Method and system for remote sensing inversion of coastal nutrients and chemical oxygen demand

Abstract

The present application belongs to the technical field of remote sensing monitoring of water pollution, and is a remote sensing inversion method and system for nitrogen and phosphorus nutrients and chemical oxygen demand in near-shore sea areas. Remote sensing data, reanalysis data and in-situ water quality monitoring data are obtained, and spatio-temporal matching is performed to construct a star-ground spatio-temporal data set. Spatio-temporal information is added, and the selected features are expanded through feature engineering to construct a star-ground spatio-temporal big data set that integrates spatio-temporal features and marine physicochemical properties. A basic model is constructed to determine the input features for predicting inorganic nitrogen, active phosphate and chemical oxygen demand, respectively, to form water quality parameter models, and to train and optimize the inversion models of the water quality parameters to depict the complex nonlinear relationship between nitrogen and phosphorus nutrients and chemical oxygen demand in water bodies and satellite data, marine physicochemical data and spatio-temporal information. The present application realizes long-term, spatially fully covered, and retroactive water quality remote sensing monitoring with high observation frequency.

Description

Technical Field

[0001] This invention belongs to the field of water pollution remote sensing monitoring technology, specifically relating to a remote sensing inversion method and system for nitrogen and phosphorus nutrients and chemical oxygen demand in nearshore waters. Background Technology

[0002] Nearshore waters are inextricably linked to human life, playing a vital role in recreation, transportation, human health, and economic development. However, since the mid-20th century, the discharge of industrial wastewater and domestic sewage, as well as the inflow of agricultural fertilizers, have led to the degradation of nearshore water quality, a sharp increase in eutrophication, severe damage to ecosystem functions, and the exponential spread of oxygen-deficient areas in coastal waters. This has resulted in frequent marine disasters such as red tides, seriously impacting marine water safety and the lives and livelihoods of people along the coast. Therefore, the deterioration of the nearshore marine environment has attracted widespread attention from policymakers and scientists. Understanding the distribution patterns of nearshore water quality and tracking its long-term changes is of great significance for water resource management.

[0003] Nitrogen and phosphorus nutrients and chemical oxygen demand (COD) are core indicators for monitoring nearshore marine environments. Excessive discharge of these nutrients has led to eutrophication, resulting in frequent red tides and severe economic losses. Traditional water quality monitoring methods primarily include ship-based monitoring and automatic buoy monitoring. However, these methods are limited by cost and sea conditions, and the spatiotemporal synchronicity and coverage of the data obtained are also significantly insufficient. Existing water quality monitoring methods clearly cannot meet the needs for rapid monitoring, comprehensive coverage, and scientific management of nearshore marine pollution.

[0004] On the other hand, satellite remote sensing has the advantages of large-scale, long-term, retrospective, and quasi-synchronous observation, providing abundant data sources for monitoring marine water quality and eutrophication. MODIS imagery has advantages such as short revisit periods, high spectral resolution, and global open access, and has become widely used data in marine ecological environment monitoring research. Summary of the Invention

[0005] To address the problems described in the prior art, this invention provides a remote sensing inversion method and system for nearshore nutrients and chemical oxygen demand (COD). By using machine learning algorithms to remotely monitor nitrogen and phosphorus nutrients and COD in nearshore waters, the operational level of intelligent monitoring of the nearshore marine environment on a large scale is improved.

[0006] The method of this invention is achieved through the following technical solution: a remote sensing inversion method for nitrogen and phosphorus nutrients and chemical oxygen demand in nearshore waters, comprising the following steps:

[0007] S1. Data collection and processing: acquiring remote sensing data, reanalysis data, and in-situ water quality monitoring data;

[0008] S2. Spatiotemporal matching of remote sensing data, reanalysis data and in-situ water quality monitoring data is performed to construct a space-ground spatiotemporal dataset;

[0009] S3. Add spatiotemporal information to the space-ground spatiotemporal dataset, expand the alternative features through feature engineering, and construct a space-ground spatiotemporal big data dataset that integrates spatiotemporal features and ocean physicochemical properties.

