Offshore water quality remote sensing inversion and classification method based on small sample deep learning

Abstract

The invention relates to the technical field of ocean remote sensing, in particular to an offshore water quality remote sensing inversion and classification method based on small sample deep learning, which comprises the following steps: acquiring offshore in-situ observation, satellite remote sensing and ocean reanalysis data; carrying out atmospheric correction, carrying out space-time reconstruction on the missing remote sensing reflectivity by adopting a data interpolation empirical orthogonal function, carrying out mathematical transformation on input features, and screening out an optimal feature subset by adopting a strategy based on an average absolute percentage error; constructing a deep learning network based on a generative small sample to invert the offshore soluble inorganic nitrogen and inorganic phosphorus concentration, and determining the water quality grade; and analyzing an inversion mechanism by using an SHAP method. According to the method, the problem of data space-time discontinuity is solved through the data interpolation empirical orthogonal function, the capture capability and inversion precision of the nonlinear relation under the small sample condition are remarkably improved by utilizing the generative network, the model interpretability is realized in combination with SHAP analysis, and scientific support is provided for offshore water quality fine management.

Description

Technical Field

[0001] This invention relates to the field of marine remote sensing technology, specifically to a method for nearshore water quality remote sensing inversion and classification based on small-sample deep learning. Background Technology

[0002] With the acceleration of industrialization and continuous urbanization, human activities have significantly increased the transport flux of nutrients such as nitrogen and phosphorus into the ocean, leading to increasingly severe water pollution problems in coastal waters. Eutrophication-induced ecological disasters such as algal blooms and red tides not only cause global economic losses but also seriously threaten human health. Nitrogen and phosphorus nutrients, as key nutrients in coastal waters, are important indicators for measuring the degree of water pollution. Therefore, achieving accurate monitoring of the spatiotemporal distribution and variation patterns of nitrogen and phosphorus nutrients is of significant practical importance for water quality assessment, environmental management decisions, and pollution prevention and control.

[0003] Currently, the main methods for obtaining seawater quality parameters include in-situ monitoring and satellite remote sensing. However, existing monitoring methods and inversion models still have many shortcomings in practical applications. Existing in-situ monitoring methods are costly, have sparse stations, and are difficult to capture the significant spatial heterogeneity of seawater, failing to meet the needs of large-scale, high spatiotemporal resolution surveys. While satellite remote sensing technology can achieve wide-area observation, there is a highly nonlinear relationship between non-optically active parameters such as nitrogen and phosphorus and remote sensing data, which traditional empirical models cannot effectively fit, resulting in limited inversion accuracy. Furthermore, existing deep learning models are mostly designed for inland waters and have poor adaptability when directly transferred to the more complex open sea environment. Due to the above-mentioned insufficient inversion accuracy, there is currently a lack of automated, high-precision marine water quality classification methods based on remote sensing data, making it difficult to effectively support the refined management of the marine environment. To address this, a method for nearshore water quality remote sensing inversion and classification based on small-sample deep learning is proposed. Summary of the Invention

[0004] The purpose of this invention is to provide a method for remote sensing inversion and classification of nearshore water quality based on few-sample deep learning, which uses few-sample deep learning to perform remote sensing inversion and classification of nearshore water quality.

[0005] To achieve the above objectives, the present invention provides the following technical solution: Nearshore water quality remote sensing inversion and classification methods based on few-sample deep learning include: Acquire nearshore in-situ water quality observation data, satellite remote sensing data, and ocean reanalysis data; perform atmospheric correction on the satellite remote sensing data to generate remote sensing reflectance; use data interpolation empirical orthogonal functions to spatiotemporally reconstruct and fill in the missing remote sensing reflectance; resample the processed remote sensing reflectance and ocean reanalysis data; and complete spatiotemporal matching with the in-situ water quality observation data to generate input features. The input features are mathematically transformed, and effective variables are selected using a feature selection strategy based on the mean absolute percentage error to determine the optimal subset of input features for inversion. Based on the optimal subset of input features, a few-sample deep learning network based on generative Transformer is constructed. By capturing the nonlinear dependencies of the input features, the concentrations of dissolved inorganic nitrogen and dissolved inorganic phosphorus in the nearshore waters are output in a generative prediction manner. Based on the concentrations of dissolved inorganic nitrogen and dissolved inorganic phosphorus in the nearshore waters, the water quality classification level of the study area is determined by referring to the seawater quality standards. The SHAP method is used to quantify the contribution of each input feature to the output of a few-shot deep learning network based on generative Transformer, and the inversion mechanism is analyzed.

[0006] Preferably, the specific process for obtaining remote sensing reflectance is as follows: Calculate the radiance reflected by water surface foam and the radiance of Rayleigh scattering, and perform ozone absorption correction on the total radiance received from the satellite remote sensing data; based on the assumption that the water-free radiance in the short-wave infrared band is zero, calculate the ratio of atmospheric path radiation to Rayleigh scattering radiance in the short-wave infrared band; determine the aerosol mode and its weight using an aerosol scattering lookup table; calculate the aerosol optical thickness, atmospheric path radiation radiance, and atmospheric diffuse transmittance in the visible light band using the aerosol mode weight interpolation; subtract the atmospheric path radiation radiance and the water surface foam reflection radiance from the total radiance after ozone absorption correction, divide the result by the atmospheric diffuse transmittance to obtain the water-free radiance, and obtain the remote sensing reflectance after water-free radiance bidirectional correction.

[0007] Preferably, the specific process of using data interpolation empirical orthogonal functions to perform spatiotemporal reconstruction and filling of missing data is as follows: The remote sensing reflectance is preprocessed into a matrix consisting of spatial units and time series. Pure ocean pixels are extracted using a land mask, the overall mean is subtracted, and the initial value of all missing pixels is set to zero. Empirical orthogonal function analysis is performed on the preprocessed matrix. The initial predicted value of the missing pixel is calculated using the spatial and temporal functions of the first-order mode. The initial value of the missing pixel is then replaced to complete the initial filling. The empirical orthogonal function modes are iteratively updated. The values ​​of missing pixels are updated using the matrix values ​​reconstructed in each iteration. The optimal number of modes to be retained is determined by cross-validation until the iteration results converge. The matrix after filling in the missing data is then restored to the original data format.

