High-sediment-concentration remote sensing inversion method and device based on sentinel-2 image

Abstract

The application discloses a kind of high sediment concentration remote sensing inversion method and device of suspended load based on Sentinel-2 image, first, the satellite multispectral data and measured sediment concentration data of target water area are collected, and are preprocessed;Then the initial spectral feature set including the band feature combination of visible light-short wave infrared range is constructed, then the cross-validation recursive feature elimination algorithm (RFECV) is used to carry out preliminary screening from initial feature library, then according to spectral sensitivity analysis, the optimal feature subset is obtained by fine screening;Again, using random forest (RF) model is based on optimal spectral feature and sediment concentration data set training, constructs suspended load sediment concentration inversion model;Finally, the model is applied to realize the remote sensing inversion of high sediment concentration of suspended load, generates the sediment concentration time series of target water area.The present application effectively overcomes the common spectral saturation problem under high sediment concentration condition, can realize high-precision inversion in 0-45 kg / m³ concentration range, significantly improves the robustness and accuracy of suspended load high sediment concentration remote sensing inversion.

Description

Technical Field

[0001] This invention relates to the field of remote sensing inversion technology for water sediment content, and more specifically, to a method and apparatus for remote sensing inversion of high suspended sediment content based on Sentinel-2 imagery. Background Technology

[0002] Suspended sediment concentration is an important indicator for measuring the intensity of sediment transport in rivers, and it is crucial for river channel evolution, reservoir siltation, aquatic ecological environment, and coastal zone management. Traditional sediment concentration measurement usually relies on manual field sampling and laboratory analysis. Although it can obtain relatively accurate point data, it suffers from problems such as low efficiency, high cost, and limited spatial representativeness. It is particularly difficult to implement during flood season or under high sediment concentration conditions, making it difficult to achieve large-scale, high-frequency sediment monitoring.

[0003] Satellite remote sensing technology has been widely used for monitoring water sediment concentration due to its advantages such as wide coverage, short revisit cycles, and relatively low cost. Existing remote sensing inversion methods mainly include theoretical analysis, empirical modeling, and machine learning. Theoretical analysis is based on radiative transfer theory, with a clear physical mechanism, but it is computationally complex and requires high data quality. Empirical modeling achieves inversion by establishing a statistical relationship between spectral reflectance and sediment concentration, which is simple and easy to implement but has poor universality. Machine learning can uncover the nonlinear relationship between spectrum and sediment concentration, significantly improving inversion accuracy, but it still has the following problems under high sediment concentration conditions: 1. Spectral saturation phenomenon: When the sand content increases (e.g., exceeding 5 kg / m³), the reflectance of most bands tends to stabilize, resulting in a significant decrease in the inversion accuracy of traditional models; 2. Difficulty in feature selection: The spectra of water bodies with high sediment content exhibit complex changes such as "red shift," making single spectral features no longer applicable in high sediment content ranges. Furthermore, manual feature selection is insufficient to comprehensively capture sensitive information across different concentration ranges. 3. Weak model generalization ability: Existing models are mostly designed for medium to low sediment content scenarios, and have poor adaptability to coarse-grained sediment with a sediment content of more than 20 kg / m³. Summary of the Invention

[0004] The purpose of this invention is to provide a remote sensing inversion method for suspended sediment with high sediment content based on Sentinel-2 imagery and RFECV-RF algorithm, so as to solve the problems of low inversion accuracy, difficulty in feature selection and insufficient generalization ability of existing technologies in high sediment content ranges.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: The first aspect provides a remote sensing inversion method for high suspended sediment content based on Sentinel-2 imagery, including: Collect Sentinel-2 satellite multispectral remote sensing image data of the target water area and on-site measured suspended sediment concentration data for the corresponding period, and preprocess the satellite multispectral data; An initial spectral feature set is constructed based on preprocessed satellite multispectral data, wherein the initial spectral feature set includes a combination of band features in the visible light to shortwave infrared range; A two-stage adaptive screening framework is used for feature screening, including: applying a cross-validation recursive feature elimination algorithm to evaluate the feature importance and perform initial screening on the initial spectral feature set, and then performing fine screening based on the feature-suspended mass sensitivity test to obtain the optimal spectral features; The dataset consisting of the optimal spectral features and the corresponding suspended sediment concentration is used as input to construct and train a random forest regression model, thereby obtaining a trained remote sensing inversion model for suspended sediment concentration. Using a trained remote sensing inversion model for suspended sediment concentration, the suspended sediment concentration of the target water body is inverted to generate a time series of sediment concentration.

[0006] In one embodiment, preprocessing the satellite multispectral data includes: Atmospheric correction, radiometric calibration, band superposition, cloud masking, and water body extraction were performed on the satellite multispectral data, and the reflectance data of the target water body was extracted in the study area. The extracted target water body reflectance data is spatiotemporally matched with the measured sediment concentration data; The isolated forest algorithm was used to detect and clean outliers in the matched dataset, and a reflectance-sand content sample dataset was constructed.

[0007] In one implementation, the atmospheric correction is performed using an atmospheric radiative transfer model; the water body extraction is achieved using a normalized water body index combined with a threshold segmentation method.

[0008] In one embodiment, the combination of band features in the visible-shortwave infrared range includes single-band reflectance features, band difference features, and band ratio features, wherein the single-band reflectance features include: surface reflectance of the B1, B2, B3, B4, B5, B6, B7, B8, B8A, B11, and B12 bands of the Sentinel-2 satellite. Band difference characteristics include: B12–B2, B11–B8; The band ratio characteristics include: B4 / B8, B3 / B8, B3 / B5, B7 / B5, B5 / B6, B8A / B6, and B11 / B12.