[0010] S4. Select the XGBoost machine learning ensemble algorithm based on CART decision tree as the underlying algorithm, construct the basic XGBoost model, and determine the input features of the basic XGBoost model when it is used to predict inorganic nitrogen, reactive phosphate and chemical oxygen demand, so as to form the water quality parameter model.

[0011] S5. Using the features of the space-ground spatiotemporal big data set constructed in step S3, train the basic XGBoost model, retain all features that contribute to the improvement of model accuracy during the training process, and use them as the final input features of each water quality parameter model to construct a new XGBoost model.

[0012] S6. Improve the parameterized loss function and use the improved parameterized loss function to optimize the constructed new XGBoost model;

[0013] S7. The hyperparameters of the newly constructed XGBoost model are tuned to obtain the final inversion model of water quality parameters.

[0014] S8. Using the constructed final inversion model, the complex nonlinear relationship between nitrogen and phosphorus nutrients and chemical oxygen demand in water bodies and satellite data, marine physicochemical data and spatiotemporal information is characterized.

[0015] The system of this invention is achieved through the following technical solution: a remote sensing inversion system for nitrogen and phosphorus nutrients and chemical oxygen demand in nearshore waters, comprising the following modules:

[0016] The data collection and processing module is used to acquire remote sensing data, reanalysis data, and in-situ water quality monitoring data.

[0017] The dataset construction module is used to perform spatiotemporal matching of remote sensing data, reanalysis data and in-situ water quality monitoring data to construct a space-ground spatiotemporal dataset.

[0018] The data fusion module is used to add spatiotemporal information to the satellite-ground spatiotemporal dataset, expand the alternative features through feature engineering, and construct a satellite-ground spatiotemporal big data dataset that integrates spatiotemporal features and ocean physicochemical properties.

[0019] The basic model building module is used to select the XGBoost machine learning ensemble algorithm based on CART decision tree as the underlying algorithm, build the basic XGBoost model, and determine the input features of the basic XGBoost model when it is used to predict inorganic nitrogen, reactive phosphate and chemical oxygen demand, respectively, to form the models of each water quality parameter.

[0020] The feature selection module uses features from the space-ground spatiotemporal big data set constructed in the data fusion module to train the basic XGBoost model, retaining all features that contribute to the improvement of model accuracy during the training process, and using them as the final input features for each water quality parameter model to construct a new XGBoost model.

[0021] The model optimization module improves the parametric loss function and uses the improved parametric loss function to optimize the newly constructed XGBoost model.

[0022] The parameter tuning module tunes the hyperparameters of the newly constructed XGBoost model to obtain the final inversion model of water quality parameters.

[0023] The inversion module uses the constructed final inversion model to characterize the complex nonlinear relationships between nitrogen and phosphorus nutrients and chemical oxygen demand in water bodies and satellite data, marine physicochemical data and spatiotemporal information.

[0024] This invention, based on MODIS multispectral imagery and utilizing machine learning algorithms, develops a long-term, large-scale, near-real-time remote sensing monitoring solution for excessive pollutants in nearshore waters, including inorganic nitrogen (DIN), reactive phosphate (SRP), and chemical oxygen demand (COD). This improves the operational level of intelligent monitoring of the nearshore marine environment on a large scale. Compared with existing technologies, the beneficial effects achieved by this invention include:

[0025] 1. Compared with on-site observation and UAV hyperspectral water quality remote sensing methods, this invention utilizes the MODIS multispectral satellite remote sensing inversion method, which has outstanding advantages in terms of spatiotemporal coverage, observation frequency, and spatiotemporal synchronization of observation data. It can achieve long-term, full spatial coverage, and traceable water quality remote sensing monitoring, and the observation frequency is higher.

[0026] 2. Based on the XGBoost algorithm, this invention employs feature engineering and IncMRE for feature selection, constructing an interpretable artificial intelligence (XAI) model of nitrogen and phosphorus nutrients and organic matter in water bodies based on optical remote sensing. Compared with existing models, the modeling method of this invention is more advanced, enabling it to more quickly and effectively capture and deeply characterize the complex nonlinear relationships between nitrogen and phosphorus nutrients and chemical oxygen demand (COD) in water bodies and satellite data.

[0027] 3. This invention is the first to propose integrating spatiotemporal features and marine physicochemical properties into the model input. This method significantly improves the overall accuracy of the remote sensing inversion model of non-optically active water quality parameters.

