Water chlorophyll concentration inversion method based on hyperspectral remote sensing image

Abstract

The invention provides a water chlorophyll a concentration inversion method based on a hyperspectral remote sensing image, which relates to the technical field of remote sensing inversion and comprises the steps of image preprocessing, water pixel extraction, actually measured sample space registration, spectral feature construction, model score optimization, concentration prediction and the like. According to the method, an optimal scheme is screened in a feature-model combination through cross validation and a unified scoring function, the optimal scheme is applied to whole image calculation, and a chlorophyll a concentration spatial distribution map and a matched quality control map layer are output. The method has the characteristics of high precision, self-adaption and high engineering practicability, and is suitable for a water quality remote sensing inversion scene driven by multi-source hyperspectral data.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing inversion, and in particular to a method for inverting chlorophyll a concentration in water bodies based on hyperspectral remote sensing images. Background Art

[0002] Currently, remote sensing inversion of chlorophyll a (Chl-a) concentration in water bodies mostly relies on semi-empirical band ratio models constructed from multispectral satellites (such as Sentinel-2 and Landsat-8) or empirical models based on single algorithms. Although multispectral data are convenient to acquire, their limited spectral resolution makes it difficult to separate the aliased information between chlorophyll, suspended matter, and dissolved organic matter. This results in poor model transferability across different water bodies and large fluctuations in accuracy.

[0003] The new generation of Earth observation constellations (such as Resource-1 02E / 02D) provides operational hyperspectral imagery with 30m spatial resolution, 10nm (VNIR) to 20nm (SWIR) spectral resolution, and 60km swath width. These images, combined with airborne and in-situ surface observations, form a multi-source data stream with high temporal resolution. By leveraging machine learning, ensemble learning, and cloud-based parallel processing frameworks, high-dimensional spectral signatures are deeply coupled with field-measured water quality samples, significantly improving the accuracy and spatial and temporal coverage of Chl-a retrieval.

[0004] Although the hyperspectral + machine learning method has been proven to have higher accuracy, there are still problems: (1) Traditional algorithms use a fixed single model and cannot adaptively select the optimal model based on data characteristics; (2) There are spectral response differences between the training samples and the entire image, and the atmosphere-water surface-sensor chain correction process is not coupled with the model training closed loop; (3) There is a lack of a unified method system for the entire processing process (preprocessing-feature optimization-model construction-accuracy assessment-large-area extrapolation), resulting in a high threshold for engineering application and poor versatility. Summary of the Invention

[0005] In order to overcome the shortcomings of the existing technology, the purpose of the present invention is to provide a method for inverting the chlorophyll a concentration of water bodies based on hyperspectral remote sensing images, which realizes remote sensing inversion of the chlorophyll a concentration of water bodies with high precision, high adaptability and controllable quality.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A method for inverting chlorophyll a concentration in water bodies based on hyperspectral remote sensing images, comprising:

[0008] Preprocessing and water pixel extraction are performed on the original hyperspectral remote sensing images covering the target water area to obtain a water image subset;

[0009] Performing spatial registration on the chlorophyll a concentration points measured during the satellite transit with the water body pixel subset to form a labeled sample set;

[0010] Applying each candidate feature construction method in the preset spectral feature library to the annotated sample set in turn to generate corresponding feature matrices respectively;

[0011] For each feature matrix, each algorithm in the machine learning model library is called one by one to perform K-fold cross validation and calculate the unified scoring function value;

[0012] Comparing all the unified scoring function values, selecting the feature matrix and algorithm pairing combination with the highest score as the optimal feature-model combination;

[0013] The water body pixel subset is input into the optimal feature-model combination to perform concentration prediction processing to generate a chlorophyll a concentration spatial distribution map and a matching quality control layer.

[0014] Preferably, the acquired original hyperspectral remote sensing image covering the target water area is preprocessed and water body pixels are extracted to obtain a water body image subset, including:

[0015] Performing radiometric calibration and atmospheric correction on the original hyperspectral remote sensing image to obtain an above-water reflectance image;

[0016] Calculating a water index value for each pixel in the reflectance image;

[0017] Comparing the water index value with a preset discrimination threshold, when the water index value is greater than the discrimination threshold, classifying the corresponding pixel as a water pixel and assigning a value of 1; otherwise, classifying it as a non-water pixel and assigning a value of 0;

[0018] Generate a binary mask layer consistent with the spatial size of the hyperspectral image based on the classification results of all pixels;

[0019] The binary mask layer is used to identify water body areas in the reflectance image, and corresponding pixel data are extracted as the water body image subset.

[0020] Preferably, the chlorophyll a concentration points measured during the satellite transit are spatially registered with the water body pixel subset to form a labeled sample set, including:

[0021] Obtaining ground-measured data corresponding to the target water area on the day of the satellite's transit; the ground-measured data includes sampling time, latitude and longitude coordinates, and chlorophyll a concentration value;

[0022] Based on the geo-registration information of the latitude and longitude coordinates and the original hyperspectral remote sensing image, searching for a valid pixel nearest to each of the latitude and longitude coordinates in the water body pixel subset;

[0023] Extracting the hyperspectral reflectance vector of the corresponding pixel for each of the latitude and longitude coordinates, and pairing it with the chlorophyll a concentration value of the latitude and longitude coordinates to form a sample entry;

[0024] All paired sample entries are combined into the labeled sample set.

[0025] Preferably, based on the geo-registration information of the latitude and longitude coordinates and the original hyperspectral remote sensing image, searching for the valid pixel nearest to each of the latitude and longitude coordinates in the water body pixel subset includes:

[0026] Convert each of the latitude and longitude coordinates into projection plane coordinates used by the image according to the geo-registration matrix of the original hyperspectral remote sensing image;

[0027] Calculate the row index and column index of the target pixel using the projection plane coordinates, the pixel resolution in the georeferencing matrix, and the coordinates of the upper left corner of the image;

[0028] Creating a 3×3 pixel window with the row index and the column index as the center, and screening pixels with a mask value of 1 and complete data within the window as candidate pixels;

[0029] The Euclidean distance between the latitude and longitude coordinates and the center point of each candidate pixel is calculated, and the pixel with the smallest distance is selected as the nearest valid neighbor pixel.