[0008] Preferably, the mathematical transformation includes: calculating the natural logarithm of the input feature and the negative value of the square of the input feature to expand the nonlinear expression of the data distribution; introducing spatial geographic attributes to obtain the longitude and latitude information corresponding to the input feature, calculating the product of the input feature with the cosine value of the longitude and the product of the input feature with the cosine value of the latitude respectively, and using the calculation results as enhanced features that fuse spatiotemporal information.

[0009] Preferably, the feature selection strategy based on mean absolute percentage error is to calculate the basic mean absolute percentage error based on the input features after the mathematical transformation. The validity of each input feature is verified separately, including keeping other input features unchanged, assigning random values ​​to the current input feature to be tested, generating updated input features, and calculating the updated mean absolute percentage error based on the updated input features. The updated mean absolute percentage error is compared with the base mean absolute percentage error. If the updated mean absolute percentage error is greater than the base mean absolute percentage error, the input feature is determined to be a valid variable, and all variables determined to be valid are combined into the optimal subset of input features.

[0010] Preferably, the core structure of the generative Transformer-based few-shot deep learning network adopts a context learning architecture, which normalizes the input features and combines them with random feature embedding to generate an initial embedding vector containing context information; it constructs a dual attention mechanism, including inter-feature attention and inter-sample attention, wherein the inter-feature attention is used to capture the dependencies between different features within a single sample, and the inter-sample attention is used to learn the association between the same features across samples; the outputs of the two attention mechanisms are processed by nonlinear transformation, residual connection and layer normalization to form the final feature representation; A generative prediction approach is adopted, which uses a piecewise constant distribution to model the regression task, outputs the probability weights of the piecewise intervals, and calculates the mean of the probability distribution to generate the concentrations of dissolved inorganic nitrogen and dissolved inorganic phosphorus in the nearshore waters.

[0011] Preferably, the step of using the SHAP method to quantify the contribution of each input feature to the output of the generative Transformer-based few-shot deep learning network involves constructing a simplified input space and using binary vectors to represent the state of the input features in the feature subset. To calculate the marginal contribution, for each target input feature to be analyzed, the difference between the feature subset containing the input feature and the remaining feature subset after removing the target input feature in the model prediction value is calculated as the marginal contribution of the target input feature. The final contribution value is obtained, and the marginal contribution is weighted and summed according to the size of the feature subset to obtain the SHAP value of the input feature. The magnitude and sign of the SHAP value are used to characterize the intensity and direction of the influence of the target input feature on the concentration of dissolved inorganic nitrogen and dissolved inorganic phosphorus in the nearshore waters.

[0012] Preferably, the analytical inversion mechanism specifically involves: taking the absolute value of the calculated SHAP value of each input feature and calculating the average value, using the result as a global importance index to measure the contribution of each input feature; sorting all input features in descending order based on the global importance index to screen out the dominant factors affecting concentration changes; and analyzing the positive and negative correlations between the dominant factors and the concentrations of dissolved inorganic nitrogen and dissolved inorganic phosphorus in nearshore waters, based on the positive and negative distribution characteristics of the SHAP values ​​corresponding to the dominant factors.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention addresses the nonlinear relationship between non-optically active parameters such as nitrogen and phosphorus in nearshore waters and remote sensing data. It constructs a small-sample deep learning network with a generative Transformer as its core. By capturing the complex dependencies of input features, it effectively overcomes the problems of poor fitting ability of traditional empirical models and weak adaptability of conventional deep learning models in open sea areas, achieving high-precision concentration inversion. At the same time, to address the data loss problem caused by cloud cover, it uses data interpolation empirical orthogonal functions to spatiotemporally reconstruct and fill in the gaps in remote sensing reflectance, ensuring the spatiotemporal continuity of the input dataset and further guaranteeing the stability and predictive reliability of the model in complex nearshore environments.

[0014] 2. This invention introduces the SHAP method to quantify the contribution intensity and direction of input features such as mixed layer depth and sea surface temperature to the model output results, effectively solving the problem of the black box nature of deep learning models and providing a scientific basis for analyzing the spatiotemporal variation mechanism of nearshore nutrient concentration. By combining feature engineering with mathematical transformations and feature selection strategies based on mean absolute percentage error, this invention can select the most effective subset of feature variables from multi-source data and eliminate redundant information, thereby significantly improving the model's inversion performance while reducing computational complexity.

[0015] 3. Based on the high-precision inversion results of dissolved inorganic nitrogen and dissolved inorganic phosphorus, this invention can directly determine the water quality level of the study area, realizing full automation from parameter inversion to water quality classification. Compared with traditional costly and limited-coverage in-situ monitoring methods, this invention can capture the spatiotemporal heterogeneity of water quality on a large scale, providing near real-time water quality classification results, and providing comprehensive and timely decision support for the formulation of pollution prevention and control measures and water quality assessment. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a method for nearshore water quality remote sensing inversion and classification based on few-sample deep learning. Figure 2 This is a schematic diagram of the architecture of a nearshore water quality remote sensing inversion and classification method based on few-sample deep learning; Figure 3 This is a coastal water quality classification map based on a nearshore water quality remote sensing inversion and classification method using small-sample deep learning. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 and Figure 2 This invention provides a method for nearshore water quality remote sensing inversion and classification based on few-sample deep learning, and the technical solution is as follows: Nearshore water quality remote sensing inversion and classification methods based on few-sample deep learning include: Acquire nearshore in-situ water quality observation data, satellite remote sensing data, and ocean reanalysis data; perform atmospheric correction on the satellite remote sensing data to generate remote sensing reflectance; use data interpolation empirical orthogonal functions to spatiotemporally reconstruct and fill in the missing remote sensing reflectance; resample the processed remote sensing reflectance and ocean reanalysis data; and complete spatiotemporal matching with the in-situ water quality observation data to generate input features. The input features are mathematically transformed, and effective variables are selected using a feature selection strategy based on the mean absolute percentage error to determine the optimal subset of input features for inversion. Based on the optimal subset of input features, a few-sample deep learning network based on generative Transformer is constructed. By capturing the nonlinear dependencies of the input features, the concentrations of dissolved inorganic nitrogen and dissolved inorganic phosphorus in the nearshore waters are output in a generative prediction manner. Based on the concentrations of dissolved inorganic nitrogen and dissolved inorganic phosphorus in the nearshore waters, the water quality classification level of the study area is determined by referring to the seawater quality standards. The SHAP method is used to quantify the contribution of each input feature to the output of a few-shot deep learning network based on generative Transformer, and the inversion mechanism is analyzed.

