A method and device for predicting chlorophyll concentration in the ocean surface

Abstract

The present invention provides a method and device for predicting chlorophyll concentration in the ocean surface, and relates to the technical field of ocean data analysis and processing; the method is to first import water body chlorophyll data and satellite remote sensing Rrs reflectivity, then perform time matching, standardization processing and logarithmic conversion, then screen satellite data set features to construct a target data set, and then integrate a chlorophyll concentration prediction model based on multiple gradient models, and finally input the satellite target data set and the in-situ data set as independent variables and dependent variables respectively into the model and MDN network prediction, and merge the results using a regression method. The present invention effectively extracts key information from complex data, improves model capture and prediction capabilities, makes up for feature deficiencies, reduces the impact of data distribution, improves the limitations of fusion algorithms, and uses data reasonably, the model is highly targeted, and the prediction results are scientific and practical.

Description

Technical Field

[0001] The present invention mainly relates to the technical field of ocean data analysis and processing, and in particular to a method and device for predicting chlorophyll concentration in the ocean surface. Background Art

[0002] In the research and monitoring of marine ecosystems, chlorophyll a concentration is a key indicator, which is directly related to the assessment of marine primary productivity, the health of the marine ecological environment, and important research areas such as the global carbon cycle. Accurately grasping the changes in chlorophyll a concentration is of irreplaceable significance for understanding the structure and function of marine ecosystems and predicting the evolution trend of marine ecological environment. Therefore, the development of a high-precision and reliable chlorophyll concentration prediction method has become an urgent need in the field of marine scientific research.

[0003] Traditional chlorophyll a concentration measurement mainly relies on field sampling and laboratory analysis. Although this method has high accuracy, it has obvious limitations. On the one hand, field sampling requires a lot of manpower, material resources and time costs, and can only obtain limited sampling point data, making it difficult to achieve large-scale, continuous monitoring; on the other hand, traditional methods have poor timeliness and cannot reflect the dynamic changes in chlorophyll a concentration in a timely manner, making it difficult to meet the needs of real-time monitoring and rapid assessment of marine ecosystems. With the continuous deepening of marine scientific research and the increasing awareness of marine ecological protection, traditional methods can no longer meet the requirements for comprehensive, timely and accurate monitoring of marine ecosystems.

[0004] The emergence of satellite remote sensing technology has provided a new means for monitoring chlorophyll a concentration. It can obtain spectral information of the ocean surface over a large area and continuously, thereby realizing macroscopic monitoring of chlorophyll a concentration. However, satellite remote sensing data faces many challenges in inverting chlorophyll a concentration. For example, satellite remote sensing Rrs reflectivity is interfered by many factors, including atmospheric scattering and absorption, suspended particles in water bodies, dissolved organic matter, etc. These factors make it difficult to accurately extract chlorophyll a concentration information. In addition, the relationship between satellite remote sensing Rrs reflectivity and chlorophyll a concentration in different bands is complex and nonlinear. Traditional analysis methods are difficult to fully explore the inherent connection, resulting in limited prediction accuracy. Therefore, how to effectively use satellite remote sensing data and improve the accuracy of chlorophyll a concentration prediction has become an urgent problem to be solved in the current field of ocean data analysis and processing technology. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a method and device for predicting the chlorophyll concentration of the ocean surface in view of the deficiencies of the prior art.

[0006] The technical solution of the present invention to solve the above technical problem is as follows: A method for predicting chlorophyll concentration in the ocean surface comprises the following steps:

[0007] S1. Import multiple pre-obtained water body chlorophyll data and multiple satellite remote sensing Rrs reflectances of different bands;

[0008] S2, setting the difference between multiple satellite remote sensing Rrs reflectivity and multiple water body chlorophyll data to be less than a preset time difference, completing time matching, standardizing the multiple satellite remote sensing Rrs reflectivity after time matching to obtain a satellite data set, and performing logarithmic transformation on the multiple water body chlorophyll data after time matching to obtain an in-situ data set;

[0009] S3, performing feature screening on the satellite data set, and constructing a satellite target data set through the screened satellite remote sensing Rrs reflectivity;

[0010] S4, integrated chlorophyll concentration prediction model based on XGBoost gradient model, LightBoost gradient model and CatBoost gradient model;

[0011] S5. Input the satellite target data set as an independent variable and the in-situ data set as a dependent variable into the chlorophyll concentration prediction model and the MDN mixed density network to predict chlorophyll concentration. Based on the regression method, the prediction results output by the chlorophyll concentration prediction model and the prediction results output by the MDN mixed density network are merged to obtain the chlorophyll concentration prediction result.

[0012] Another technical solution of the present invention to solve the above technical problem is as follows: a device for predicting chlorophyll concentration in the surface layer of the ocean, comprising:

[0013] The data import module is used to import pre-acquired chlorophyll data of multiple water bodies and multiple satellite remote sensing Rrs reflectances of different bands;

[0014] A preprocessing module is used to set the difference between multiple satellite remote sensing Rrs reflectances and multiple water body chlorophyll data to be less than a preset time difference, complete time matching, perform standardization processing on the multiple satellite remote sensing Rrs reflectances after time matching to obtain a satellite data set, and perform logarithmic transformation processing on the multiple water body chlorophyll data after time matching to obtain an in-situ data set;

[0015] A feature screening module is used to perform feature screening on the satellite data set and construct a satellite target data set through the screened satellite remote sensing Rrs reflectivity;

[0016] Model integration module, used to integrate chlorophyll concentration prediction models based on XGBoost gradient model, LightBoost gradient model and CatBoost gradient model;

[0017] The prediction processing module is used to input the satellite target data set as an independent variable and the in-situ data set as a dependent variable into the chlorophyll concentration prediction model and the MDN mixed density network to predict the chlorophyll concentration, and to merge the prediction results output by the chlorophyll concentration prediction model and the prediction results output by the MDN mixed density network based on the regression method to obtain the chlorophyll concentration prediction result.