[0030] Preferably, the calculation formulas for the row index and the column index are respectively:

[0031] c idx =round(c)

[0032] r idx =round(r)

[0033] Among them, c idx and r idx are the column index and the row index respectively, c and r are the continuous column and row coordinate values ​​respectively, ΔX and ΔY are the horizontal and vertical distances between the latitude and longitude coordinates and the center of the upper left corner pixel, respectively. x is the pixel resolution of the original hyperspectral remote sensing image in the column direction; R y is the pixel resolution in the row direction of the original hyperspectral remote sensing image, θ x and θ ydenotes the horizontal and vertical rotational coupling between the column and row directions, respectively, and is used to represent the geometric offset relationship when the original hyperspectral remote sensing image has a non-orthogonal posture. detA denotes the determinant of the original hyperspectral remote sensing image registration affine matrix, which is used for normalized coordinate transformation. round(·) denotes the rounding function.

[0034] Preferably, each candidate feature construction method in the preset spectral feature library is applied to the annotated sample set in turn to generate corresponding feature matrices, including:

[0035] Reading the function definition and parameter template of each candidate feature construction method from the spectral feature library according to a preset order;

[0036] For each candidate feature construction method, the hyperspectral reflectance vectors in the labeled sample set are called one by one, feature calculation is performed, and the corresponding sample feature row vector is obtained;

[0037] Stacking all sample feature row vectors row by row in a predetermined order in the labeled sample set to form a feature matrix with a feature dimension consistent with the output dimension of the current candidate feature construction method;

[0038] The feature matrix is ​​associated with the unique identifier of the candidate feature construction method and stored to complete a one-to-one mapping of features and methods.

[0039] Preferably, for each candidate feature construction method, the hyperspectral reflectance vectors in the labeled sample set are called one by one, feature calculation is performed, and the corresponding sample feature row vector is obtained, including:

[0040] All the hyperspectral reflectance vectors of the labeled sample set are spliced ​​into a two-dimensional reflectance matrix in the preset order. Among them, S is the number of samples and B is the number of bands;

[0041] Read the band weight vector from the current candidate feature construction method And the nonlinear transformation function ψ preset by the feature construction method k k (·), forming a vectorized feature operator;

[0042] The two-dimensional reflectivity matrix Z and the band weight vector w k Perform element-wise multiplication to obtain the intermediate matrix Among them, 1 S is a column vector of length S, which is used to convert w k Expand to a dimension that matches the number of rows in Z;

[0043] In the intermediate matrix T (k) Call the feature transformation function element by element to get the feature matrix U(k) =ψ k (T (k) ); wherein the sth row is the sample feature row vector z s is the hyperspectral reflectance vector of the sth sample.

[0044] Preferably, for each feature matrix, each algorithm in the machine learning model library is called one by one to perform K-fold cross validation and calculate the unified scoring function value, including:

[0045] The feature matrix U (k) The corresponding chlorophyll a concentration label vector Perform corresponding combinations to obtain training data pairs (U (k) ,y);

[0046] Select each candidate regression algorithm model from the machine learning model library in turn And based on the training data pair (U (k) ,y) Perform K-fold cross validation training and verification process to obtain the prediction vector under the candidate regression algorithm model

[0047] Calculate the prediction vector The coefficient of determination between the chlorophyll a concentration label vector y Root mean square error (RMSE) (k,j) and mean absolute error MAE (k,j) ;

[0048] The coefficient of determination Root mean square error (RMSE) (k,j) and mean absolute error MAE (k,j) Substitute the unified scoring function Get the unified scoring function value; where α, β, and γ are weight coefficients, satisfying α+β+γ=1, and α>β>γ>0, RMSE max 、MAE max The maximum RMSE and maximum MAE of the corresponding errors of all candidate regression algorithm models under the feature matrix are respectively; Score (k,j) The unified scoring function value corresponding to each pair of feature construction method k and model algorithm j.

[0049] Preferably, all the unified scoring function values ​​are compared, and the paired combination of the feature matrix and the algorithm with the highest score is selected as the optimal feature-model combination, including:

[0050] Establishing a mapping relationship between the unified scoring function value and the corresponding combination number;

[0051] Identifying the combination number having the highest unified scoring function value;

[0052] The corresponding feature construction method and machine learning model algorithm are determined according to the combination number with the highest unified scoring function value as the optimal feature-model combination for the current task.

[0053] Preferably, the water pixel subset is input into the optimal feature-model combination to perform concentration prediction processing to generate a chlorophyll a concentration spatial distribution map and a supporting quality control layer, including:

[0054] Based on the optimal feature construction method in the optimal feature-model combination, feature conversion is performed on the hyperspectral reflectance vector of each pixel in the water body pixel subset to generate a corresponding pixel feature data matrix;

[0055] Inputting the pixel feature data matrix into the optimal machine learning model in the optimal feature-model combination, performing concentration prediction, and obtaining the predicted chlorophyll a concentration values ​​of all water body pixels;

[0056] Reconstructing the predicted chlorophyll a concentration values ​​of all water body pixels according to their spatial positions in the water body pixel subset to generate the chlorophyll a concentration spatial distribution map having spatial consistency with the original hyperspectral remote sensing image;

[0057] The feature integrity status, model response confidence level, and cloud cover mark information of each pixel during the prediction process are synchronously recorded to generate the corresponding quality control layer.