[0019] The specific process for obtaining remote sensing reflectance is as follows: Calculate the radiance reflected by water foam and the radiance of Rayleigh scattering, and perform ozone absorption correction on the total radiance received from the satellite remote sensing data; based on the assumption that the water-free radiance in the short-wave infrared band is zero, calculate the ratio of atmospheric path radiation to Rayleigh scattering radiance in the short-wave infrared band; determine the aerosol mode and its weight using an aerosol scattering lookup table; calculate the aerosol optical thickness, atmospheric path radiation radiance, and atmospheric diffuse transmittance in the visible light band using the aerosol mode weight interpolation; subtract the atmospheric path radiation radiance and the water foam reflection radiance from the total radiance after ozone absorption correction, divide the result by the atmospheric diffuse transmittance to obtain the water-free radiance, and obtain the remote sensing reflectance after water-free radiance bidirectional correction; Specifically, the in-situ water quality observation data used in this embodiment comes from a comprehensive spatiotemporal dataset covering China for forty years, mainly using the monthly water quality records for the whole country from 2017 to 2022 provided by the National Marine Environmental Monitoring Center; the satellite remote sensing data used is MODIS-Aqua 1B level data, specifically including raw data in the 412, 443, 488, 555, and 667 nm bands; the ocean reanalysis data used comes from the Copernicus Ocean Service data repository, specifically including sea surface salinity, sea surface height, sea surface temperature, sea surface current velocity, and mixed layer depth. The basic components required for atmospheric correction are calculated based on the satellite remote sensing data, including the radiance of water surface foam reflection and Rayleigh scattering radiance; The water surface foam reflectance is obtained by multiplying the following parameters by pi: water absorption correction coefficient, foam reflectance, solar irradiance at the top of the atmosphere, solar zenith angle, atmospheric diffuse transmittance from the water surface to the remote sensor, and atmospheric diffuse transmittance from the top of the atmosphere to the water surface. The foam reflectance is determined based on the wind speed at a height of 10 meters above the water surface; the two have a power function relationship, i.e., foam reflectance equals a constant (0.00000026) multiplied by the wind speed to the power of 3.52. The atmospheric diffuse transmittance is obtained by using a natural constant as the base and an exponent containing the tangent, sine, and cosine of the solar zenith angle. Approximate formula calculation: To calculate Rayleigh scattering radiance, an accurate Rayleigh scattering lookup table generated by the ocean-atmosphere coupled vector radiative transfer numerical calculation model is used. The solar zenith angle of the current pixel, the remote sensor observation zenith angle, and the relative azimuth angle between the sun and the remote sensor observation are used as inputs for interpolation retrieval, and correction is made according to the actual air pressure. On this basis, ozone two-way absorption correction is performed. The influence of water surface foam reflection radiance and transmittance is subtracted from the total radiance received by the remote sensor, and then multiplied by an exponential attenuation term determined by ozone optical thickness, solar zenith angle, and remote sensor observation zenith angle to obtain the total radiance after ozone correction. Aerosol model extrapolation based on the shortwave infrared band and visible light band was conducted, with atmospheric correction. Based on the black water assumption of zero water-leaving radiance in the shortwave infrared band, the ratio of atmospheric path radiance to Rayleigh scattering radiance in the shortwave infrared band was calculated. Using an aerosol scattering lookup table, a quadratic polynomial relationship between this ratio and aerosol optical thickness was established, thereby calculating the aerosol optical thickness in the shortwave infrared band. By comparing the measured ratio with the theoretical ratios of 20 preset aerosol models in the lookup table, the two closest aerosol models were selected, and calculations were performed. The weights of the first mode reflect the degree of bias in the actual aerosol properties between the two modes. Using these weights and the parameters of the two modes, the aerosol optical thickness in the visible light band and the ratio of atmospheric path radiation to Rayleigh scattering radiance are interpolated and calculated, thereby obtaining the atmospheric path radiation radiance in the visible light band. At the same time, using the atmospheric diffuse transmittance lookup table and the calculated aerosol optical thickness, the uplink and downlink atmospheric diffuse transmittance corresponding to the two modes are obtained respectively, and then weighted and summed using the same weights to obtain the final atmospheric diffuse transmittance. The water-leaving radiance is calculated using the ozone-corrected total radiance, atmospheric path radiance, and atmospheric diffuse transmittance, and then converted into remote sensing reflectance. The formula for calculating the water-leaving radiance is as follows:

[0020] Further obtain the normalized water ionization radiance:

[0021] in, For water radiance, The total radiance is after correction for ozone and white foam. This refers to the atmospheric path radiance in the visible light band. The atmospheric diffuse transmittance from the sea surface to the remote sensor. The zenith angle of the sun. The instantaneous Earth-Sun distance The mean Earth-Sun distance is given. Based on this, a bidirectional correction for water-leaving radiation is performed to obtain the corrected value. The remote sensing reflectance is obtained by dividing by the solar irradiance at the top of the atmosphere at the average Earth-Sun distance. By employing refined physical radiative transfer correction and a dual aerosol mode weighted interpolation strategy based on the shortwave infrared band, the problem of aerosol extrapolation in complex nearshore atmospheric environments was effectively solved, significantly improving the accuracy of remote sensing reflectance acquisition. This provided high signal-to-noise ratio input features for subsequent deep learning models and effectively ensured the accuracy and robustness of non-optically active water quality parameters (nitrogen and phosphorus) inversion.

[0022] The specific process of using empirical orthogonal functions for data interpolation to perform spatiotemporal reconstruction and filling of missing data is as follows: The remote sensing reflectance is preprocessed into a matrix consisting of spatial units and time series. Pure ocean pixels are extracted using a land mask, the overall mean is subtracted, and the initial value of all missing pixels is set to zero. Empirical orthogonal function analysis is performed on the preprocessed matrix. The initial predicted value of the missing pixel is calculated using the spatial and temporal functions of the first-order mode. The initial value of the missing pixel is then replaced to complete the initial filling. The empirical orthogonal function modes are iteratively updated. The values ​​of missing pixels are updated using the matrix values ​​reconstructed in each iteration. The optimal number of modes to be retained is determined by cross-validation until the iteration results converge. The matrix after filling in the missing data is then restored to the original data format.