[0018] Another technical solution of the present invention to solve the above-mentioned technical problem is as follows: a device for predicting the chlorophyll concentration of the ocean surface, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method for predicting the chlorophyll concentration of the ocean surface as described above is implemented.

[0019] The beneficial effects of the present invention are: 1) by importing pre-acquired water body chlorophyll data and satellite remote sensing Rrs reflectivity, and performing time matching, standardization processing and logarithmic transformation, a solid foundation is laid for subsequent model training and prediction. Time matching ensures the timeliness and relevance of the data, so that the satellite remote sensing data and the water body chlorophyll data are consistent in time, reducing the error caused by time difference. Standardization processing makes the satellite remote sensing Rrs reflectivity data have a unified scale, which is convenient for subsequent feature screening and model training. Logarithmic transformation optimizes the water body chlorophyll data to make it more in line with the assumptions and requirements of the model, improves the stability and interpretability of the data, and thus improves the accuracy and reliability of the entire prediction method.

[0020] 2) The method of feature screening, model integration and mixed density network is integrated to realize the comprehensive prediction of chlorophyll concentration changes. The method and device can extract more effective information from complex data, improve the model's ability to capture chlorophyll concentration changes and prediction accuracy, effectively make up for the lack of input features, reduce the negative impact of data distribution, and improve the limitations of complex fusion algorithms. By inputting the satellite target data set and the in-situ data set into the model as independent variables and dependent variables respectively, the effective use of data and the targeted training of the model are realized, making the prediction results more scientific and practical. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 A schematic diagram of a process for predicting chlorophyll concentration provided by an embodiment of the present invention;

[0022] Figure 2 Scatter density plots and evaluation indicators of each model provided in the embodiment of the present invention on the validation set A;

[0023] Figure 3 Scatter density plots and evaluation indicators of each model provided in the embodiment of the present invention on the validation set B;

[0024] Figure 4 A distribution diagram of prediction results of each model provided in an embodiment of the present invention on a validation set A;

[0025] Figure 5 A distribution diagram of prediction results of each model provided in an embodiment of the present invention on a validation set B;

[0026] Figure 6 Chla scatter distribution and difference diagram of SynthRidge and OCI at specified latitudes provided by the embodiment of the present invention;

[0027] Figure 7 A radar chart of performance indicators of each model provided by the embodiment of the present invention in the study area;

[0028] Figure 8 Normalized Taylor diagrams of the models provided in the embodiments of the present invention in the study area;

[0029] Fig. 9 The inversion result diagram of each model provided in the embodiment of the present invention at Roi_1;

[0030] Fig.10 The inversion result diagram of each model provided in the embodiment of the present invention in Roi_2;

[0031] Fig.11 This is a functional module block diagram of a chlorophyll concentration prediction device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0032] The principles and features of the present invention are described below in conjunction with the accompanying drawings. The examples given are only used to explain the present invention and are not used to limit the scope of the present invention.

[0033] The present invention aims to provide a method and device for predicting chlorophyll concentration in the ocean surface, which is suitable for analyzing chlorophyll a concentration in water bodies of different optical types, and improves the inversion consistency of chlorophyll a concentration by integrating machine learning feature engineering and model integration. The following is a detailed description through specific embodiments.

[0034] Embodiment 1:

[0035] like Figure 1 As shown, the embodiment of the present invention provides a method for predicting chlorophyll concentration in the ocean surface, comprising the following steps:

[0036] S1. Import multiple pre-obtained water body chlorophyll data and multiple satellite remote sensing Rrs reflectances of different bands;

[0037] S2, setting the difference between multiple satellite remote sensing Rrs reflectivity and multiple water body chlorophyll data to be less than a preset time difference, completing time matching, standardizing the multiple satellite remote sensing Rrs reflectivity after time matching to obtain a satellite data set, and performing logarithmic transformation on the multiple water body chlorophyll data after time matching to obtain an in-situ data set;

[0038] S3, performing feature screening on the satellite data set, and constructing a satellite target data set through the screened satellite remote sensing Rrs reflectivity;

[0039] S4, integrated chlorophyll concentration prediction model based on XGBoost gradient model, LightBoost gradient model and CatBoost gradient model;

[0040] S5. Input the satellite target data set as an independent variable and the in-situ data set as a dependent variable into the chlorophyll concentration prediction model and the MDN mixed density network to predict chlorophyll concentration. Based on the regression method, the prediction results output by the chlorophyll concentration prediction model and the prediction results output by the MDN mixed density network are merged to obtain the chlorophyll concentration prediction result.