[0058] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0059] By integrating hyperspectral remote sensing imagery with ground-based data, this paper constructs a comprehensive water chlorophyll a concentration inversion process, encompassing key steps such as image preprocessing, water extraction, sample registration, feature construction, model optimization, and concentration prediction. By utilizing a candidate feature approach and a multi-model combination for scoring and screening, the paper adaptively selects the optimal inversion solution, improving model accuracy and generalization. The resulting concentration map and quality control layer have a consistent spatial structure, facilitating downstream applications and analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0061] Figure 1A flow chart of a method provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0062] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0063] The purpose of the present invention is to provide a method for inverting chlorophyll a concentration in water bodies based on hyperspectral remote sensing images, which realizes remote sensing inversion of chlorophyll a concentration in water bodies with high precision, high adaptability and controllable quality.

[0064] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0065] Figure 1 A flow chart of the method provided in the embodiment of the present invention is shown in FIG. Figure 1 As shown, the present invention provides a method for inverting chlorophyll a concentration in water bodies based on hyperspectral remote sensing images, comprising:

[0066] Step 100: preprocessing and extracting water body pixels from the acquired original hyperspectral remote sensing image covering the target water area to obtain a water body image subset;

[0067] Step 200: spatially registering the chlorophyll a concentration points measured during the satellite transit with the water pixel subset to form a labeled sample set;

[0068] Step 300: applying each candidate feature construction method in the preset spectral feature library to the annotated sample set in turn to generate corresponding feature matrices respectively;

[0069] Step 400: For each feature matrix, call each algorithm in the machine learning model library one by one to perform K-fold cross validation and calculate the unified scoring function value;

[0070] Step 500: Compare all unified scoring function values ​​and select the feature matrix and algorithm pairing combination with the highest score as the optimal feature-model combination;

[0071] Step 600: Input the water pixel subset into the optimal feature-model combination to perform concentration prediction processing to generate a chlorophyll a concentration spatial distribution map and a corresponding quality control layer.

[0072] Preferably, the acquired original hyperspectral remote sensing image covering the target water area is preprocessed and water body pixels are extracted to obtain a water body image subset, including:

[0073] Performing radiometric calibration and atmospheric correction on the original hyperspectral remote sensing image to obtain an above-water reflectance image;

[0074] Calculating a water index value for each pixel in the reflectance image;

[0075] Comparing the water index value with a preset discrimination threshold, when the water index value is greater than the discrimination threshold, classifying the corresponding pixel as a water pixel and assigning a value of 1; otherwise, classifying it as a non-water pixel and assigning a value of 0;

[0076] Generate a binary mask layer consistent with the spatial size of the hyperspectral image based on the classification results of all pixels;

[0077] The binary mask layer is used to identify water body areas in the reflectance image, and corresponding pixel data are extracted as the water body image subset.

[0078] In this implementation, a raw hyperspectral remote sensing image covering the target water area is first acquired. A radiometric calibration algorithm is used to convert the digital values ​​(DN) of each band into physical reflectance units. The calibration results are then atmospherically corrected using the 6S atmospheric radiation transfer model or an equivalent atmospheric correction algorithm to obtain a true reflectance image above the water surface. This reflectance image retains multi-band spectral information, serving as the basis for subsequent water body identification and spectral feature extraction.

[0079] Based on the reflectance image, a typical water body identification index (such as the Normalized Difference Water Index (NDWI)) is calculated and a binary classification is performed for each pixel, either water or non-water, based on a preset threshold. When the water index value of a pixel exceeds the threshold, it is identified as a water pixel and assigned a value of 1; otherwise, it is assigned a value of 0. This generates a binary mask layer with the same spatial resolution as the original image. By applying this mask layer to the reflectance image, pixel data belonging to water areas can be effectively extracted, forming a water image subset for subsequent chlorophyll a concentration inversion analysis.

[0080] Furthermore, this embodiment uses the Ziyuan-1 02E (ZY1-02E) satellite, which is equipped with a new-generation AHSI hyperspectral camera with 166 spectral bands. It acquires image data with a spectral resolution of 10nm and 20nm in the visible-near infrared (VN) and shortwave infrared (SW) ranges, respectively. Furthermore, its spatial resolution reaches 30m, and the image width is 60km. In dual-satellite network operation mode, the satellite system further enhances its revisit monitoring capabilities. The data used in this embodiment comes from the hyperspectral imagery acquired by the Ziyuan-1 02E satellite.

[0081] Since the original ZY1-02EASHI image cannot be used directly, this embodiment requires a series of image processing through ENVI. The reflectance of the ground object after FLAASH atmospheric correction needs to be corrected for water reflectance to obtain accurate remote sensing reflectance. The calculation formula is as follows:

[0082]

[0083] Where, represents the remote sensing reflectance estimated approximately on the image, and ρ represents the surface reflectance.

[0084] After preprocessing the image, the Normalized Difference Water Index (NDWI) is calculated and the threshold is selected to extract the water body in the study area. The NDWI calculation formula is as follows:

[0085] NDWI=(Green-NIR) / (Green+NIR)

[0086] Where Green is the green light band and NIR is the near infrared band.

[0087] Preferably, the chlorophyll a concentration points measured during the satellite transit are spatially registered with the water body pixel subset to form a labeled sample set, including:

[0088] Obtaining ground-measured data corresponding to the target water area on the day of the satellite's transit; the ground-measured data includes sampling time, latitude and longitude coordinates, and chlorophyll a concentration value;

[0089] Based on the geo-registration information of the latitude and longitude coordinates and the original hyperspectral remote sensing image, searching for a valid pixel nearest to each of the latitude and longitude coordinates in the water body pixel subset;

[0090] Extracting the hyperspectral reflectance vector of the corresponding pixel for each of the latitude and longitude coordinates, and pairing it with the chlorophyll a concentration value of the latitude and longitude coordinates to form a sample entry;

[0091] All paired sample entries are combined into the labeled sample set.

[0092] In this implementation, ground-based sample data is first collected to match the satellite transit time. The measured data includes the latitude and longitude coordinates of multiple sample points in the target water area, the corresponding sampling time, and the measured chlorophyll a concentration. These latitude and longitude coordinates are obtained using a GPS device, and the time accuracy should meet the requirement of no more than two hours of error from the image acquisition time to ensure data synchronization. After formatting, the sampled data forms a basic table of measured samples, which is then used for subsequent spatial matching with the hyperspectral imagery.