[0023] Specifically, preprocessing the remote sensing reflectance into a matrix composed of spatial units and time series involves defining the dimensions of the matrix and reshaping the satellite remote sensing reflectance data in the time series into... A two-dimensional matrix, where The number of spatial units (i.e., the total number of valid ocean pixels after removing land). This represents the number of data files in the time series. During this process, a land mask is applied to extract pure ocean pixels to eliminate land signal interference; the overall mean of all valid observations in the matrix is ​​calculated, and each value in the matrix is ​​subtracted from this overall mean to eliminate baseline bias. Finally, for missing pixels in the matrix caused by cloud cover, atmospheric correction failure, etc., their initial values ​​are uniformly set to zero, serving as the starting input for subsequent iterative calculations. The step of performing empirical orthogonal function analysis on the preprocessed matrix and calculating the initial predicted values ​​of missing pixels using the spatial and temporal functions of the first-order mode specifically involves performing singular value decomposition on the preprocessed and zero-padded matrix, decomposing it into a product of the spatial EOF matrix, the singular value diagonal matrix, and the adjoint matrix of the temporal EOF matrix. In the first round of calculation, only the first-order EOF mode containing the maximum variation information is extracted. The initial predicted values ​​of the missing positions are calculated using the product of the spatial feature vector (spatial function), singular values, and temporal feature vector (temporal function) corresponding to this mode. These calculated initial predicted values ​​are used to replace the missing pixel values ​​set to zero in the original matrix, while the existing valid observation values ​​remain unchanged, thus completing the initial data filling. The initially filled complete matrix is ​​used as new input for empirical orthogonal function analysis. In each subsequent iteration, the currently selected number of EOF modes (i.e., the previous...) are used... The principal components are used to reconstruct the matrix, and the values ​​at the locations of the missing data filled in the previous round are updated using the values ​​in the reconstructed matrix. This is to determine the optimal number of modes to retain. The cross-validation strategy is adopted: a portion of the original valid data is randomly selected as the validation set and treated as missing data to participate in the imputation process. The root mean square error between the imputed values ​​of these validation points and the true observations is calculated under different modal numbers. The modal number that minimizes the validation error is selected as the optimal modal number for the final iterative calculation. The above iterative update steps are repeated until the difference between the imputation results of two adjacent iterations is less than a preset threshold, which means that the result has converged. Once the iteration process meets the convergence condition, the calculation stops. Each value in the final filled matrix is ​​then added to the overall mean subtracted during the preprocessing stage to restore the original magnitude of the data. Finally, the two-dimensional matrix is ​​reverse-engineered according to its original longitude, latitude, and time dimensions to restore it to the original satellite remote sensing image format. The DINEOF method is used to reconstruct missing remote sensing data. Its significant advantage lies in the fact that it does not require the introduction of external prior data such as sea surface temperature or wind field; it can achieve high-precision missing value filling solely based on the spatiotemporal correlation of the remote sensing data itself. This method automatically determines the optimal number of modes through iterative optimization and cross-validation, preserving the nonlinear spatiotemporal variation trend of water body optical characteristics to the maximum extent, and effectively solving the data discontinuity problem caused by cloud obstruction in nearshore areas.

[0024] The mathematical transformation includes: calculating the natural logarithm of the input feature and the negative value of the square of the input feature to expand the nonlinear expression of the data distribution; introducing spatial geographic attributes to obtain the longitude and latitude information corresponding to the input feature, calculating the product of the input feature with the cosine value of the longitude and the product of the input feature with the cosine value of the latitude respectively, and using the calculation results as enhanced features that fuse spatiotemporal information; The calculation of the natural logarithm of the input feature and the negative of the square of the input feature is as follows: For each original input feature variable generated through spatiotemporal matching, its natural logarithm is first calculated as a first-type transformation feature to adjust the data distribution and mitigate the impact of numerical magnitude differences on model training; simultaneously, the square of the input feature variable is calculated and negatively taken as a second-type transformation feature. This step, by introducing logarithmic and higher-order terms, explicitly expands the nonlinear dimension of the feature, helping deep learning networks better capture the complex nonlinear response relationship between input variables and water quality parameters. Obtain the geographic longitude and latitude values ​​corresponding to each input feature data point. To couple spatial location with physical parameters, first calculate the cosine value of the trigonometric function of the longitude of the input feature variable, and multiply the longitude cosine value with the input feature variable to obtain the third type of transformation feature; similarly, calculate the cosine value of the trigonometric function of the latitude of the input feature variable, and multiply the latitude cosine value with the input feature variable to obtain the fourth type of transformation feature; For each original input feature variable, four corresponding derived feature variables are generated following the process described above. Then, a feature dimension concatenation operation is performed, combining all derived feature variables generated for all original input features with the original input feature variables themselves to construct an expanded feature matrix. In this matrix, each row represents a spatiotemporal observation sample, and each column represents an original feature or its corresponding derived feature, thereby expanding the original feature space into a high-dimensional feature space containing all candidate variables, constructing a complete candidate feature set including the original features and all derived transformation features. By employing natural logarithmic, negative squared, and latitude-longitude cosine weighted transformations, the nonlinear dimension of the features is explicitly extended and deeply coupled with spatial geographic information. This not only enhances the model's ability to capture nonlinear relationships of non-optically active water quality parameters but also enables it to automatically perceive the spatial heterogeneity between nearshore and offshore areas, thereby significantly improving the generalization performance and prediction accuracy of the water quality inversion model in complex marine environments.

[0025] The feature selection strategy based on mean absolute percentage error is to calculate the basic mean absolute percentage error based on the input features after mathematical transformation. The validity of each input feature is verified separately, including keeping other input features unchanged, assigning random values ​​to the current input feature to be tested, generating updated input features, and calculating the updated mean absolute percentage error based on the updated input features. The updated mean absolute percentage error is compared with the base mean absolute percentage error. If the updated mean absolute percentage error is greater than the base mean absolute percentage error, the input feature is determined to be a valid variable, and all variables determined to be valid are combined into the optimal subset of input features.