[0041] In the above embodiments, 1) by importing pre-acquired water body chlorophyll data and satellite remote sensing Rrs reflectivity, and performing time matching, standardization processing and logarithmic transformation, a solid foundation is laid for subsequent model training and prediction. Time matching ensures the timeliness and relevance of the data, so that the satellite remote sensing data and the water body chlorophyll data are consistent in time, reducing the error caused by time difference. Standardization processing makes the satellite remote sensing Rrs reflectivity data have a unified scale, which is convenient for subsequent feature screening and model training. Logarithmic transformation optimizes the water body chlorophyll data to make it more in line with the assumptions and requirements of the model, improves the stability and interpretability of the data, and thus improves the accuracy and reliability of the entire prediction method.

[0042] 2) The method of feature screening, model integration and mixed density network is integrated to realize the comprehensive prediction of chlorophyll concentration changes. The method and device can extract more effective information from complex data, improve the model's ability to capture chlorophyll concentration changes and prediction accuracy, effectively make up for the lack of input features, reduce the negative impact of data distribution, and improve the limitations of complex fusion algorithms. By inputting the satellite target data set and the in-situ data set into the model as independent variables and dependent variables respectively, the effective use of data and the targeted training of the model are realized, making the prediction results more scientific and practical.

[0043] Preferably, in step S1, the plurality of water body chlorophyll data are obtained by measuring the ocean surface using a SeaWiFS instrument;

[0044] The satellite remote sensing Rrs reflectivity of multiple bands is the satellite remote sensing Rrs reflectivity of different bands of the ocean based on Terra satellite images downloaded from the specified platform. Specifically, the 10 bands of Rrs data of the ocean based on Terra satellite images and the Rrs values ​​measured by the SeaWiFS instrument are downloaded. The 10 bands are 412nm, 443nm, 469nm, 488nm, 531nm, 547nm, 555nm, 645nm, 667nm and 678nm, and the values ​​measured by SeaWiFS are 412nm, 443nm, 490nm, 510nm, 555nm and 670nm respectively.

[0045] In the above embodiment, 1) it is clear that the water chlorophyll data comes from the measurement of the ocean surface by the SeaWiFS instrument. This measurement method has high accuracy and pertinence, and can obtain real data reflecting the chlorophyll a concentration (i.e., InsituChla) of the ocean surface. The SeaWiFS instrument is widely used in the field of marine chlorophyll measurement, and its measurement results have certain authority and reliability, providing high-quality basic data for subsequent model training, which helps to improve the accuracy of model prediction.

[0046] 2) Specify the acquisition of satellite remote sensing Rrs reflectivity from different bands of the ocean based on Terra satellite images. Terra satellite images have high spatial resolution and spectral resolution, and can provide rich water optical information. Reflectivity data in different bands can reflect the optical properties of different substances in the water body, and have a potential correlation with chlorophyll a concentration. Selecting a suitable source of satellite remote sensing data makes the acquired data more adaptable to the chlorophyll concentration prediction needs of the present invention, enhances the adaptability of the model to different water optical types, and improves the versatility of the prediction method in different marine environments.

[0047] Preferably, in step S2, logarithmic transformation is performed on the time-matched chlorophyll data of the multiple water bodies, including:

[0048] The chlorophyll data of multiple water bodies after time matching were converted into log10 logarithmic form.

[0049] In step S2, the difference between multiple satellite remote sensing Rrs reflectivity and multiple water body chlorophyll data is set to be less than a preset time difference, and time matching is completed. Specifically, the time matching difference is set to be less than 3 hours.

[0050] In step S2, the multiple satellite remote sensing Rrs reflectivities after time matching are standardized to obtain a satellite data set. Specifically, the multiple satellite remote sensing Rrs reflectivities are feature expanded according to the set standardization requirements to complete the standardization process.

[0051] In the above embodiment, the log10 logarithmic transformation of the water body chlorophyll data after time matching has many benefits. First, the logarithmic transformation can make the distribution of data closer to the normal distribution, meet the assumptions of many statistical analysis and machine learning models, and improve the fitting effect of the model. Secondly, the logarithmic transformation can compress the scale of the data, reduce the heteroscedasticity of the data, make the change trend of the data in different concentration ranges more obvious, and help the model better capture the characteristics of chlorophyll concentration changes. In addition, the logarithmic transformation can also reduce the impact of outliers on the model, improve the stability and robustness of the model, and thus improve the accuracy of chlorophyll concentration prediction.

[0052] Preferably, in step S3, the satellite data set is subjected to feature screening, and a satellite target data set is constructed through the screened satellite remote sensing Rrs reflectivity, including:

[0053] The characteristic score corresponding to each satellite remote sensing Rrs reflectivity in the satellite data set is calculated using the Z-Score calculation formula, and the Z-Score calculation formula is:

[0054]

[0055] Among them, z(i) is the feature score of the i-th satellite remote sensing Rrs reflectivity, x(i) is the score of the i-th feature in the satellite dataset, μ is the average score of the feature in the satellite dataset, and σ is the standard deviation of the satellite remote sensing Rrs reflectivity in the satellite dataset;

[0056] The importance score of each satellite remote sensing Rrs reflectivity in the satellite data set is calculated using the importance comparison formula, and the importance comparison formula is:

[0057] imp(i)=RF(i)-max(M(i)),

[0058] Where imp(i) is the importance score of the i-th feature, RF(i) is the importance score of the i-th feature in the satellite dataset, and M(i) is the maximum importance score of the i-th feature in the satellite dataset;