[0028] 4. This invention combines the measured distribution range of inorganic nitrogen, reactive phosphate, and chemical oxygen demand data, and improves the loss function of the machine learning algorithm to the Huber Loss parameterized loss function, thereby improving the accuracy and robustness of the model. Attached Figure Description

[0029] Figure 1 This is a flowchart of the remote sensing inversion method for nearshore nutrients and chemical oxygen demand provided in this embodiment of the invention.

[0030] Figure 2 This is a schematic diagram illustrating the importance of each feature in the XGB-DIN model, XGB-SRP model, and XGB-COD model in embodiments of the present invention.

[0031] Figure 3 This is a comparison chart of the measured values ​​and model predictions of inorganic nitrogen in the test set in the embodiments of the present invention.

[0032] Figure 4 This is a comparison chart of the measured values ​​and model predictions of reactive phosphate in the test set in the embodiments of the present invention.

[0033] Figure 5 This is a comparison chart of the measured values ​​and model predictions of chemical oxygen demand in the test set in the embodiments of the present invention.

[0034] Figure 6 This is a spatial distribution map of inorganic nitrogen concentration retrieved from MODIS-Aqua imagery dated November 28, 2021, in an embodiment of the present invention.

[0035] Figure 7 This is a spatial distribution map of reactive phosphate concentration retrieved from MODIS-Aqua imagery dated November 28, 2021, in an embodiment of the present invention.

[0036] Figure 8 This is a spatial distribution map of chemical oxygen demand concentration retrieved from MODIS-Aqua imagery dated November 28, 2021, in an embodiment of the present invention. Detailed Implementation

[0037] To enable those skilled in the art to better understand the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings and implementation cases in the nearshore waters of Guangdong Province. However, the implementation methods of the present invention are not limited thereto. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention are within the scope of protection of the present invention.

[0038] Example 1

[0039] Please see Figure 1 The remote sensing inversion method for nearshore nutrients and chemical oxygen demand provided in this invention, based on MODIS optical sensing and supplemented by HYCOM reanalysis of marine physicochemical properties, remotely monitors inorganic nitrogen, reactive phosphate, and chemical oxygen demand in nearshore waters. Specifically, it includes the following steps:

[0040] S1. Data collection and processing: acquiring remote sensing data, reanalysis data, and in-situ water quality monitoring data.

[0041] In-situ water quality monitoring data were collected from the nearshore waters of Guangdong Province. These data came from the spring, summer, and autumn seawater quality monitoring conducted by the Guangdong Provincial Environmental Monitoring Center in the nearshore waters of Guangdong from 2014 to 2021, and from the ecological environment survey and assessment conducted by the South China Institute of Environmental Sciences of the Ministry of Ecology and Environment in key sea areas of Guangdong Province from 2020 to 2021. A total of 4,039 water samples were collected, including observations of inorganic nitrogen, reactive phosphate, and chemical oxygen demand concentrations.

[0042] Collect remote sensing data. Download the MODISL2 water color band surface optical and intrinsic optical data, which are quasi-synchronous with the water quality sampling time, from the NASA Ocean Color website. The surface optical data includes remotely sensed reflectance (Rrsλ), and the intrinsic optical data includes absorption coefficient (aλ) and backscattering coefficient (bλ) obtained based on the GIOP model framework. b (λ), where λ is 412nm, 443nm, 469nm, 488nm, 531nm, 547nm, 555nm, 645nm, 667nm, and 678nm. Geometric correction was performed on the remote sensing data using IDL 8.5; atmospheric correction was performed on the images, and each image was resampled to a standard grid using ENVI 5.3.

[0043] Sea surface salinity (SSS) and sea surface current field (SSC) data from the HYCOM model reanalysis data that are quasi-synchronous with MODIS were downloaded as auxiliary input features for the model. The spatial resolution is 0.08°. These data were then resampled to the same 1km resolution (approximately 0.005°) as MODIS through bilinear interpolation. Based on the zonal and meridional currents of the sea surface, sea surface velocity (SSV) and sea surface current direction (SSD) data were further calculated as the final reanalysis data.