[0026] Specifically, a basic deep learning model is constructed based on the complete candidate feature set. The basic deep learning model is used to predict the validation set data, and the mean absolute percentage error between the prediction result and the in-situ measured data is calculated. This is defined as the basic mean absolute percentage error. The method for calculating the mean absolute percentage error is as follows: calculate the absolute value of the difference between the predicted value and the measured value for each sample point, divide the absolute value by the measured value to obtain the relative error, and finally calculate the average value of the relative errors of all samples and convert it into a percentage form. The validity verification of each input feature includes keeping other input features unchanged, assigning random values ​​to the current input feature to be tested, generating updated input features, and calculating the updated mean absolute percentage error based on the updated input features. Specifically, this involves performing sensitivity tests on each feature variable in the candidate feature set sequentially. When testing a specific variable, all other column values ​​in the feature matrix are kept unchanged, while random values ​​are assigned to the current input feature to be tested (in this embodiment, the original data order of the feature column is shuffled), thereby destroying the physical or spatial information carried by the feature and generating an updated input feature matrix. This perturbated updated feature matrix is ​​used for prediction, and the mean absolute percentage error is recalculated according to the aforementioned method, and defined as the updated mean absolute percentage error. The process of comparing the updated mean absolute percentage error (MAAb) with the base MAb is as follows: if the updated MAb is greater than the base MAb, the input feature is determined to be a valid variable. The optimal subset of input features is formed by comparing the updated error of each feature with the base error. If randomization of a feature causes an increase in the model's prediction error (i.e., the updated MAb is greater than the base MAb), it indicates that the feature contains information beneficial to the model's prediction, and it is determined to be a valid variable and retained. Conversely, if the error does not increase or even decreases, it indicates that the feature does not contribute to the model or interferes with it, and it is removed. Finally, all the verified and determined valid variables are grouped together to form the optimal subset of input features. By adopting a feature selection strategy based on model performance changes, we can directly quantify and evaluate the actual contribution of each feature to the final prediction accuracy of the model, and objectively eliminate redundant or noisy features that may be generated after mathematical transformation. This not only significantly reduces the dimensionality of the input data, but also reduces the computational burden of model training.

[0027] The core structure of the generative Transformer-based few-shot deep learning network adopts a context learning architecture. It normalizes the input features and combines them with random feature embedding to generate an initial embedding vector containing context information. It constructs a dual attention mechanism, including inter-feature attention and inter-sample attention. The inter-feature attention is used to capture the dependencies between different features within a single sample, and the inter-sample attention is used to learn the associations between the same features across samples. The outputs of the two attention mechanisms are processed by nonlinear transformation, residual connection, and layer normalization to form the final feature representation. A generative prediction approach is adopted, which uses a piecewise constant distribution to model the regression task, outputs the probability weights of the piecewise intervals, and calculates the mean of the probability distribution to generate the concentrations of dissolved inorganic nitrogen and dissolved inorganic phosphorus in the nearshore waters.

[0028] Specifically, based on the optimal input feature subset, preprocessing is performed using the Z-normalization method to calculate the mean and standard deviation of the features. The original feature values ​​are subtracted from the mean and then divided by the standard deviation to eliminate differences in different physical dimensions. Subsequently, the normalized features are linearly encoded, and a learnable linear encoding weight matrix and bias vector are used to map them to a high-dimensional space. Building upon this, to enhance the model's ability to perceive feature location and distribution, random feature embedding is introduced. A fixed random vector is generated for each feature, and a learned projection matrix is ​​used to transform this random vector. The transformation result is then superimposed on the linearly encoded vector to obtain the final initial embedding vector. The construction of the dual attention mechanism includes inter-feature attention and inter-sample attention. The outputs of the two attention mechanisms are processed through nonlinear transformation, residual connection, and layer normalization to form the final feature representation. Specifically, a dual attention layer is designed. First, inter-feature attention calculation is performed, using the query matrix, key matrix, and value matrix to capture the nonlinear dependencies between different physical features within a single sample. Simultaneously, inter-sample attention calculation is performed, using the query matrix of all samples and the key and value matrices of the training samples to learn the association patterns between the same features across samples. The outputs of the two attention mechanisms are concatenated and input into a two-layer multilayer perceptron containing a nonlinear activation function (ReLU in this embodiment) for transformation. A residual connection mechanism is introduced to add the output of the multilayer perceptron to the initial input, and half-precision layer normalization is used to standardize the result to form a high-level feature representation containing rich contextual information and nonlinear relationships. The generative prediction method employs a piecewise constant distribution model for regression tasks, outputting probability weights for each segment interval and calculating the mean of the probability distribution to generate the concentrations of dissolved inorganic nitrogen and dissolved inorganic phosphorus in nearshore waters. Specifically, unlike traditional regression models that directly output numerical values, this model uses a piecewise constant distribution to fit the probability density of the target water quality parameters (nitrogen and phosphorus concentrations). The Softmax function generates probability weights for the target values ​​falling within each preset numerical interval, and the expected value of this probability distribution is calculated as the final prediction result. Specifically, the center value of each preset interval (i.e., the average of the upper and lower bounds of the interval) is multiplied by its corresponding prediction probability weight, and the product results of all intervals are summed to obtain the predicted values ​​of dissolved inorganic nitrogen and dissolved inorganic phosphorus concentrations in nearshore waters. The generative Transformer-based few-sample deep learning network employs a dual attention mechanism to simultaneously capture the physical coupling mechanism between features and the spatiotemporal correlation pattern between samples, greatly enhancing the model's feature extraction capability under small-sample conditions. The unique random feature embedding and generative prediction method not only avoids the overfitting risk of traditional regression models on non-Gaussian distributed data, but also significantly improves the robustness and accuracy of non-optically active water quality parameters (nitrogen and phosphorus) inversion through probability-weighted averaging.

[0029] The method of using SHAP to quantify the contribution of each input feature to the output of the generative Transformer-based few-shot deep learning network is used to construct a simplified input space, and to represent the state of the input feature in the feature subset using binary vectors. To calculate the marginal contribution, for each target input feature to be analyzed, the difference between the feature subset containing the input feature and the remaining feature subset after removing the target input feature in the model prediction value is calculated as the marginal contribution of the target input feature. The final contribution value is obtained, and the marginal contribution is weighted and summed according to the size of the feature subset to obtain the SHAP value of the input feature. The magnitude and sign of the SHAP value are used to characterize the intensity and direction of the influence of the target input feature on the concentration of dissolved inorganic nitrogen and dissolved inorganic phosphorus in the nearshore waters.