[0059] The SHAP calculation formula is used to calculate the SHAP value of each satellite remote sensing Rrs reflectivity in the satellite data set. The SHAP calculation formula is:

[0060] SHAP i =∑[p(s)*(f(s U{x_i})-f(s)],

[0061] Among them, SHAP irepresents the SHAP value of feature x_i, p(s) represents the weighted probability of all possible permutations of feature subset s, f(s) represents the output of the prediction model on feature subset s, and f(s U{x_i}) represents the predicted output after adding feature x_i to feature subset s;

[0062] According to the preset values ​​and the results of the feature scores, importance scores and SHAP values ​​corresponding to each satellite remote sensing Rrs reflectivity, the satellite remote sensing Rrs reflectivity with high importance is screened out, and the satellite target data set is constructed through the screened satellite remote sensing Rrs reflectivity.

[0063] For example, if imp(i) is greater than 0 (the default value), it means that the feature has significant importance in the satellite dataset. If imp(i) is less than or equal to 0, it means that the feature is not significant in the satellite dataset.

[0064] It should be understood that BorutaShap is a machine learning tool for feature selection and interpretation, which combines Boruta's feature selection capabilities with SHAP's interpretation capabilities to provide more comprehensive feature selection and interpretation results. Its core is mainly Boruta's Z-Score calculation and importance comparison and SHAP value calculation.

[0065] In the above embodiment, 1) the satellite data set is feature screened using methods such as Z-Score calculation, importance comparison and SHAP value calculation, which can accurately identify the satellite remote sensing Rrs reflectivity features that are closely related to the change in chlorophyll concentration. Z-Score calculation can measure the relative importance of each feature in the data set, importance comparison further determines the significance of the feature, and SHAP value calculation can deeply explain the contribution of the feature to the model prediction results. Through this multi-dimensional feature evaluation, the features with high importance screened out can more effectively represent the relationship between the chlorophyll a concentration of the water body and the satellite remote sensing Rrs reflectivity, reduce the interference of redundant and irrelevant features, and improve the training efficiency and prediction accuracy of the model.

[0066] 2) The satellite target dataset is constructed by screening the satellite remote sensing Rrs reflectivity, making the dataset more streamlined and targeted. The optimized satellite target dataset can better reflect the intrinsic relationship between the optical properties of water bodies and the changes in chlorophyll concentration, providing better data input for subsequent model training, helping the model to more accurately learn and capture the laws in the data, thereby improving the accuracy of chlorophyll concentration prediction.

[0067] Preferably, in step S4, a chlorophyll concentration prediction model is integrated based on the XGBoost gradient model, the LightBoost gradient model and the CatBoost gradient model, including:

[0068] The initial chlorophyll concentration prediction model was obtained by integrating the XGBoost gradient model, the LightBoost gradient model and the CatBoost gradient model by voting.

[0069] The hyperparameter networks in the XGBoost gradient model, the LightBoost gradient model, and the CatBoost gradient model were set to obtain a chlorophyll concentration prediction model.

[0070] It should be understood that by integrating the XGBoost, LightBoost and CatBoost models by voting to obtain GBDT (ensemble), setting the hyperparameter network of three different gradient boosting tree algorithms, XGBoost, LightBoost and CatBoost, to find the optimal hyperparameter combination, and further improve the performance of the GBDT (ensemble) model.

[0071] In the above embodiment, 1) the XGBoost, LightBoost and CatBoost models are integrated by voting to obtain the initial chlorophyll concentration prediction model, and the optimal combination is found by setting the hyperparameter network, which significantly improves the performance of the model. The voting method can integrate the prediction results of multiple models, give full play to the advantages of each model, and improve the stability and accuracy of the prediction. The optimization and adjustment of the hyperparameter network enables the model to better adapt to the data characteristics, further explore the potential information in the data, and enhance the model's ability to model complex nonlinear relationships, thereby improving the prediction ability of chlorophyll concentration changes.

[0072] 2) Build an integrated model based on different gradient boosting tree algorithms so that the model can adapt to different types of data distribution and feature relationships. XGBoost, LightBoost and CatBoost have their own characteristics in processing large-scale data and complex nonlinear relationships. Their integration can make up for the limitations of a single model and improve the model's adaptability to different water environments and chlorophyll concentration change patterns. This model integration strategy increases the flexibility and robustness of the prediction method, enabling it to achieve better prediction results in various practical application scenarios.

[0073] Preferably, in step S5, the prediction result output by the chlorophyll concentration prediction model and the prediction result output by the MDN mixed density network are combined based on a regression method to obtain a chlorophyll concentration prediction result, including:

[0074] The regression coefficient is calculated using the loss function formula group, which is:

[0075] coefficients=argmin(‖Xβ-y‖∧2+α‖β‖∧2),

[0076] coefficients=(XTX+αI)∧(-1)XTy,

[0077] Where coefficients is the regression coefficient, X is the characteristic matrix of the satellite target data set, y is the target variable, β is the model parameter vector, α is the regularization parameter, XTX is the product of the transpose of X and X, I is the identity matrix, αI is the matrix with α as the diagonal element, and XTy is the product of the transpose of X and y;

[0078] The prediction results output by the chlorophyll concentration prediction model and the prediction results output by the MDN mixed density network are combined into y merge , multiply the regression coefficient by the regression coefficient to obtain the chlorophyll concentration prediction result:

[0079] y ridge =y merge *coefficients.