[0044] Download ETOPO water depth data of the nearshore waters of Guangdong Province and resample it to 1km resolution using bilinear interpolation.

[0045] S2. Spatiotemporal matching of remote sensing data, reanalysis data and in-situ water quality monitoring data is performed to construct a space-ground spatiotemporal dataset.

[0046] Based on latitude and longitude and monitoring time, a time window of ±1 day and a spatial window of 3*3 pixels were adopted. Abnormal pixels (greater than 1.5 times the standard deviation) were removed. The remote sensing data, reanalysis data and in-situ water quality monitoring data collected in step S1 were spatiotemporally matched to construct a satellite-ground spatiotemporal dataset.

[0047] S3. Based on the space-ground spatiotemporal dataset, add spatiotemporal information, expand the alternative features through feature engineering, and construct a space-ground spatiotemporal big data dataset that integrates spatiotemporal features and ocean physicochemical properties.

[0048] Spatiotemporal information (i.e., spatial and temporal information) is added to the constructed satellite-ground spatiotemporal dataset. Spatial information includes longitude, latitude, and local water depth, while temporal information includes month (MOY), day of the week (DOW), Julian day (DOY), cyclic month (cMOY), cyclic week (cDOW), and cyclic day (cDOY). In this embodiment, time parameters MOY, DOW, DOY, cMOY, cDOW, and cDOY are calculated based on the time of image acquisition, and spatial information, including longitude, latitude, and local water depth, is obtained based on the pixel's latitude and longitude.

[0049] Marine physicochemical properties include sea surface salinity, sea surface current field, and their associated sea surface current velocity and direction.

[0050] The model's candidate features were expanded through feature engineering, including mathematical transformations of all data such as reciprocals, logarithms, exponents, square roots, and squares, as well as dual-band and tri-band combinations and band normalization of remote sensing data, to construct a large-scale space-ground spatiotemporal dataset integrating spatiotemporal features and ocean physicochemical properties. The dual-band and tri-band combination forms are shown in Table 1. Rrs(λ), a(λ), and b were respectively... b (λ) Generate band-normalized data nRrs(λ), na(λ), and nb using band normalization. b (λ), the band normalization method is shown in formula (1).

[0051]

[0052] Table 1. Input values ​​for remote sensing reflectance Rrs, absorption coefficient a, and backscattering coefficient b in MODIS remote sensing products in the model input. b The band combination form, where R can represent Rrs, a, b b Rrs, a, b normalized to band b

[0053]

[0054] S4. Select the XGBoost machine learning ensemble algorithm based on CART decision tree as the underlying algorithm, construct the basic XGBoost model, and determine the input features of the basic XGBoost model when it is used to predict inorganic nitrogen, reactive phosphate and chemical oxygen demand, so as to form the water quality parameter model.

[0055] All features from a large dataset integrating spatiotemporal characteristics and marine physicochemical properties were used as inputs for pre-training to construct a basic XGBoost model. This model yielded water quality parameter models for predicting inorganic nitrogen, reactive phosphate, and chemical oxygen demand. The pre-training hyperparameters of the basic XGBoost model are shown in Table 2, and the mean relative error (MRE) of the model was calculated.

[0056] Table 2. Pre-training parameters of the basic XGBoost model (for Sigma, please refer to δ in the loss function in step 3 below).

[0057]

[0058] In this embodiment, the pre-training process of the basic XGBoost model is repeated multiple times (e.g., ten times) to enhance the reliability of the results.

[0059] S5. Using the features of the space-ground spatiotemporal big data set constructed in step S3, train the basic XGBoost model, retain all features that contribute to the improvement of model accuracy during the training process, and use them as the final input features of each water quality parameter model to construct a new XGBoost model.

[0060] Randomly assign values ​​to feature i in the space-ground spatiotemporal big data set constructed in step S3, and construct a test XGBoost model; calculate IncMRE, the change in mean relative error (MRE) of the test XGBoost model relative to the base XGBoost model. i As shown in the following formula (2).

[0061] IncMRE i =MRE i TM -MRE i BM (2)

[0062] Among them, MRE i TM MRE represents the mean relative error of the XGBoost model during testing. i BM This represents the average relative error of the base XGBoost model.