[0030] The construction of the simplified input space, using binary vectors to represent the state of input features in a feature subset, specifically involves defining a simplified binary input space for the high-dimensional input feature space used for inversion. In this binary input space, binary vectors are generated that correspond one-to-one with the original input features, where each element takes only a value of 0 or 1. A value of 1 indicates that the corresponding original input feature exists in the current feature combination and participates in model prediction; a value of 0 indicates that the corresponding original input feature is missing or masked and does not participate in the current feature combination. In this way, complex continuous numerical inputs are transformed into enumerable combinations of feature states. The calculation of marginal contribution involves, for each target input feature to be analyzed, calculating the difference in model prediction values ​​between the feature subset containing the target feature and the remaining feature subset after removing the target input feature. Specifically, this marginal contribution of the target input feature is calculated by selecting a feature subset containing the target feature, mapping this simplified subset back to the original input space using a mapping function, inputting it into the trained deep learning network to obtain the first prediction value, removing the target feature from the feature subset (i.e., changing the corresponding binary bit from 1 to 0), forming the remaining feature subset, mapping it back to the original space, and inputting it into the network to obtain the second prediction value. Subtracting the second prediction value from the first prediction value yields the marginal contribution of the target feature in the context of the current specific feature subset. The process of obtaining the final contribution value involves weighted summation of the marginal contributions based on the size of the feature subsets to obtain the SHAP value of the input feature. The magnitude and sign of the SHAP value characterize the strength and direction of the target input feature's influence on nearshore dissolved inorganic nitrogen and phosphorus concentrations, respectively. Specifically, since features are interdependent, the marginal contribution of a single subset is insufficient to represent global importance. Therefore, it is necessary to traverse all possible feature subset combinations and calculate the marginal contribution of the target feature under all combinations. These marginal contributions are then weighted and summed. The weights depend on the size of the feature subsets and the number of permutations of the total number of features to ensure that each feature receives a fair contribution. The final weighted sum is the SHAP value of the input feature. The absolute value of the SHAP value directly reflects the strength of the feature's contribution to the model's prediction results (nitrogen and phosphorus concentrations). The sign of the SHAP value reveals the direction of the influence: a positive value indicates that an increase in the feature's value will increase the predicted concentration, while a negative value indicates that an increase in the feature's value will decrease the predicted concentration. By introducing the SHAP interpretability analysis method, the mechanism of the nonlinear deep learning inversion model is analyzed. By decomposing the complex inversion results into the sum of the independent contributions of various physical features (such as sea surface temperature, remote sensing reflectance, etc.), it is possible not only to quantitatively identify the key driving factors that dominate the changes in nitrogen and phosphorus concentrations, but also to reveal the positive and negative correlations between features and concentrations.

[0031] The analytical inversion mechanism is as follows: the absolute value of the calculated SHAP value of each input feature is taken and the average value is calculated. The result is used as a global importance index to measure the contribution of each input feature. Based on the global importance index, all input features are sorted in descending order to screen out the dominant factors affecting concentration changes. Combining the positive and negative distribution characteristics of the SHAP values ​​corresponding to the dominant factors, the positive and negative correlations between the dominant factors and the concentrations of dissolved inorganic nitrogen and dissolved inorganic phosphorus in the nearshore area are analyzed. The process of taking the absolute value of the SHAP value of each input feature and averaging the results, and using the result as a global importance index to measure the contribution of each input feature, specifically involves: for each input feature, traversing all samples in the validation set, obtaining the SHAP value of the feature at each sample point, taking the absolute value of each SHAP value to eliminate the mutual cancellation effect of positive and negative directions, and then calculating the arithmetic mean of these absolute values. The arithmetic mean directly reflects the average marginal contribution of the feature to the model output, and is defined as a global importance index to measure the contribution of the feature. The process of sorting all input features in descending order based on the global importance index to identify the dominant factors affecting concentration changes specifically involves: sorting all input features from largest to smallest according to the global importance index values ​​calculated above. Features with larger values ​​indicate a greater correction to the model's prediction results, i.e., higher importance. Based on the sorting results, the top-ranked features (top ten in this embodiment) are selected as the dominant factors affecting nearshore dissolved inorganic nitrogen and dissolved inorganic phosphorus concentration changes. Analysis of actual inversion results shows that the mixing layer depth, spatially adjusted sea surface temperature, and remote sensing reflectance in specific bands are typically among the top factors and are identified as key dominant variables. The analysis of the positive and negative correlation between the dominant factors and the concentrations of dissolved inorganic nitrogen and dissolved inorganic phosphorus in nearshore waters, based on the positive and negative distribution characteristics of the corresponding SHAP values ​​of the dominant factors, specifically involves: for each selected dominant factor, plotting its SHAP value distribution map; analyzing the distribution of the SHAP value of the feature on the positive and negative axes in the distribution map; if the SHAP value of a feature is mainly distributed in the positive range, it indicates that the presence of the feature tends to increase the predicted concentration; if it is mainly distributed in the negative range, it tends to decrease the predicted concentration; and combining this with the numerical magnitude of the feature itself (in this embodiment, high and low values ​​are distinguished by color), refer to... Figure 3 To further analyze its physical mechanism.

[0032] To address the challenge of retrieving non-optically active water quality parameters in nearshore waters, a comprehensive solution integrating multi-source data and few-sample deep learning is proposed. By employing empirical orthogonal functions for data interpolation to spatiotemporally reconstruct remotely sensed reflectance, the solution effectively addresses data loss caused by cloud cover, ensuring the spatiotemporal continuity of the input data. A feature subset selection strategy based on mean absolute percentage error is used to optimize the feature set. Combined with a few-sample deep learning network based on generative Transformers, this significantly improves the ability to capture nonlinear dependencies and the accuracy of inversion under small-sample conditions. Furthermore, the SHAP method enables interpretability analysis of the model, overcoming the black-box limitations of traditional deep learning models. This provides scientific and precise decision support for water quality classification and management in complex nearshore water environments.