[0093] The latitude and longitude coordinates of each measured point are converted to projected plane coordinates consistent with the original hyperspectral remote sensing image through a unified projection transformation. The initial row and column index position of each point in the image is calculated based on the image's georeferencing matrix. Within a local pixel window centered on this index, water pixels with a mask value of 1 and no missing data are selected. The nearest valid neighbor pixel is determined based on the principle of minimum center-point distance. Subsequently, the corresponding hyperspectral reflectance vector is extracted from the selected pixel and paired one-to-one with the measured chlorophyll a concentration value to form a complete sample entry. All paired sample entries are organized sequentially to form a labeled sample set for subsequent feature construction and model training.

[0094] Preferably, based on the geo-registration information of the latitude and longitude coordinates and the original hyperspectral remote sensing image, searching for the valid pixel nearest to each of the latitude and longitude coordinates in the water body pixel subset includes:

[0095] Convert each of the latitude and longitude coordinates into projection plane coordinates used by the image according to the geo-registration matrix of the original hyperspectral remote sensing image;

[0096] Calculate the row index and column index of the target pixel using the projection plane coordinates, the pixel resolution in the georeferencing matrix, and the coordinates of the upper left corner of the image;

[0097] Creating a 3×3 pixel window with the row index and the column index as the center, and screening pixels with a mask value of 1 and complete data within the window as candidate pixels;

[0098] The Euclidean distance between the latitude and longitude coordinates and the center point of each candidate pixel is calculated, and the pixel with the smallest distance is selected as the nearest valid neighbor pixel.

[0099] In this implementation, the six-parameter affine geographic transformation matrix is ​​first parsed based on the georeferencing information attached to the original hyperspectral remote sensing image. This matrix includes the projected coordinates of the image's upper left corner, the spatial resolution in the row and column directions, and two rotation terms. For each ground-based measurement point with latitude and longitude coordinates, a unified projective transformation method (e.g., WGS84 to UTM) is first used to convert these coordinates into the image's corresponding projected plane coordinates. Based on the affine transformation relationship, the row and column index of the coordinate in the image is calculated using the relative position of these coordinates to a reference point in the image's upper left corner.

[0100] After obtaining the image row and column index corresponding to the target point, a 3×3 pixel window is constructed in the water pixel subset with it as the center. All pixels within the window are screened, and invalid pixels with a mask value of 0 or missing spectral data are eliminated. The set of candidate pixels with complete data and identified as water bodies is retained. Subsequently, the spatial Euclidean distance between the center coordinates of each candidate pixel and the target latitude and longitude coordinates is compared, and the pixel with the closest distance is selected as the nearest neighbor valid pixel corresponding to the measured point and used for subsequent extraction of the hyperspectral reflectance vector. This method ensures accurate pairing of the image space and the ground sampling location, effectively improving the sample annotation accuracy and the representativeness of the model training.

[0101] Preferably, the calculation formulas for the row index and the column index are respectively:

[0102] c idx =round(c)

[0103] r idx =round(r)

[0104] Among them, c idx and r idx are the column index and the row index respectively, c and r are the continuous column and row coordinate values ​​respectively, ΔX and ΔY are the horizontal and vertical distances between the latitude and longitude coordinates and the center of the upper left corner pixel, respectively. x is the pixel resolution of the original hyperspectral remote sensing image in the column direction; R y is the pixel resolution in the row direction of the original hyperspectral remote sensing image, θ x and θ y denotes the horizontal and vertical rotational coupling between the column and row directions, respectively, and is used to represent the geometric offset relationship when the original hyperspectral remote sensing image has a non-orthogonal posture. detA denotes the determinant of the original hyperspectral remote sensing image registration affine matrix, which is used for normalized coordinate transformation. round(·) denotes the rounding function.

[0105] This formula is used to convert the latitude and longitude coordinates of a ground-based measurement point into pixel row and column indices in the original hyperspectral remote sensing imagery, enabling spatial registration and pixel location. Specifically, the latitude and longitude coordinates of the measured point are first converted to projected coordinates using a unified projection transformation (e.g., WGS84 to UTM or the image's corresponding projection coordinate system). Next, six affine transformation parameters are derived from the image's accompanying georeferencing information: the horizontal and vertical projection coordinates of the pixel center in the upper left corner of the image, the pixel spatial resolution in the column and row directions, and the rotational coupling in both directions. Based on these parameters and the coordinate difference between the measured point and the image's upper left corner, the continuous column and row coordinates of the measured point in the image are calculated using the inverse affine transformation. Finally, these continuous coordinate values ​​are rounded to the nearest integer to obtain the final integer column and row indices, representing the pixel location of the measured point in the image data matrix. The affine parameters relied upon in this process are typically provided directly by the remote sensing image provider in the image metadata or image header file, eliminating the need for user calculation and allowing for programmatic coordinate conversion and index location. This method is applicable to any remote sensing image product containing georeferenced information.

[0106] Preferably, each candidate feature construction method in the preset spectral feature library is applied to the annotated sample set in turn to generate corresponding feature matrices, including:

[0107] Reading the function definition and parameter template of each candidate feature construction method from the spectral feature library according to a preset order;

[0108] For each candidate feature construction method, the hyperspectral reflectance vectors in the labeled sample set are called one by one, feature calculation is performed, and the corresponding sample feature row vector is obtained;

[0109] Stacking all sample feature row vectors row by row in a predetermined order in the labeled sample set to form a feature matrix with a feature dimension consistent with the output dimension of the current candidate feature construction method;

[0110] The feature matrix is ​​associated with the unique identifier of the candidate feature construction method and stored to complete a one-to-one mapping of features and methods.