[0063] Repeat the above random assignment operation on all features in the space-ground spatiotemporal big data constructed in step S3 to obtain the IncMRE of all features. i Value. IncMRE i The larger the value, the more important the feature i is for the inversion of water quality parameters.

[0064] IncMRE for all features i The values ​​are arranged in descending order, meaning that the more important a feature is to the model, the higher it appears.

[0065] The features i are input one by one into the constructed base XGBoost model in descending order for training. If the model accuracy improves, the input features are retained; if the model accuracy decreases or remains unchanged, the input features are discarded. This process is repeated until all features have been detected. The final retained features are the input features of the model, as shown in Table 3.

[0066] Table 3. Input features of each selected model based on XGBoost algorithm and IncMRE.

[0067]

[0068] S6. Improve the parameterized loss function and use the improved parameterized loss function to optimize the constructed new XGBoost model.

[0069] Considering the range of measured inorganic nitrogen, reactive phosphate, and chemical oxygen demand data, and to improve the robustness of the model, the loss function of the machine learning algorithm is selected as the Huber Loss parameterized loss function, as shown in the following formula (3), and it is improved as shown in the following formula (4) to make it more suitable for this embodiment.

[0070]

[0071]

[0072] Where y is the true value and f(x) is the predicted value of the model. Formula (4) is differentiable when y≠0. When the prediction deviation is less than δ, the squared error is used; when the prediction deviation is greater than δ, the linear error is used. The improved loss function improves the accuracy of the model.

[0073] S7. The hyperparameters of the newly constructed XGBoost model are tuned to obtain the final inversion model of water quality parameters.

[0074] Based on the final input features of each water quality parameter model determined in step S5, the hyperparameters of the new XGBoost model are optimized using grid search (GridSearchCV). The optimization parameters include n_estimator (number of learners), max_depth (maximum depth of the learner tree), reg_alpha (L1 regularization), reg_lambda (L2 regularization), and colsample_bytree (feature sampling rate), etc. Please refer to Table 4 for details. This ensures that the model has both excellent fitting performance and better generalization performance, thus obtaining the final inversion models of water quality parameters as XGB-DIN, XGB-SRP, and XGB-COD models.

[0075] Table 4. Hyperparameters of the optimal XGBoost model

[0076]

[0077]

[0078] S8. Using the constructed final inversion model, the complex nonlinear relationship between nitrogen and phosphorus nutrients and chemical oxygen demand in water bodies and satellite data, marine physicochemical data and spatiotemporal information is characterized.

[0079] This invention can effectively capture and characterize the complex nonlinear relationships between nitrogen and phosphorus nutrients and chemical oxygen demand in water bodies and satellite data, marine physicochemical data, and spatiotemporal information. This embodiment utilizes SHAP (SHapley Additive exPlanations) values ​​for characterization, such as... Figure 2 As shown, the larger the SHAP value, the more important the corresponding feature is for model inversion.

[0080] Regarding the contribution of remote sensing data to the three models XGB-DIN, XGB-SRP, and XGB-COD, the XGB-DIN model is mainly composed of remote sensing reflectance (R0) in the MODIS blue and green bands. rs The data contribution is specifically for 443nm, 488nm, and 555nm. Figure 2 (Figure (a) in the text). The XGB-SRP model is mainly composed of MODIS visible light band R rs The data contribution, specifically at 412nm, 443nm, 555nm, and 667nm, generates information gain for the model through the ratio of the two bands. Figure 2 (Figure (b) in the text). In the XGB-COD model, besides utilizing MODIS's R... rs In addition to the data, the absorption coefficient (a) and backscattering coefficient (b) were also utilized. bThis involves wavelengths of 412nm, 443nm, 488nm, 555nm, 645nm, 667nm, and 678nm. Figure 2 (Figure (c) in the middle).

[0081] In addition to MODIS remote sensing data, the contributions of ocean physical and chemical properties (salinity, current field) and spatiotemporal information (water depth, longitude, latitude, Julian Day) to the model are also significant.