[0033] Example 2: This embodiment applies a nearshore water quality remote sensing inversion and classification method based on small sample deep learning to a summer water eutrophication monitoring scenario in a certain sea area, and performs inversion and water quality classification on the concentration of dissolved inorganic nitrogen and dissolved inorganic phosphorus in the sea area. We acquired MODIS-Aqua satellite 1B level remote sensing data covering the study area during this period, as well as ocean reanalysis data such as sea surface salinity, sea surface temperature, and mixed layer depth provided by the Copernicus Ocean Service Center, and combined them with in-situ water quality observation data from some stations in the area provided by the National Marine Environmental Monitoring Center as true references. A refined atmospheric correction process is implemented for the acquired satellite remote sensing data. This process first calculates the reflectance of water foam based on the 3.52nd power function relationship of sea surface wind speed, then subtracts the reflectance radiance of water foam, and combines a Rayleigh scattering lookup table to subtract the influence of Rayleigh scattering and perform ozone absorption correction. Based on the assumption that the water radiance in the short-wave infrared band is zero, the ratio of atmospheric path radiation to Rayleigh scattering in the short-wave infrared band is calculated. By matching it with the theoretical ratio of 20 preset modes in the aerosol scattering lookup table, the two closest aerosol modes and their weights are determined. Using these weights, the aerosol optical thickness and atmospheric diffuse transmittance in the visible light band are interpolated to accurately remove atmospheric path radiation. After water-leaving radiance bidirectional correction, remote sensing reflectance data that can truly reflect the spectral characteristics of the water body is generated. The aerosol scattering lookup table contains 20 standard aerosol models pre-calculated based on Mie scattering theory. Each model includes optical parameters such as extinction coefficient, single-scattering albedo, and asymmetry factor in the short-wave infrared and visible light bands. During the comparison phase, the relative error between the measured ratio and the theoretical ratio of each aerosol model is calculated, and the aerosol model with the smallest relative error is selected as the optimal model. If the difference between the smallest relative error and the second-best model is less than a preset threshold (5% in this embodiment), both models are selected simultaneously for weight calculation. The weight calculation uses an inverse distance weighted interpolation method to perform a weighted average of the visible light band aerosol optical thickness and atmospheric path radiance of the two models to obtain the final aerosol parameters for that pixel. Due to frequent cloud cover in this sea area, a large amount of data is missing from satellite imagery. This embodiment employs an empirical orthogonal function for data interpolation to reconstruct the generated remote sensing reflectance in a spatiotemporal manner. The remote sensing reflectance dataset containing missing values ​​is constructed as a two-dimensional matrix composed of spatial pixels and time series. After applying a high-precision land mask to remove land interference, the effective ocean pixels are mean-reduced, and missing values ​​at cloud-covered locations are initialized to zero. Singular value decomposition is then performed on the matrix, and the initial predicted values ​​at the missing locations are calculated using the spatial and temporal eigenvectors of the first-order EOF mode. Based on this, an iterative loop is entered, using the top k optimal mode reconstruction matrices to update the missing values ​​each time. A cross-validation strategy—randomly masking some known data and calculating the reconstruction error—is used to dynamically determine the optimal number of modes to retain until the difference between the results of two adjacent iterations meets the convergence threshold. Finally, the reconstructed matrix is ​​restored, obtaining a spatiotemporally continuous and seamless remote sensing reflectance field for this sea area. This field is then resampled to 1km resolution with ocean reanalysis data, achieving spatiotemporal matching with in-situ observation data. Specific mathematical transformations and filtering were performed on the matched input features. On the one hand, the natural logarithm and negative square value of remote sensing reflectance and environmental factors for each band were calculated to enhance the nonlinear expression of the data distribution. On the other hand, spatial geographic attributes were introduced, and the product of each feature value and the cosine of its longitude and latitude was calculated to integrate geospatial information. Subsequently, a strategy based on mean absolute percentage error (MAPE) was used to filter the amplified high-dimensional feature set: keeping other features unchanged, the numerical order of a certain feature was shuffled one by one, and the change of MAPE on the validation set was observed. Only those features that caused a significant increase in model error after shuffling were retained, and the optimal subset of input features, including spatially corrected sea surface temperature, reflectance of specific bands, and their nonlinear transformation terms, was determined. Based on the selected optimal feature subset, a few-sample deep learning network based on generative Transformer was constructed for concentration inversion. This network processes the input data through normalization and random feature embedding, and utilizes an inter-feature attention mechanism to capture the complex nonlinear coupling relationship between sediment, chlorophyll, and nitrogen and phosphorus nutrients in the Yangtze River estuary. Simultaneously, it uses an inter-sample attention mechanism to learn the temporal correlation of water quality changes at different times. Using a generative prediction approach, it models the probability distribution of dissolved inorganic nitrogen and dissolved inorganic phosphorus concentrations in the sea area through piecewise constant distribution, and takes the average as the final inversion result. Based on the inverted concentration data, the water quality in the study area is classified into Class I to Class V. The results show that this area, influenced by runoff input, has high nitrogen and phosphorus concentrations and is classified as Class IV or V (worst quality), while the water quality gradually improves in the offshore area. The SHAP method was used to analyze the mechanism of the inversion model. By constructing a binary simplified feature input space and calculating the marginal contribution, the SHAP values ​​of each feature on the predicted nitrogen and phosphorus values ​​were quantified. Global importance analysis showed that sea surface salinity (SSS) and mixing layer depth are the dominant factors affecting the summer nitrogen and phosphorus distribution in this region. The positive and negative distribution characteristics of SHAP values ​​revealed the physical rationality of the inversion mechanism: the SHAP value of sea surface salinity is significantly negatively correlated with the predicted concentration, accurately reflecting the dilution and mixing process spreading to the open sea; while the contribution of mixing layer depth reflects the inhibitory effect of vertical stratification on the upwelling of bottom nutrients in summer. This analysis not only verified the reliability of the model but also provided direct scientific evidence for the analysis of the causes of summer eutrophication in this region.