[0111] In this embodiment, the definition information for each candidate feature construction method is first read sequentially from a preset spectral feature library. Each construction method contains a corresponding function expression, the required input band index, the feature output dimension, and a description of the operation logic. For example, it may include normalized difference features (such as band ratios), principal component features, spectral position difference features, and first-order derivative or second-order derivative transformations. Each construction method is assigned a unique identifier to ensure its traceability and result mapping consistency during the model construction process. The candidate feature construction methods in the spectral feature library are pre-configured in a structured manner, each method is assigned a unique number and a fixed calling order. This preset order can be set based on historical experience, feature expression ability, physical meaning, or computational complexity, or it can be sorted by feature group category, such as executing linear combination methods first and then nonlinear transformation methods. Once the preset order is determined, it will be read and applied in strict accordance with this order during the model construction process to ensure controllable feature construction and experimental repeatability. This order can be configured by the user during initialization or set by default to a standard order list.

[0112] For a certain candidate feature construction method, all hyperspectral reflectance vectors in the annotated sample set are used as input, and feature construction operations are performed in batches according to the sample order. During the feature construction process, a one-time feature transformation is performed on each reflectance vector according to the predefined band combination and calculation rules of the method to obtain the corresponding feature vector. The feature vectors generated by all samples are arranged in rows to form a feature matrix with consistent dimensions. After the matrix generation is completed, the matrix is ​​bound to the identifier of the current feature method and stored as the basis for training input and model scoring. By traversing the entire feature library, multiple feature matrix sets with unified structure but different feature expressions can be generated in sequence, supporting subsequent model optimization and generalization performance comparison.

[0113] For example, each candidate feature construction method explicitly specifies the required input band indexes and specific mathematical operation rules in its definition template. Band combinations can be set based on typical water quality spectral response characteristics, such as the red-edge band, green band, or near-infrared band associated with chlorophyll a absorption. Calculation rules can include standard algorithms such as normalized difference, linear weighting, principal component compression, and derivative calculation. This template is set by domain experts or model designers during feature library configuration. It can use a fixed combination or be dynamically generated based on the spectral response range of the target sensor, ensuring that the constructed features are both physically interpretable and discriminative.

[0114] Preferably, for each candidate feature construction method, the hyperspectral reflectance vectors in the labeled sample set are called one by one, feature calculation is performed, and the corresponding sample feature row vector is obtained, including:

[0115] All the hyperspectral reflectance vectors of the labeled sample set are spliced ​​into a two-dimensional reflectance matrix in the preset order. Among them, S is the number of samples and B is the number of bands;

[0116] Read the band weight vector from the current candidate feature construction method And the nonlinear transformation function ψ preset by the feature construction method k k (·), forming a vectorized feature operator;

[0117] The two-dimensional reflectivity matrix Z and the band weight vector w k Perform element-wise multiplication to obtain the intermediate matrix Among them, 1 S is a column vector of length S, which is used to convert w k Expand to a dimension that matches the number of rows in Z;

[0118] In the intermediate matrix T (k) Call the feature transformation function element by element to get the feature matrix U (k) =ψ k (T (k) ); wherein the sth row is the sample feature row vector z s is the hyperspectral reflectance vector of the sth sample.

[0119] In this implementation, all hyperspectral reflectance vectors in the annotated sample set are first read in a preset order and concatenated into a two-dimensional reflectance matrix in the order in which the samples are arranged. Each row represents the reflectance vector for a sample, and each column corresponds to a specific band. Assuming the number of samples is S and the number of bands is B, the dimensions of the concatenated matrix are S rows and B columns. For the current candidate feature construction method, a preset band weight vector is obtained from the feature definition template. This vector has a dimension that matches the number of bands, and its corresponding feature constructor is read in conjunction with the vector. This involves performing nonlinear transformations, normalized mappings, or mathematical combination processing on the input values.

[0120] The two-dimensional reflectivity matrix is ​​element-wise multiplied with the band weight vector of the current candidate method. To achieve batch operations, the weight vector is first expanded in the sample dimension to a dimension consistent with the number of rows in the reflectivity matrix, so that it can be multiplied one-to-one with the reflectivity vector of each sample to form an intermediate matrix. Subsequently, the nonlinear feature constructor is executed element-by-element on the intermediate matrix to generate a feature matrix with consistent dimensions. Each row in the matrix is ​​the feature row vector corresponding to the sample under the feature construction method. The feature matrix will be used as input for subsequent modeling and evaluation and stored together with the unique identifier of the current feature construction method to ensure that the results are traceable and consistent with the method matching.

[0121] Preferably, for each feature matrix, each algorithm in the machine learning model library is called one by one to perform K-fold cross validation and calculate the unified scoring function value, including:

[0122] The feature matrix U (k) The corresponding chlorophyll a concentration label vector Perform corresponding combinations to obtain training data pairs (U (k) ,y);

[0123] Select each candidate regression algorithm model from the machine learning model library in turn And based on the training data pair (U (k) ,y) Perform K-fold cross validation training and verification process to obtain the prediction vector under the candidate regression algorithm model

[0124] Calculate the prediction vector The coefficient of determination between the chlorophyll a concentration label vector y Root mean square error (RMSE) (k,j) and mean absolute error MAE (k,j) ;

[0125] The coefficient of determination Root mean square error (RMSE) (k,j) and mean absolute error MAE (k,j) Substitute the unified scoring function Get the unified scoring function value; where α, β, and γ are weight coefficients, satisfying α+β+γ=1, and α>β>γ>0, RMSE max 、MAE max The maximum RMSE and maximum MAE of the corresponding errors of all candidate regression algorithm models under the feature matrix are respectively; Score (k,j) The unified scoring function value corresponding to each pair of feature construction method k and model algorithm j.

[0126] In this implementation, a feature matrix generated by a feature construction method is first combined with its corresponding chlorophyll a concentration label vector to construct training data pairs. Each row of the feature matrix corresponds one-to-one with each element in the label vector, representing the input variable and target output value for the same sample. This combination forms a structured training set that can be used by each candidate machine learning model.