[0082] In terms of their overall contribution to the model, the main contributing features to the XGB-DIN model are latitude, salinity, longitude, and R. rs (443nm, 488nm, 555nm); the main contributing features to the XGB-SRP model are latitude, Julian day, and R. rs (412nm, 555nm); the main contributing features to the XGB-COD model are salinity, water depth, absorption coefficients (412nm, 488nm, 645nm, 667nm, 678nm), ocean current, and R. rs (555nm and 678nm), longitude, and Julian Day, etc. In this embodiment, the model-related inputs are obtained through feature engineering including mathematical transformations such as reciprocal, logarithm, exponent, square root, and square, as well as combinations of dual-band and three-band, and are subjected to maximum and minimum normalization using the following formula (5). The inputs for XGB-DIN, XGB-SRP, and XGB-COD models are shown in Table 3.

[0083]

[0084] The selected features are all converted into one-dimensional matrices and input into the trained XGB-DIN, XGB-SRP and XGB-COD models respectively. The model loss function is defined as Huber Loss with parameters. The value of parameter Sigma can be found in Table 4. The outputs are the concentrations of inorganic nitrogen, reactive phosphate and chemical oxygen demand respectively.

[0085] In this embodiment, 20% of the dataset is randomly selected as a test set to verify the model's accuracy.

[0086] Comparison of inorganic nitrogen concentration based on field observation and inversion based on XGBoost algorithm, for example Figure 3 As shown, its coefficient of determination is 0.88, the average relative error is 24.39%, and the root mean square error is 0.119 mg / L.

[0087] Comparison of reactive phosphate concentrations based on field observations and XGBoost algorithm inversion: Figure 4As shown, its coefficient of determination is as high as 0.92, the average relative error is 33.27%, and the root mean square error is 0.004 mg / L.

[0088] Comparison of chemical oxygen demand concentration based on field observation and inversion based on XGBoost algorithm, for example Figure 5 As shown, its coefficient of determination is 0.76, the average relative error is 18.58%, and the root mean square error is 0.263 mg / L.

[0089] This embodiment uses MODIS-Aqua satellite remote sensing imagery of the nearshore waters of Guangdong Province dated November 28, 2021. A machine learning model based on the XGBoost algorithm was used to retrieve the concentrations of inorganic nitrogen, reactive phosphate, and chemical oxygen demand, respectively. Their spatial distributions are shown in the figure below. Figure 6 , Figure 7 , Figure 8 .

[0090] This embodiment of the remote sensing inversion method utilizes high-revisit-rate MODIS remote sensing imagery and the XGBoost machine learning algorithm, supplemented by HYCOM model reanalysis of marine physicochemical parameter products. First, feature engineering is used to expand the model's candidate features. Then, IncMRE is used to select model input factors. Finally, grid search is performed to optimize the hyperparameters of the XGBoost model, thereby constructing inversion models suitable for inorganic nitrogen, reactive phosphate, and chemical oxygen demand in nearshore waters. This enables long-term, large-scale, and near-real-time remote sensing monitoring of pollutants exceeding standards in nearshore waters, improving the operational level of large-scale intelligent monitoring of nearshore marine environments. Furthermore, the model input simultaneously integrates spatiotemporal features and marine physicochemical properties, a first in the field, significantly improving the overall accuracy of the model.

[0091] Example 2

[0092] Based on the same inventive concept as Example 1, this example provides a remote sensing inversion system for nitrogen and phosphorus nutrients and chemical oxygen demand in nearshore waters, comprising the following modules:

[0093] The data collection and processing module is used to acquire remote sensing data, reanalysis data, and in-situ water quality monitoring data.

[0094] The dataset construction module is used to perform spatiotemporal matching of remote sensing data, reanalysis data and in-situ water quality monitoring data to construct a space-ground spatiotemporal dataset.

[0095] The data fusion module is used to add spatiotemporal information to the satellite-ground spatiotemporal dataset, expand the alternative features through feature engineering, and construct a satellite-ground spatiotemporal big data dataset that integrates spatiotemporal features and ocean physicochemical properties.

[0096] The basic model building module is used to select the XGBoost machine learning ensemble algorithm based on CART decision tree as the underlying algorithm, build the basic XGBoost model, and determine the input features of the basic XGBoost model when it is used to predict inorganic nitrogen, reactive phosphate and chemical oxygen demand, respectively, to form the models of each water quality parameter.