[0080] Specifically, the input neurons of the MDN mixed density network are set to 1000, and the activation function of the input layer is set to 'relu'; the number of hidden layers is 5, the number of neurons in each layer is set to 100, and the activation function of the hidden layer is set to 'relu'; the optimization algorithm selects 'adam'.

[0081] The Ridge Regression method is used to construct the optimal prediction model, and the regularization term is introduced to solve the overfitting problem that may exist in the ordinary least squares method. The basic idea is to add a penalty term to the loss function to limit the size of the model parameters. The goal of ridge regression is to find the optimal model parameters to minimize the loss function.

[0082] In this embodiment, the Ridge regression method is used to construct the optimal prediction model, and the regularization term is introduced to solve the overfitting problem that may exist in the ordinary least squares method. When calculating the ridge regression coefficient, the model parameters are determined by minimizing a specific loss function (the loss function formula group provided in this embodiment), wherein the regularization parameter is set to 0.1 (setting in this embodiment). In this way, ridge regression can effectively avoid the model from overfitting complex data relationships when processing satellite remote sensing data and water body chlorophyll data, improve the generalization ability of the model, and thus more accurately predict the change in chlorophyll concentration. The regularization characteristics of ridge regression make the model more robust to noise and outliers in the data. In the actual marine environment, satellite remote sensing data and water body chlorophyll data may be disturbed by various factors, and there are certain noise and outliers. Ridge regression reduces the impact of these abnormal data on the model by constraining the coefficients, making the prediction results of the model more stable and reliable. This helps to maintain the consistency of model performance under different data sets and environmental conditions, and improves the practicality and effectiveness of the chlorophyll concentration prediction method.

[0083] In the above embodiment, 1) the prediction results of the chlorophyll concentration prediction model and the MDN mixed density network are merged based on the regression method, which makes full use of the advantages of the two models and improves the accuracy and reliability of the prediction results. By reasonably designing the loss function formula group to calculate the regression coefficient, the prediction results of the two models are effectively integrated, so that the final prediction result can comprehensively consider the influence of multiple factors and reduce the possible deviation of a single model. This merging processing method can better capture the trend and uncertainty of chlorophyll concentration changes, and provide more comprehensive and accurate prediction information for marine ecological research and management.

[0084] 2) Reasonable parameter settings are made for the MDN network, such as input neurons, activation functions, hidden layer structures, and optimization algorithms, so that it can work better with the chlorophyll concentration prediction model. The characteristics of the MDN network enable it to handle data uncertainty. Combined with the integrated chlorophyll concentration prediction model, it further enriches the model's expressive power and improves its ability to fit complex data distributions. The introduction of regularization terms through the ridge regression method effectively solves the overfitting problem, enhances the generalization ability of the model, ensures that the model can maintain good prediction performance on different data sets, and improves the stability and reliability of the entire prediction method.

[0085] The above is a comprehensive introduction to the chlorophyll concentration prediction method of the present invention. The following is an evaluation of the accuracy of the chlorophyll concentration prediction model in the present method.

[0086] First, construct model evaluation indicators.

[0087] Using R 2To evaluate the goodness of fit of the regression model, RMSLE is used to measure the difference between the predicted value and the true value, while MAE, Bias, and Slope calculate the average of the absolute errors between the predicted value and the true value, the average error of the predicted value, and the linear correlation degree, respectively. In addition, the present invention also calculates the RMSE of the Chla concentration without logarithmic transformation to evaluate the magnitude of the standard deviation. Their formulas are as follows:

[0088]

[0089]

[0090] Taking the satellite target dataset (including satellite remote sensing Rrs reflectance) as validation set A and the in-situ dataset (including water body chlorophyll data) as validation set B, first, the performance of the chlorophyll concentration prediction model is evaluated using the constructed evaluation indicators, and secondly, the generality of the chlorophyll concentration prediction model on different sensors is evaluated.

[0091] 1) Calculate the evaluation indicators for 5 common and widely recognized algorithms in ocean chlorophyll a inversion, namely OCI, SVR, RFR, GBDT (ensemble), MDN, and SynthRidge, so as to conduct an overall performance evaluation.

[0092] 2) Based on the validation sets of validation set A and validation set B, scatter density plots (as shown in Figure 2 and Figure 3 ), predicted concentration distribution plots (as shown in Figure 4 and Figure 5 ) are drawn, and they are divided into oligotrophic water bodies (Chla ≤ 0.1 mg / m3), mesotrophic water bodies (0.1 < Chla ≤ 1 mg / m3), and eutrophic water bodies (Chla > 1 mg / m3) according to the trophic level to evaluate the performance of SnthRidge in different nutrient areas.

[0093] 3) On the monthly time scale, based on the Terra-MODIS level 3 product data (February 2004), the Chla spatial mapping capabilities of each model are compared. Draw the Chla scatter distribution and difference plots of SynthRidge and OCI at the specified latitude (as shown in Figure 6 ).

