is as follows: ; in, i This represents the i-th piece of raw data that has not undergone standardization. It is the mean of the dataset; It is the standard deviation of the dataset; This represents a standardized dataset; The standardized dataset is shuffled to form a matching dataset, and the matching dataset is divided into three subsets: training set, validation set, and test set; these are used for training, optimization, and evaluation of the Bayesian deep learning model, respectively, to obtain a Bayesian deep learning model with hierarchical feature embedding. To eliminate potential biases in the order, distribution, and local features of the data, the standardized dataset is shuffled to form the matching dataset. The matching dataset is divided into three subsets: training set, validation set, and test set. The matching dataset is sequentially entered into these three subsets, and the training, optimization, and evaluation of the Bayesian deep learning model are performed sequentially. During the training phase, the Bayesian deep learning model has an inference phase. The reasoning stage is processed as follows: ; in, The model parameters of the Bayesian deep learning model during the inference phase are represented. and This represents a variational approximation of the posterior distribution of the parameters; This represents the identity matrix sampled from the standard normal distribution; The Bayesian deep learning model training uses mean squared error (MSE) as the main loss function; mean squared error (MSE) is used as the regularization penalty term in the loss function during variational Bayesian learning. The total loss function is: ; in, These are regularization weight coefficients. The number of Monte Carlo samplings is indicated; Mean Squared Error (MSE): reflects the overall difference between the predicted and measured values; Variational Lower Bound (ELBO): serves as a "penalty term"; that is, Total Loss = MSE + Penalty Term (ELBO × Coefficient); When using the aforementioned verification set, the verification phase is processed as follows: ; The model parameters represent the Bayesian deep learning model during the validation phase; It is the approximate mean of the variational posterior distribution of the parameters; When using the test set, the performance of all methods in chlorophyll a concentration inversion was comprehensively evaluated, and their overall performance and performance under different nutrient levels were analyzed. Evaluation indicators included fit slope, fit bias, correlation coefficient, mean absolute error, median absolute error, and root mean square error. The Bayesian deep learning model is used to obtain real-time chlorophyll a concentration data in the ocean area based on real-time spectral remote sensing reflectance data acquired by the ocean color satellite.

2. The chlorophyll concentration inversion method based on hierarchical feature embedding according to claim 1, characterized in that, The step of acquiring measured chlorophyll a concentration data within a preset ocean area and using the measured chlorophyll a concentration data as target reference data for training a Bayesian deep learning model includes: Collect measured chlorophyll a concentration data within a pre-defined marine area; The collected measured chlorophyll a concentration data were cleaned and impurities removed. High performance liquid chromatography was used to measure chlorophyll a concentration data. Logarithmic transformation was performed on the processed measured chlorophyll a concentration data to obtain the target reference data.

3. The chlorophyll concentration inversion method based on hierarchical feature embedding according to claim 1, characterized in that, The spectral remote sensing reflectance data and the measured chlorophyll a concentration data are matched temporally and spatially, and integrated to construct a dataset for training a Bayesian deep learning model, including: The spectral remote sensing reflectance data corresponding to the measured chlorophyll a concentration data were extracted using a 3 × 3 pixel bounding box. Two different time windows, "daily window" and "8-day window", are used to match the data in time and space. A corresponding maximum coefficient of variation (CV) threshold is set for each time window. If the CV of the matched spectral remote sensing reflectance data exceeds the specified threshold, it is excluded.

4. The chlorophyll concentration inversion method based on hierarchical feature embedding according to claim 1, characterized in that, The dataset is standardized to eliminate the influence of different units of measurement, resulting in a standardized dataset, including: The standardization process is as follows: ; in, i This represents the i-th piece of raw data that has not undergone standardization. It is the mean of the dataset; It is the standard deviation of the dataset; This represents a standardized dataset; that is, the standardized data = (original data - mean) ÷ standard deviation.

5. The chlorophyll concentration inversion method based on hierarchical feature embedding according to claim 1, characterized in that, The inference operation of the Bayesian deep learning model includes: During the model inference phase, a standardized processing method consistent with the model training is adopted, and the spectral remote sensing reflectance data to be estimated is preprocessed using the constructed matching dataset; The following weight sampling method is used for model inference: ; in, These represent the model parameters during the inference process. and This represents a variational approximation of the posterior distribution of the parameters. This represents the identity matrix sampled from the standard normal distribution; Based on the Bayesian inference framework, the Monte Carlo sampling method is used to repeatedly sample the parameters, thereby obtaining the estimated chlorophyll a concentration and its uncertainty quantification results, mainly including the standard deviation (SD) and CV, which are calculated as follows: ； ； ； in, , and These represent the mean, SD, and CV of the estimated chlorophyll a concentration, respectively. The sample variance of the estimated results is represented by the number of Monte Carlo samplings and the confidence level. and This indicates that the values ​​are set to 10 and 0.95 respectively. It is based on the specified confidence level and degrees of freedom The obtained critical value of the t-distribution; The final estimated values ​​of chlorophyll a concentration from different ocean color satellites were compared with other publicly available chlorophyll a concentration data products.

6. The chlorophyll concentration inversion method based on hierarchical feature embedding according to claim 1, characterized in that, The evaluation of the Bayesian deep learning model includes: It is compared with a variety of chlorophyll a concentration inversion methods, including three widely accepted empirical algorithms: OCI-Hu, OCI-Wang, CI-OC3M, and five machine learning models: support vector machine, random forest, gradient boosting tree, extreme gradient boosting tree, and Bayesian neural network. During the evaluation process, the empirical algorithm refits the matching dataset using the least squares method; the training process of the machine learning model is basically the same as that of this method, and the respective optimization strategies of these models are used. A comprehensive evaluation of the chlorophyll a concentration inversion performance of all methods was conducted, and the overall performance and performance under different nutrient levels were analyzed.

7. A chlorophyll concentration inversion system based on hierarchical feature embedding, characterized in that, A method for processing chlorophyll concentration inversion based on hierarchical feature embedding as described in any one of claims 1-6 includes: A data input module is set up to acquire spectral remote sensing reflectance data from ocean color satellites; and to acquire measured chlorophyll a concentration data within a preset ocean area, using the measured chlorophyll a concentration data as target reference data for training a Bayesian deep learning model. A data processing module is set up to perform time and space matching between the spectral remote sensing reflectance data and the measured chlorophyll a concentration data, and integrate them to construct a dataset for training a Bayesian deep learning model. The dataset is standardized to eliminate the influence of different units of measurement, resulting in a standardized dataset. The standardization process is as follows: ; in, i represents the i-th piece of raw data that has not undergone standardization; It is the mean of the dataset; It is the standard deviation of the dataset; This represents a standardized dataset; The system includes a hierarchical feature embedding module, a loss calculation and optimization module, and an evaluation module. The standardized dataset is shuffled to form a matching dataset, which is then divided into three subsets: a training set, a validation set, and a test set, which are used for training, optimization, and evaluation of the Bayesian deep learning model, respectively. The result output module is set up to obtain a Bayesian deep learning model with hierarchical feature embedding. The Bayesian deep learning model is used to obtain real-time chlorophyll a concentration data in the ocean area based on real-time spectral remote sensing reflectance data obtained by ocean color satellite.

Images