[0097] The feature selection module uses features from the space-ground spatiotemporal big data set constructed in the data fusion module to train the basic XGBoost model, retaining all features that contribute to the improvement of model accuracy during the training process, and using them as the final input features for each water quality parameter model to construct a new XGBoost model.

[0098] The model optimization module improves the parametric loss function and uses the improved parametric loss function to optimize the newly constructed XGBoost model.

[0099] The parameter tuning module tunes the hyperparameters of the newly constructed XGBoost model to obtain the final inversion model of water quality parameters.

[0100] The inversion module uses the constructed final inversion model to characterize the complex nonlinear relationships between nitrogen and phosphorus nutrients and chemical oxygen demand in water bodies and satellite data, marine physicochemical data and spatiotemporal information.

[0101] In this embodiment, the data collection and processing module acquires reanalysis data as follows: the sea surface salinity data and sea surface current field data in the MODIS quasi-synchronous HYCOM model reanalysis data are resampled to the same 1km resolution as MODIS through bilinear interpolation, and the sea surface velocity data and sea surface current direction data are further calculated based on the zonal and meridional currents of the sea surface as the final reanalysis data.

[0102] The improved parametric loss function in the model optimization module is as follows:

[0103]

[0104] Where y is the true value and f(x) is the model's predicted value.

[0105] Each module in this embodiment is used to implement the corresponding steps in embodiment 1. For detailed implementation process, please refer to embodiment 1, which will not be repeated here.

[0106] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A remote sensing inversion method for nearshore nutrients and chemical oxygen demand, characterized in that, Includes the following steps: S1. Data collection and processing: acquiring remote sensing data, reanalysis data, and in-situ water quality monitoring data; S2. Spatiotemporal matching of remote sensing data, reanalysis data and in-situ water quality monitoring data is performed to construct a space-ground spatiotemporal dataset; S3. Add spatiotemporal information to the space-ground spatiotemporal dataset, expand the alternative features through feature engineering, and construct a space-ground spatiotemporal big data dataset that integrates spatiotemporal features and ocean physicochemical properties. S4. Select the XGBoost machine learning ensemble algorithm based on CART decision tree as the underlying algorithm, construct the basic XGBoost model, and determine the input features of the basic XGBoost model when it is used to predict inorganic nitrogen, reactive phosphate and chemical oxygen demand, so as to form the water quality parameter model. S5. Using the features of the space-ground spatiotemporal big data set constructed in step S3, train the basic XGBoost model, retain all features that contribute to the improvement of model accuracy during the training process, and use them as the final input features of each water quality parameter model to construct a new XGBoost model. S6. Improve the parameterized loss function and use the improved parameterized loss function to optimize the constructed new XGBoost model; S7. The hyperparameters of the newly constructed XGBoost model are tuned to obtain the final inversion model of water quality parameters. S8. Using the constructed final inversion model, the complex nonlinear relationship between nitrogen and phosphorus nutrients and chemical oxygen demand in water bodies and satellite data, marine physicochemical data and spatiotemporal information is characterized. The process of obtaining reanalysis data in step S1 is as follows: the sea surface salinity data and sea surface current field data in the MODIS quasi-synchronous HYCOM model reanalysis data are resampled to the same 1km resolution as MODIS through bilinear interpolation, and the sea surface velocity data and sea surface current direction data are further calculated based on the sea surface zonal and meridional currents, which are used as the final reanalysis data. The spatiotemporal information added in step S3 includes spatial information such as longitude, latitude, and local water depth, and temporal information such as month, day of the week, Julian day, cyclic month, cyclic week, and cyclic day. The model's alternative features are expanded through feature engineering, including mathematical transformations of all data such as reciprocal, logarithm, exponent, square root, and square, as well as dual-band combination, three-band combination, and band normalization of remote sensing data. Marine physicochemical properties include sea surface salinity, sea surface current field, and their associated sea surface current velocity and direction.

2. The remote sensing inversion method according to claim 1, characterized in that, The remote sensing data acquired in step S1 consists of apparent optical and intrinsic optical data of the MODIS L2 water color band, which are quasi-synchronous with the water quality sampling time. The apparent optical data includes remote sensing reflectance. The intrinsic optical quantity data includes the absorption coefficient obtained based on the GIOP model framework. and backscattering coefficient .