[0127] Subsequently, each regression algorithm model is selected from the built-in machine learning model library for evaluation, such as random forest regression, support vector regression, gradient boosted regression tree, extreme random tree, etc. For each model, K-fold cross-validation is performed on the training data. This means that the training set is divided into K equally divided data subsets, one of which is used as the validation set and the remaining subsets as the training set in turn. The training-prediction process is repeated K times to obtain a stable prediction result for the model on the feature matrix, and a prediction vector consisting of a set of prediction values ​​equal to the number of samples is output.

[0128] After the prediction is complete, the predicted vector is compared with the original label vector for error evaluation. Three evaluation metrics are calculated: the coefficient of determination (CDR), the root mean square error (RMSE), and the mean absolute error (MAE). The CDR measures the model's fit, the RMSE reflects the squared average of the prediction deviations, and the MAE measures the average difference between the predicted value and the true value. These three metrics collectively characterize the model's performance under the current feature construction method and provide a basis for subsequent unified scoring.

[0129] In order to achieve objective comparison between multiple models, this embodiment constructs a unified scoring function and integrates the above three indicators in a weighted manner. Specifically, the scoring function uses the coefficient of determination as a positive indicator, and at the same time performs maximum normalization processing on the root mean square error and the mean absolute error, and then performs reverse weighted integration to form a standardized scoring value. Among them, each weight coefficient satisfies the constraint condition that the sum is 1, and the weight of the coefficient of determination is greater than the weights of the other two error terms to ensure that the scoring result gives priority to the fitting ability of the model. Finally, the scoring value is bound to the corresponding feature-model combination for scoring ranking and decision selection in the subsequent optimization stage.

[0130] Specifically, the machine learning model library of this embodiment includes:

[0131] (1) Random Forest Model

[0132] Random forest is a classic algorithm in bagging. Its basic principle is to construct multiple weak estimators in parallel and combine their predictions to obtain a final output. Random forest samples the training set with replacement to generate multiple sample subsets and independently constructs a decision tree on each subset. At each node in the tree, the algorithm randomly selects a subset of features and finds the optimal feature to split the node. Ultimately, the output of the random forest is the integration of the predictions of all decision trees.

[0133]

[0134] Where: T is the number of decision trees, x is the independent variable, which comes from the training set, is the result output by the random forest in the training set, and DT is the result output by a single decision tree.

[0135] (2) XGBoost model

[0136] XGBoost (EXtreme Gradient Boosting) is a new generation algorithm based on GBDT. Its base tree model can fit nonlinear data well. The calculation formula is as follows:

[0137]

[0138] Where represents the model prediction value, represents the input data, K is the number of trees in the model, is the output result of a single tree, and F is the tree space (usually CART tree).

[0139] The basic core idea of ​​XGBoost is consistent with GBDT, but it achieves a balance between accuracy and complexity. This is mainly reflected in the fact that XGBoost adds a structural risk term to the loss function to form the objective function, which is expressed as:

[0140]

[0141] Where SRI is the structural risk term. This change makes XGBoost train in the direction of minimizing the objective function rather than minimizing the loss function.

[0142] Preferably, all the unified scoring function values ​​are compared, and the paired combination of the feature matrix and the algorithm with the highest score is selected as the optimal feature-model combination, including:

[0143] Establishing a mapping relationship between the unified scoring function value and the corresponding combination number;

[0144] Identifying the combination number having the highest unified scoring function value;

[0145] The corresponding feature construction method and machine learning model algorithm are determined according to the combination number with the highest unified scoring function value as the optimal feature-model combination for the current task.

[0146] For example, after determining the optimal feature-model combination, this embodiment also evaluates and verifies the machine model in the combination. Preferably, in order to evaluate the inversion effect of the random forest model and the XGBoost model applied to remote sensing images, this embodiment uses the measured chlorophyll a concentration to evaluate the accuracy of the inversion results. This embodiment uses the determination coefficient R 2, Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) evaluation indicators, taking the measured chlorophyll a concentration as the true value, evaluate the accuracy of the image inversion results. 2 The calculation formulas for , RMSE and MAE are:

[0147]

[0148] Where n is the number of samples, y i is the i-th observation, is the i-th predicted value, is the mean of the observations.

[0149] This example selects all bands from 400 nm to 900 nm for water reflectance correction. Correlation analysis is performed between the corrected ground reflectance and the measured chlorophyll-a concentration to obtain the top three bands with the best correlation. A band ratio model is then constructed using all corrected bands. Correlation analysis is then performed between this constructed band ratio model and the measured chlorophyll-a concentration to obtain the top three band combinations with the best correlation.

[0150] In this embodiment, for each candidate feature construction method and each machine learning model algorithm combination, after completing cross-validation and calculating the unified scoring function value, this embodiment automatically establishes a one-to-one mapping relationship between each scoring value and its corresponding combination number. This combination number can be represented in the form of a "feature method number + model algorithm number" structure to ensure that each scoring value can be uniquely mapped to a specific feature-model combination. All scoring values ​​and their corresponding numbers are uniformly recorded in a structured format during the storage process for call and analysis in the optimization stage.

[0151] After calculating and recording all combined scores, this embodiment iterates through all scores, identifies the highest value, and obtains its corresponding combination number. This number directly indexes the optimal feature construction method and machine learning model pairing, which the system determines as the optimal feature-model combination for the current task. This optimal combination not only has the highest unified scoring function performance but also has the best adaptability to the specific water body data sample set. This combination will be used as the basis for subsequent concentration inversion processing of water body pixels across the entire region.

[0152] Preferably, the water pixel subset is input into the optimal feature-model combination to perform concentration prediction processing to generate a chlorophyll a concentration spatial distribution map and a supporting quality control layer, including:

[0153] Based on the optimal feature construction method in the optimal feature-model combination, feature conversion is performed on the hyperspectral reflectance vector of each pixel in the water body pixel subset to generate a corresponding pixel feature data matrix;

[0154] Inputting the pixel feature data matrix into the optimal machine learning model in the optimal feature-model combination, performing concentration prediction, and obtaining the predicted chlorophyll a concentration values ​​of all water body pixels;

[0155] Reconstructing the predicted chlorophyll a concentration values ​​of all water body pixels according to their spatial positions in the water body pixel subset to generate the chlorophyll a concentration spatial distribution map having spatial consistency with the original hyperspectral remote sensing image;

[0156] The feature integrity status, model response confidence level, and cloud cover mark information of each pixel during the prediction process are synchronously recorded to generate the corresponding quality control layer.