[0094] 4) In the coastal waters, due to the presence of suspended solids, pigments, and dissolved organic matter, as well as the influence of waves, vortices, river inputs, and tides, the propagation and reflection of light become complex. In order to further analyze the performance of SynthRidge in optically complex waters (since OCI performs poorly in coastal waters, it is not included in the comparison), the present invention selects two coastal waters as the study areas (Roi_1, Roi_2) and draws a radar chart (as shown in Figure 7As shown) predict the normalized Taylor diagram (as shown Figure 8 shown).

[0095] 5) Taking the study area as the scope, the spatial scale was reduced, and based on the Terra-MODIS level 3 product data in February 2004, the spatial mapping capabilities of each model for Chla in Roi_1 and Roi_2 areas were compared on a monthly time scale. At the same time, the standard deviation stretching technique was used to process the image, and the results are shown in Fig. 9 and Fig.10 middle.

[0096] Embodiment 2:

[0097] like Fig.11 As shown, another embodiment of the present invention provides a device for predicting chlorophyll concentration in the ocean surface, comprising:

[0098] The data import module is used to import pre-acquired chlorophyll data of multiple water bodies and multiple satellite remote sensing Rrs reflectances of different bands;

[0099] A preprocessing module is used to set the difference between multiple satellite remote sensing Rrs reflectances and multiple water body chlorophyll data to be less than a preset time difference, complete time matching, perform standardization processing on the multiple satellite remote sensing Rrs reflectances after time matching to obtain a satellite data set, and perform logarithmic transformation processing on the multiple water body chlorophyll data after time matching to obtain an in-situ data set;

[0100] A feature screening module is used to perform feature screening on the satellite data set and construct a satellite target data set through the screened satellite remote sensing Rrs reflectivity;

[0101] Model integration module, used to integrate chlorophyll concentration prediction models based on XGBoost gradient model, LightBoost gradient model and CatBoost gradient model;

[0102] The prediction processing module is used to input the satellite target data set as an independent variable and the in-situ data set as a dependent variable into the chlorophyll concentration prediction model and the MDN mixed density network to predict the chlorophyll concentration, and to merge the prediction results output by the chlorophyll concentration prediction model and the prediction results output by the MDN mixed density network based on the regression method to obtain the chlorophyll concentration prediction result.

[0103] Preferably, in the feature screening module, the satellite data set is subjected to feature screening, and a satellite target data set is constructed through the screened satellite remote sensing Rrs reflectivity, including:

[0104] The characteristic score corresponding to each satellite remote sensing Rrs reflectivity in the satellite data set is calculated using the Z-Score calculation formula, and the Z-Score calculation formula is:

[0105]

[0106] Among them, Z(i) is the feature score of the i-th satellite remote sensing Rrs reflectivity, x(i) is the score of the i-th feature in the satellite dataset or satellite dataset, μ is the average score of the feature in the satellite dataset, and σ is the standard deviation of the satellite remote sensing Rrs reflectivity in the satellite dataset;

[0107] The importance score of each satellite remote sensing Rrs reflectivity in the satellite data set is calculated using the importance comparison formula, and the importance comparison formula is:

[0108] imp(i)=RF(i)-max(M(i)),

[0109] Where imp(i) is the importance score of the i-th feature, RF(i) is the importance score of the i-th feature in the satellite dataset, and M(i) is the maximum importance score of the i-th feature in the satellite dataset;

[0110] The SHAP calculation formula is used to calculate the SHAP value of each satellite remote sensing Rrs reflectivity in the satellite data set. The SHAP calculation formula is:

[0111] SHAP i =∑[p(s)*(f(s U{x_i})-f(s)],

[0112] Among them, SHAP i represents the SHAP value of feature x_i, p(s) represents the weighted probability of all possible permutations of feature subset s, f(s) represents the output of the prediction model on feature subset s, and f(s U{x_i}) represents the predicted output after adding feature x_i to feature subset s;

[0113] According to the preset values ​​and the results of the feature scores, importance scores and SHAP values ​​corresponding to each satellite remote sensing Rrs reflectivity, the satellite remote sensing Rrs reflectivity with high importance is screened out, and the satellite target data set is constructed through the screened satellite remote sensing Rrs reflectivity.

[0114] Preferably, in the model integration module, a chlorophyll concentration prediction model is integrated based on the XGBoost gradient model, the LightBoost gradient model and the CatBoost gradient model, including:

[0115] The initial chlorophyll concentration prediction model was obtained by integrating the XGBoost gradient model, the LightBoost gradient model and the CatBoost gradient model by voting.

[0116] The hyperparameter networks in the XGBoost gradient model, the LightBoost gradient model, and the CatBoost gradient model were set to obtain a chlorophyll concentration prediction model.

[0117] Embodiment 3:

[0118] Another embodiment of the present invention provides a device for predicting chlorophyll concentration in the ocean surface, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the chlorophyll concentration prediction method described above is implemented.

[0119] The advantages of the present invention are:

[0120] 1) The band ratio method and BorutaShap were used to expand and screen features, and then an integration strategy was adopted to combine the three gradient boosting decision tree models of XGBoost, LightBoost and CatBoost with MDN to obtain a chlorophyll concentration prediction model (also known as the Synth Ridge Framework model (abbreviated as SRF)). SRF can effectively make up for the lack of input features, reduce the negative impact of data distribution, and improve the limitations of complex fusion algorithms. In addition, the performance of SRF on the SeaWiFS dataset proves its versatility on different sensors. Therefore, SRF can be regarded as a framework that can be used to achieve unified inversion of Chla across different optical types of water bodies.