[0034] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for nearshore water quality remote sensing inversion and classification based on small-sample deep learning, characterized in that, include: Acquire nearshore in-situ water quality observation data, satellite remote sensing data, and ocean reanalysis data; perform atmospheric correction on the satellite remote sensing data to generate remote sensing reflectance; use data interpolation empirical orthogonal functions to spatiotemporally reconstruct and fill in the missing remote sensing reflectance; resample the processed remote sensing reflectance and ocean reanalysis data; and complete spatiotemporal matching with the in-situ water quality observation data to generate input features. The input features are mathematically transformed, and effective variables are selected using a feature selection strategy based on the mean absolute percentage error to determine the optimal subset of input features for inversion. Based on the optimal subset of input features, a few-sample deep learning network based on generative Transformer is constructed. By capturing the nonlinear dependencies of the input features, the concentrations of dissolved inorganic nitrogen and dissolved inorganic phosphorus in the nearshore waters are output in a generative prediction manner. Based on the concentrations of dissolved inorganic nitrogen and dissolved inorganic phosphorus in the nearshore waters, the water quality classification level of the study area is determined by referring to the seawater quality standards. The SHAP method is used to quantify the contribution of each input feature to the output of a few-shot deep learning network based on generative Transformer, and the inversion mechanism is analyzed.

2. The method for nearshore water quality remote sensing inversion and classification based on small-sample deep learning according to claim 1, characterized in that, The specific process for obtaining remote sensing reflectance is as follows: calculate the reflectance of white foam on the water surface and the Rayleigh scattering radiance, and perform ozone absorption correction on the total radiance of the satellite remote sensing data received; Based on the assumption that the water-free radiance in the short-wave infrared band is zero, the ratio of atmospheric path radiation to Rayleigh scattering radiance in the short-wave infrared band is calculated. The aerosol mode and its weight are determined using an aerosol scattering lookup table. The aerosol optical thickness, atmospheric path radiation radiance, and atmospheric diffuse transmittance in the visible light band are calculated by interpolation using the aerosol mode weights. The atmospheric path radiation radiance and the water surface foam reflection radiance are subtracted from the total radiance after ozone absorption correction. The result is divided by the atmospheric diffuse transmittance to obtain the water-free radiance. After water-free radiation bidirectional correction, the remote sensing reflectance is obtained.

3. The method for nearshore water quality remote sensing inversion and classification based on small-sample deep learning according to claim 1, characterized in that, The specific process of using empirical orthogonal functions for data interpolation to perform spatiotemporal reconstruction and filling of missing data is as follows: The remote sensing reflectance is preprocessed into a matrix consisting of spatial units and time series. Pure ocean pixels are extracted using a land mask, the overall mean is subtracted, and the initial value of all missing pixels is set to zero. Empirical orthogonal function analysis is performed on the preprocessed matrix. The initial predicted value of the missing pixel is calculated using the spatial and temporal functions of the first-order mode. The initial value of the missing pixel is then replaced to complete the initial filling. The empirical orthogonal function modes are iteratively updated. The values ​​of missing pixels are updated using the matrix values ​​reconstructed in each iteration. The optimal number of modes to be retained is determined by cross-validation until the iteration results converge. The matrix after filling in the missing data is then restored to the original data format.

4. The method for nearshore water quality remote sensing inversion and classification based on small-sample deep learning according to claim 1, characterized in that, The mathematical transformation includes: calculating the natural logarithm of the input feature and the negative value of the square of the input feature to expand the nonlinear expression of the data distribution; introducing spatial geographic attributes to obtain the longitude and latitude information corresponding to the input feature, calculating the product of the input feature with the cosine value of its longitude and the product of the input feature with the cosine value of its latitude, and using the calculation results as enhanced features that fuse spatiotemporal information.

5. The method for nearshore water quality remote sensing inversion and classification based on small-sample deep learning according to claim 1, characterized in that... The feature selection strategy based on mean absolute percentage error is to calculate the basic mean absolute percentage error based on the input features after the mathematical transformation. The validity of each input feature is verified separately, including keeping other input features unchanged, assigning random values ​​to the current input feature to be tested, generating updated input features, and calculating the updated mean absolute percentage error based on the updated input features. The updated mean absolute percentage error is compared with the base mean absolute percentage error. If the updated mean absolute percentage error is greater than the base mean absolute percentage error, the input feature is determined to be a valid variable, and all variables determined to be valid are combined into the optimal subset of input features.

6. The method for nearshore water quality remote sensing inversion and classification based on small-sample deep learning according to claim 1, characterized in that, The core structure of the generative Transformer-based few-shot deep learning network adopts a context learning architecture, which normalizes the input features and combines them with random feature embedding to generate an initial embedding vector containing context information. Construct a dual attention mechanism, It includes inter-feature attention and inter-sample attention, wherein the inter-feature attention is used to capture the dependencies between different features within a single sample, and the inter-sample attention is used to learn the association between the same feature across samples. The outputs of the two attention are processed by nonlinear transformation, residual connection and layer normalization to form the final feature representation. A generative prediction approach is adopted, which uses a piecewise constant distribution to model the regression task, outputs the probability weights of the piecewise intervals, and calculates the mean of the probability distribution to generate the concentrations of dissolved inorganic nitrogen and dissolved inorganic phosphorus in the nearshore waters.

7. The method for nearshore water quality remote sensing inversion and classification based on small-sample deep learning according to claim 1, characterized in that, The method of using SHAP to quantify the contribution of each input feature to the output of the generative Transformer-based few-shot deep learning network is used to construct a simplified input space, and to represent the state of the input feature in the feature subset using binary vectors. To calculate the marginal contribution, for each target input feature to be analyzed, the difference between the feature subset containing the input feature and the remaining feature subset after removing the target input feature in the model prediction value is calculated as the marginal contribution of the target input feature. The final contribution value is obtained, and the marginal contribution is weighted and summed according to the size of the feature subset to obtain the SHAP value of the input feature. The magnitude and sign of the SHAP value are used to characterize the intensity and direction of the influence of the target input feature on the concentration of dissolved inorganic nitrogen and dissolved inorganic phosphorus in the nearshore waters.

8. The method for nearshore water quality remote sensing inversion and classification based on small-sample deep learning according to claim 1, characterized in that, The analytical inversion mechanism is as follows: the absolute value of the calculated SHAP value of each input feature is taken and the average value is calculated. The result is used as a global importance index to measure the contribution of each input feature. Based on the global importance index, all input features are sorted in descending order to screen out the dominant factors affecting concentration changes. Combining the positive and negative distribution characteristics of the SHAP values ​​corresponding to the dominant factors, the positive and negative correlations between the dominant factors and the concentrations of dissolved inorganic nitrogen and dissolved inorganic phosphorus in the nearshore area are analyzed.

Images