3. The remote sensing inversion method according to claim 1, characterized in that, Step S5 includes: Randomly assign values ​​to feature i in the space-ground spatiotemporal big data set constructed in step S3, and construct a test XGBoost model; calculate IncMRE, the change in mean relative error (MRE) of the test XGBoost model relative to the base XGBoost model. i ; Repeat the random assignment operation on all features in the space-ground spatiotemporal big data constructed in step S3, and obtain the IncMRE of all features. i value; IncMRE for all features i Values ​​are sorted in descending order; The features i are input one by one into the base XGBoost model in descending order to train it; if the model accuracy improves, the input features are retained, and if the model accuracy decreases or remains unchanged, the input features are discarded.

4. The remote sensing inversion method according to claim 1, characterized in that, The improved parameterized loss function in step S6 is: ； in, f(x) represents the true value, and f(x) represents the predicted value of the model.

5. The remote sensing inversion method according to claim 1, characterized in that, Step S8 uses the SHAP value to characterize the complex nonlinear relationship. The larger the SHAP value, the more important the corresponding feature is to the inversion of the model.

6. A remote sensing inversion system for nearshore nutrients and chemical oxygen demand, characterized in that, Includes the following modules: The data collection and processing module is used to acquire remote sensing data, reanalysis data, and in-situ water quality monitoring data. The dataset construction module is used to perform spatiotemporal matching of remote sensing data, reanalysis data and in-situ water quality monitoring data to construct a space-ground spatiotemporal dataset. The data fusion module is used to add spatiotemporal information to the satellite-ground spatiotemporal dataset, expand the alternative features through feature engineering, and construct a satellite-ground spatiotemporal big data dataset that integrates spatiotemporal features and ocean physicochemical properties. The basic model building module is used to select the XGBoost machine learning ensemble algorithm based on CART decision tree as the underlying algorithm, build the basic XGBoost model, and determine the input features of the basic XGBoost model when it is used to predict inorganic nitrogen, reactive phosphate and chemical oxygen demand, respectively, to form the models of each water quality parameter. The feature selection module uses features from the space-ground spatiotemporal big data set constructed in the data fusion module to train the basic XGBoost model, retaining all features that contribute to the improvement of model accuracy during the training process, and using them as the final input features for each water quality parameter model to construct a new XGBoost model. The model optimization module improves the parametric loss function and uses the improved parametric loss function to optimize the newly constructed XGBoost model. The parameter tuning module tunes the hyperparameters of the newly constructed XGBoost model to obtain the final inversion model of water quality parameters. The inversion module uses the constructed final inversion model to characterize the complex nonlinear relationships between nitrogen and phosphorus nutrients and chemical oxygen demand in water bodies and satellite data, marine physicochemical data and spatiotemporal information. The process of acquiring reanalysis data is as follows: the sea surface salinity data and sea surface current field data in the MODIS quasi-synchronous HYCOM model reanalysis data are resampled to the same 1km resolution as MODIS through bilinear interpolation, and the sea surface current velocity data and sea surface current direction data are further calculated based on the sea surface zonal and meridional currents, which are used as the final reanalysis data. The added spatiotemporal information includes spatial information such as longitude, latitude, and local water depth, and temporal information such as month, day of the week, Julian day, cyclic month, cyclic week, and cyclic day. The model's alternative features are expanded through feature engineering, including mathematical transformations of all data such as reciprocal, logarithm, exponential, square root, and square, as well as dual-band combination, three-band combination, and band normalization of remote sensing data. Marine physicochemical properties include sea surface salinity, sea surface current field, and their associated sea surface current velocity and direction.

7. The remote sensing inversion system according to claim 6, characterized in that, The data collection and processing module acquires reanalysis data as follows: the sea surface salinity data and sea surface current field data in the MODIS quasi-synchronous HYCOM model reanalysis data are resampled to the same 1km resolution as MODIS through bilinear interpolation. Based on the zonal and meridional currents of the sea surface, the sea surface velocity data and sea surface current direction data are further calculated as the final reanalysis data.

8. The remote sensing inversion system according to claim 6, characterized in that, The improved parametric loss function in the model optimization module is as follows: ； in, f(x) represents the true value, and f(x) represents the predicted value of the model.

Images