[0121] 2) By importing the pre-acquired water body chlorophyll data and satellite remote sensing Rrs reflectivity, and performing time matching, standardization and logarithmic transformation, a solid foundation is laid for subsequent model training and prediction. Time matching ensures the timeliness and relevance of the data, keeps the satellite remote sensing data and the water body chlorophyll data consistent in time, and reduces the error caused by time difference. Standardization makes the satellite remote sensing Rrs reflectivity data have a unified scale, which is convenient for subsequent feature screening and model training. Logarithmic transformation optimizes the water body chlorophyll data to make it more in line with the assumptions and requirements of the model, improves the stability and interpretability of the data, and thus improves the accuracy and reliability of the entire prediction method.

[0122] 3) The method of feature screening, model integration and mixed density network is integrated to realize the comprehensive prediction of chlorophyll concentration changes. This method and device can extract more effective information from complex data, improve the model's ability to capture chlorophyll concentration changes and prediction accuracy, effectively make up for the lack of input features, reduce the negative impact of data distribution, and improve the limitations of complex fusion algorithms. By inputting the satellite target data set and the in-situ data set into the model as independent variables and dependent variables respectively, the effective use of data and the targeted training of the model are realized, making the prediction results more scientific and practical.

[0123] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described devices and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0124] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of units is only a logical function division, and there may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed.

[0125] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the embodiments of the present invention.

[0126] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for predicting chlorophyll concentration in the ocean surface, characterized in that: The steps include: S1. Import multiple pre-obtained water body chlorophyll data and multiple satellite remote sensing Rrs reflectances of different bands; S2, setting the difference between multiple satellite remote sensing Rrs reflectivity and multiple water body chlorophyll data to be less than a preset time difference, completing time matching, standardizing the multiple satellite remote sensing Rrs reflectivity after time matching to obtain a satellite data set, and performing logarithmic transformation on the multiple water body chlorophyll data after time matching to obtain an in-situ data set; S3, performing feature screening on the satellite data set, and constructing a satellite target data set through the screened satellite remote sensing Rrs reflectivity; S4, integrated chlorophyll concentration prediction model based on XGBoost gradient model, LightBoost gradient model and CatBoost gradient model; S5. Input the satellite target data set as an independent variable and the in-situ data set as a dependent variable into the chlorophyll concentration prediction model and the MDN mixed density network to predict chlorophyll concentration. Based on the regression method, the prediction results output by the chlorophyll concentration prediction model and the prediction results output by the MDN mixed density network are merged to obtain the chlorophyll concentration prediction result.

2. The chlorophyll concentration prediction method according to claim 1, characterized in that: In step S1, the chlorophyll data of the plurality of water bodies are obtained by measuring the ocean surface using the SeaWiFS instrument; The satellite remote sensing Rrs reflectivity of multiple bands in different bands is the satellite remote sensing Rrs reflectivity of different bands of the ocean based on Terra satellite images downloaded from the specified platform.

3. The chlorophyll concentration prediction method according to claim 1, characterized in that: In the step S3, the satellite data set is subjected to feature screening, and a satellite target data set is constructed through the screened satellite remote sensing Rrs reflectivity, including: The characteristic score corresponding to each satellite remote sensing Rrs reflectivity in the satellite data set is calculated using the Z-Score calculation formula, and the Z-Score calculation formula is: Among them, Z(i) is the feature score of the i-th satellite remote sensing Rrs reflectivity, x(i) is the score of the i-th feature in the satellite dataset, μ is the average score of the feature in the satellite dataset, and σ is the standard deviation of the satellite remote sensing Rrs reflectivity in the satellite dataset; The importance score of each satellite remote sensing Rrs reflectivity in the satellite data set is calculated using the importance comparison formula, and the importance comparison formula is: imp(i)=RF(i)-max(M(i)), Where imp(i) is the importance score of the i-th feature, RF(i) is the importance score of the i-th feature in the satellite dataset, and M(i) is the maximum importance score of the i-th feature in the satellite dataset; The SHAP calculation formula is used to calculate the SHAP value of each satellite remote sensing Rrs reflectivity in the satellite data set. The SHAP calculation formula is: SHAP i =∑[p(s)*(f(s U{x_i})-f(s)], Among them, SHAP i represents the SHAP value of feature x_i, p(s) represents the weighted probability of all possible permutations of feature subset s, f(s) represents the output of the prediction model on feature subset s, and f(s U{x_i}) represents the predicted output after adding feature x_i to feature subset s; According to the preset values ​​and the results of the feature scores, importance scores and SHAP values ​​corresponding to each satellite remote sensing Rrs reflectivity, the satellite remote sensing Rrs reflectivity with high importance is screened out, and the satellite target data set is constructed through the screened satellite remote sensing Rrs reflectivity.

4. The chlorophyll concentration prediction method according to claim 1, characterized in that: In step S4, a chlorophyll concentration prediction model is integrated based on the XGBoost gradient model, the LightBoost gradient model and the CatBoost gradient model, including: The initial chlorophyll concentration prediction model was obtained by integrating the XGBoost gradient model, the LightBoost gradient model and the CatBoost gradient model by voting. The hyperparameter networks in the XGBoost gradient model, the LightBoost gradient model, and the CatBoost gradient model were set to obtain a chlorophyll concentration prediction model.