[0157] In this implementation, the optimal feature construction method and machine learning model combination are applied to all hyperspectral reflectance vector processing tasks in the water pixel subset. First, the system uses the optimal feature construction method to extract features from each pixel reflectance vector in the water pixel subset, generating a complete pixel feature data matrix. The structure of this feature matrix is ​​consistent with the input feature structure used in the model training phase, ensuring input consistency and result validity during the model prediction process.

[0158] The generated pixel feature data matrix is ​​then input into the optimal machine learning model to perform concentration prediction, obtaining chlorophyll a concentration prediction values ​​corresponding to each water pixel. The system reconstructs these concentration values ​​in combination with the spatial coordinate information of the original image to generate a spatial distribution map of chlorophyll a concentration consistent with the resolution of the hyperspectral image. At the same time, the system generates a quality control layer based on whether each pixel participates in the modeling, the stability of the model output, and the mask status corresponding to the pixel, marking abnormal areas, invalid predictions, or occluded areas. The final output includes a concentration distribution map and a quality control layer, which can be used for subsequent visualization, water quality assessment, or monitoring and early warning systems.

[0159] The beneficial effects of the present invention are as follows:

[0160] This paper provides a method for inverting chlorophyll a concentration in water bodies by integrating hyperspectral remote sensing imagery with ground-based data. The method constructs a complete workflow from data preprocessing, sample registration, feature construction, model optimization, to concentration prediction output. The method features clear module divisions, with explicit data transfer logic between steps, and an overall structure that is highly versatile and scalable.

[0161] The present invention introduces a combined cross-mechanism of the spectral feature library and the machine learning model library, uses a unified scoring function to quantitatively evaluate different feature and model combinations, and adaptively selects the optimal inversion scheme, thereby significantly improving the adaptability and prediction accuracy of the model under different water conditions.

[0162] In the inversion output stage, the present invention not only generates a spatial distribution map of chlorophyll a concentration, but also simultaneously constructs a quality control layer to mark the prediction reliability of each pixel, thereby enhancing the interpretability and engineering usability of the results and facilitating subsequent monitoring system integration and decision-making support applications.

[0163] Overall, the method of the present invention has the advantages of complete process closed-loop, high model accuracy, flexible feature expression, accurate spatial positioning, and standardized result expression. It is suitable for water quality remote sensing inversion scenarios driven by multi-source hyperspectral data and has good prospects for promotion and application.

[0164] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0165] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the method and core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A method for inverting chlorophyll a concentration in water bodies based on hyperspectral remote sensing images, characterized in that: include: Preprocessing and water pixel extraction are performed on the original hyperspectral remote sensing images covering the target water area to obtain a water image subset; Performing spatial registration on the chlorophyll a concentration points measured during the satellite transit with the water body pixel subset to form a labeled sample set; Applying each candidate feature construction method in the preset spectral feature library to the annotated sample set in turn to generate corresponding feature matrices respectively; For each feature matrix, each algorithm in the machine learning model library is called one by one to perform K-fold cross validation and calculate the unified scoring function value; Comparing all the unified scoring function values, selecting the feature matrix and algorithm pairing combination with the highest score as the optimal feature-model combination; The water body pixel subset is input into the optimal feature-model combination to perform concentration prediction processing to generate a chlorophyll a concentration spatial distribution map and a matching quality control layer.

2. The method for inverting chlorophyll a concentration in water bodies based on hyperspectral remote sensing images according to claim 1, characterized in that: The original hyperspectral remote sensing images covering the target water area are preprocessed and water pixels are extracted to obtain a water image subset, including: Performing radiometric calibration and atmospheric correction on the original hyperspectral remote sensing image to obtain an above-water reflectance image; Calculating a water index value for each pixel in the reflectance image; Comparing the water index value with a preset discrimination threshold, when the water index value is greater than the discrimination threshold, classifying the corresponding pixel as a water pixel and assigning a value of 1; otherwise, classifying it as a non-water pixel and assigning a value of 0; Generate a binary mask layer consistent with the spatial size of the hyperspectral image based on the classification results of all pixels; The binary mask layer is used to identify water body areas in the reflectance image, and corresponding pixel data are extracted as the water body image subset.

3. The method for inverting chlorophyll a concentration in water bodies based on hyperspectral remote sensing images according to claim 2, characterized in that: The chlorophyll a concentration points measured during the satellite transit are spatially registered with the water body pixel subset to form a labeled sample set, including: Obtaining ground-measured data corresponding to the target water area on the day of the satellite's transit; the ground-measured data includes sampling time, latitude and longitude coordinates, and chlorophyll a concentration value; Based on the geographic registration information of the latitude and longitude coordinates and the original hyperspectral remote sensing image, searching for the valid pixel nearest to each of the latitude and longitude coordinates in the water body pixel subset; Extracting the hyperspectral reflectance vector of the corresponding pixel for each of the latitude and longitude coordinates, and pairing it with the chlorophyll a concentration value of the latitude and longitude coordinates to form a sample entry; All paired sample entries are combined into the labeled sample set.

4. The method for inverting chlorophyll a concentration in water bodies based on hyperspectral remote sensing images according to claim 3, characterized in that: Based on the geo-registration information of the latitude and longitude coordinates and the original hyperspectral remote sensing image, searching for a valid pixel nearest to each of the latitude and longitude coordinates in the water body pixel subset includes: Convert each of the latitude and longitude coordinates into projection plane coordinates used by the image according to the geo-registration matrix of the original hyperspectral remote sensing image; Calculate the row index and column index of the target pixel using the projection plane coordinates, the pixel resolution in the georeferencing matrix, and the coordinates of the upper left corner of the image; Creating a 3×3 pixel window with the row index and the column index as the center, and screening pixels with a mask value of 1 and complete data within the window as candidate pixels; The Euclidean distance between the latitude and longitude coordinates and the center point of each candidate pixel is calculated, and the pixel with the smallest distance is selected as the nearest valid neighbor pixel.

5. The method for inverting chlorophyll a concentration in water bodies based on hyperspectral remote sensing images according to claim 4, characterized in that: The calculation formulas for the row index and the column index are respectively: c idx =round(c) r idx =round(r) Among them, c idx and r idx are the column index and the row index respectively, c and r are the continuous column and row coordinate values ​​respectively, ΔX and ΔY are the horizontal and vertical distances between the latitude and longitude coordinates and the center of the upper left pixel in the projection plane, respectively. x is the pixel resolution in the original hyperspectral remote sensing image sequence direction; R y is the pixel resolution in the row direction of the original hyperspectral remote sensing image, θ x and θ y denotes the horizontal and vertical rotational coupling between the column and row directions, respectively, and is used to represent the geometric offset relationship when the original hyperspectral remote sensing image has a non-orthogonal posture. detA denotes the determinant of the original hyperspectral remote sensing image registration affine matrix, which is used for normalized coordinate transformation. round(·) denotes the rounding function.

6. The method for inverting chlorophyll a concentration in water bodies based on hyperspectral remote sensing images according to claim 1, characterized in that: Apply each candidate feature construction method in the preset spectral feature library to the annotated sample set in turn to generate corresponding feature matrices, including: Reading the function definition and parameter template of each candidate feature construction method from the spectral feature library according to a preset order; For each candidate feature construction method, the hyperspectral reflectance vectors in the labeled sample set are called one by one, feature calculation is performed, and the corresponding sample feature row vector is obtained; Stacking all sample feature row vectors row by row in a predetermined order in the labeled sample set to form a feature matrix with a feature dimension consistent with the output dimension of the current candidate feature construction method; The feature matrix is ​​associated with the unique identifier of the candidate feature construction method and stored to complete a one-to-one mapping of features and methods.

7. The method for inverting chlorophyll a concentration in water bodies based on hyperspectral remote sensing images according to claim 6, characterized in that: For each candidate feature construction method, the hyperspectral reflectance vectors in the labeled sample set are called one by one, and feature calculation is performed to obtain the corresponding sample feature row vector, including: All the hyperspectral reflectance vectors of the labeled sample set are spliced ​​into a two-dimensional reflectance matrix in the preset order. Among them, S is the number of samples and B is the number of bands; Read the band weight vector from the current candidate feature construction method And the nonlinear transformation function ψ preset by the feature construction method k k (·), forming a vectorized feature operator; The two-dimensional reflectivity matrix Z and the band weight vector w k Perform element-wise multiplication to obtain the intermediate matrix Among them, 1 S is a column vector of length S, which is used to convert w k Expand to a dimension that matches the number of rows in Z; In the intermediate matrix T (k) Call the feature transformation function element by element to get the feature matrix U (k) =ψ k (T (k) ); wherein the sth row is the sample feature row vector z s is the hyperspectral reflectance vector of the sth sample.

8. The method for inverting chlorophyll a concentration in water bodies based on hyperspectral remote sensing images according to claim 7, characterized in that: For each feature matrix, call each algorithm in the machine learning model library one by one to perform K-fold cross validation and calculate the unified scoring function value, including: The feature matrix U (k) The corresponding chlorophyll a concentration label vector Perform corresponding combinations to obtain training data pairs (U (k) ,y); Select each candidate regression algorithm model from the machine learning model library in turn And based on the training data pair (U (k) ,y) Perform K-fold cross validation training and verification process to obtain the prediction vector under the candidate regression algorithm model Calculate the prediction vector The coefficient of determination between the chlorophyll a concentration label vector y Root mean square error (RMSE) (k,j) and mean absolute error MAE (k,j) ; The coefficient of determination Root mean square error (RMSE) (k , j) and mean absolute error MAE (k,j) Substitute the unified scoring function Get the unified scoring function value; where α, β, and γ are weight coefficients, satisfying α+β+γ=1, and α>β>γ>0, RMSE max 、MAE max The maximum RMSE and maximum MAE of the corresponding errors of all candidate regression algorithm models under the feature matrix are respectively; Score (k,j) The unified scoring function value corresponding to each pair of feature construction method k and model algorithm j.

9. The method for inverting chlorophyll a concentration in water bodies based on hyperspectral remote sensing images according to claim 1, characterized in that: Compare all the unified scoring function values ​​and select the feature matrix and algorithm pairing combination with the highest score as the optimal feature-model combination, including: Establishing a mapping relationship between the unified scoring function value and the corresponding combination number; Identifying the combination number having the highest unified scoring function value; The corresponding feature construction method and machine learning model algorithm are determined according to the combination number with the highest unified scoring function value as the optimal feature-model combination for the current task.

10. The method for inverting chlorophyll a concentration in water bodies based on hyperspectral remote sensing images according to claim 1, characterized in that: The water pixel subset is input into the optimal feature-model combination to perform concentration prediction processing to generate a chlorophyll a concentration spatial distribution map and a supporting quality control layer, including: Based on the optimal feature construction method in the optimal feature-model combination, feature conversion is performed on the hyperspectral reflectance vector of each pixel in the water body pixel subset to generate a corresponding pixel feature data matrix; Inputting the pixel feature data matrix into the optimal machine learning model in the optimal feature-model combination, performing concentration prediction, and obtaining the predicted chlorophyll a concentration values ​​of all water body pixels; Reconstructing the predicted chlorophyll a concentration values ​​of all water body pixels according to their spatial positions in the water body pixel subset to generate the chlorophyll a concentration spatial distribution map having spatial consistency with the original hyperspectral remote sensing image; The feature integrity status, model response confidence level, and cloud cover mark information of each pixel during the prediction process are synchronously recorded to generate the corresponding quality control layer.

Images