5. The chlorophyll concentration prediction method according to claim 1, characterized in that: In step S5, the prediction result output by the chlorophyll concentration prediction model and the prediction result output by the MDN mixed density network are combined based on a regression method to obtain a chlorophyll concentration prediction result, including: The regression coefficient is calculated using the loss function formula group, which is: coefficients=argmin(‖Xβ-y‖ ∧ 2+α‖β‖ ∧ 2), coefficients=(XTX+αΙ)∧(-1)XTy, Where coefficients is the regression coefficient, X is the characteristic matrix of the satellite target data set, y is the target variable, β is the model parameter vector, α is the regularization parameter, XTX is the product of the transpose of X and X, I is the identity matrix, αI is the matrix with α as the diagonal element, and XTy is the product of the transpose of X and y; The prediction results output by the chlorophyll concentration prediction model and the prediction results output by the MDN mixed density network are combined into y merge , multiply the regression coefficient by the regression coefficient to obtain the chlorophyll concentration prediction result: and ridge =and merge *coefficients, Among them, y ridge This is the predicted result of chlorophyll concentration.

6. The chlorophyll concentration prediction method according to claim 1, characterized in that: In step S2, logarithmic transformation is performed on the multiple water body chlorophyll data after time matching, including: The chlorophyll data of multiple water bodies after time matching were converted into log10 logarithmic form.

7. A device for predicting chlorophyll concentration in the ocean surface, characterized in that: include: The data import module is used to import pre-acquired chlorophyll data of multiple water bodies and multiple satellite remote sensing Rrs reflectances of different bands; A preprocessing module is used to set the difference between multiple satellite remote sensing Rrs reflectances and multiple water body chlorophyll data to be less than a preset time difference, complete time matching, perform standardization processing on the multiple satellite remote sensing Rrs reflectances after time matching to obtain a satellite data set, and perform logarithmic transformation processing on the multiple water body chlorophyll data after time matching to obtain an in-situ data set; A feature screening module is used to perform feature screening on the satellite data set and construct a satellite target data set through the screened satellite remote sensing Rrs reflectivity; Model integration module, used to integrate chlorophyll concentration prediction models based on XGBoost gradient model, LightBoost gradient model and CatBoost gradient model; The prediction processing module is used to input the satellite target data set as an independent variable and the in-situ data set as a dependent variable into the chlorophyll concentration prediction model and the MDN mixed density network to predict the chlorophyll concentration, and to merge the prediction results output by the chlorophyll concentration prediction model and the prediction results output by the MDN mixed density network based on the regression method to obtain the chlorophyll concentration prediction result.

8. The chlorophyll concentration prediction device according to claim 7, characterized in that: In the feature screening module, the satellite data set is subjected to feature screening, and a satellite target data set is constructed through the screened satellite remote sensing Rrs reflectivity, including: The characteristic score corresponding to each satellite remote sensing Rrs reflectivity in the satellite data set is calculated using the Z-Score calculation formula, and the Z-Score calculation formula is: Among them, z(i) is the feature score of the i-th satellite remote sensing Rrs reflectivity, x(i) is the score of the i-th feature in the satellite dataset or satellite dataset, μ is the average score of the feature in the satellite dataset, and σ is the standard deviation of the satellite remote sensing Rrs reflectivity in the satellite dataset; The importance score of each satellite remote sensing Rrs reflectivity in the satellite data set is calculated using the importance comparison formula, and the importance comparison formula is: imp(i)=RF(i)-max(M(i)), Where imp(i) is the importance score of the i-th feature, RF(i) is the importance score of the i-th feature in the satellite dataset, and M(i) is the maximum importance score of the i-th feature in the satellite dataset; The SHAP calculation formula is used to calculate the SHAP value of each satellite remote sensing Rrs reflectivity in the satellite data set. The SHAP calculation formula is: SHAP i =∑[p(s)*(f(s∪{x_i})-f(s)], Among them, SHAP i represents the SHAP value of feature x_i, p(s) represents the weighted probability of all possible permutations of feature subset s, f(s) represents the output of the prediction model on feature subset s, and f(s U{x_i}) represents the predicted output after adding feature x_i to feature subset s; According to the preset values ​​and the results of the feature scores, importance scores and SHAP values ​​corresponding to each satellite remote sensing Rrs reflectivity, the satellite remote sensing Rrs reflectivity with high importance is screened out, and the satellite target data set is constructed through the screened satellite remote sensing Rrs reflectivity.

9. The chlorophyll concentration prediction device according to claim 7, characterized in that: In the model integration module, the chlorophyll concentration prediction model is integrated based on the XGBoost gradient model, the LightBoost gradient model and the CatBoost gradient model, including: The initial chlorophyll concentration prediction model was obtained by integrating the XGBoost gradient model, the LightBoost gradient model and the CatBoost gradient model by voting. The hyperparameter networks in the XGBoost gradient model, the LightBoost gradient model, and the CatBoost gradient model were set to obtain a chlorophyll concentration prediction model.

10. A device for predicting chlorophyll concentration in the ocean surface, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, it implements the method for predicting the chlorophyll concentration of the ocean surface as described in any one of claims 1 to 6.

Images