[0009] In one implementation, a cross-validation recursive feature elimination algorithm is used to evaluate and initially screen the initial spectral feature set, and then a fine screening is performed based on the feature-suspended mass sensitivity test to obtain the optimal spectral features, including: Using the random forest regression model as the base model, the importance scores of each feature are calculated; The number of features to be removed in each recursive step is set, and the performance of the model under different feature combinations is evaluated using a 3-fold cross-validation strategy. The coefficient of determination R² is used as the evaluation index. By recursively eliminating features whose importance does not meet the preset conditions, a globally optimal subset of candidate features is obtained. Based on the measured sediment concentration data, the suspended sediment sediment concentration samples were divided into low-value and high-value ranges according to the sediment concentration threshold. The goodness of fit and slope decay ratio between spectral characteristics and suspended sediment concentration are calculated. The goodness of fit is the coefficient of determination, and the slope decay ratio includes the changes in spectral characteristics relative to suspended sediment concentration in the high-value range and the changes in spectral characteristics relative to suspended sediment concentration in the low-value range. Set a goodness-of-fit threshold and a sensitivity threshold. If the goodness-of-fit of a certain feature is lower than the goodness-of-fit threshold and the slope decay ratio is lower than the sensitivity threshold, then remove the feature from the global optimal candidate feature subset.

[0010] In one implementation, the hyperparameters are set as follows during the training of the random forest regression model: the maximum depth of the decision tree is 15, the minimum number of samples per leaf node is 2, the number of decision trees in the forest is 150, and the bootstrap sampling ratio is 0.9.

[0011] In one implementation, after constructing and training a random forest regression model using the dataset consisting of the optimal spectral features and the corresponding suspended sediment concentration as input, the method further includes a model accuracy verification step. Specifically, the reflectance-suspended sediment concentration sample dataset is divided into a training set and a validation set using a stratified sampling method at a ratio of 7:3. The validation set is used to calculate the coefficient of determination R², root mean square error, and mean absolute error of the model inversion results. If the coefficient of determination R², root mean square error, and mean absolute error all meet preset conditions, the model accuracy is determined to meet the inversion requirements.

[0012] Based on the same inventive concept, a second aspect of the present invention provides a remote sensing inversion device for high suspended sediment content based on Sentinel-2 imagery, comprising: The data collection and preprocessing module is used to collect Sentinel-2 satellite multispectral remote sensing image data of the target water area and the corresponding field measured suspended sediment concentration data, and to preprocess the satellite multispectral data. An initial spectral feature set construction module is used to construct an initial spectral feature set based on preprocessed satellite multispectral data, wherein the initial spectral feature set includes a combination of band features in the visible light to shortwave infrared range; The feature selection module is used to perform feature selection using a two-stage adaptive selection framework, including: applying a cross-validation recursive feature elimination algorithm to evaluate the feature importance and perform initial screening on the initial spectral feature set, and then performing fine screening based on the feature-suspended mass sensitivity test to obtain the optimal spectral features; The model training module is used to take the dataset consisting of the optimal spectral features and the corresponding suspended sediment concentration as input, construct and train a random forest regression model, and obtain a trained remote sensing inversion model of suspended sediment concentration. The sediment concentration inversion module is used to invert the suspended sediment concentration of a target water body using a trained remote sensing inversion model for suspended sediment concentration, and generate a sediment concentration time series.

[0013] Based on the same inventive concept, a third aspect of the present invention provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, provides the method for remote sensing inversion of suspended sediment high sediment content based on Sentinel-2 imagery as described in the first aspect.

[0014] Based on the same inventive concept, the fourth aspect of the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, it implements the remote sensing inversion method for suspended sediment high sediment content based on Sentinel-2 imagery described in the first aspect.

[0015] Compared with the prior art, the advantages and beneficial technical effects of the present invention are as follows: (1) The inversion range is large and wide. By selecting the band feature combination of the visible light-shortwave infrared range as the initial feature library in the spectral feature engineering, it contains feature bands that are sensitive to different sediment concentration ranges. Some interference is offset by band difference, ratio, etc. Combined with the two-stage adaptive screening framework, it can automatically screen out globally sensitive and important spectral feature combinations according to the sediment concentration range in the sample. It is suitable for the inversion of suspended sediment concentration with a large range and wide range. It can solve the problem that the selected spectral features cannot be applied in high sediment concentration rivers after saturation.

[0016] (2) High inversion accuracy. Through a two-stage adaptive screening framework, important spectral features are initially screened out by RFECV, and then the optimal spectral features that remain sensitive in the global sample are finely screened out by the sensitivity threshold. It effectively captures sensitive spectral information in the high sediment content range, without the need to manually search for sensitive features. It can reflect the different spectral absorption and scattering of river water under different sediment content conditions. The established model is suitable for sediment content inversion in different concentration ranges.

[0017] (3) Good generalization ability: The model is applicable to rivers with different sediment properties, especially to coarse-grained rivers with high sediment content such as the Yellow River. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of the remote sensing inversion method for high suspended sediment content based on Sentinel-2 imagery in an embodiment of the present invention; Figure 2 This is a detailed flowchart illustrating the implementation of the remote sensing inversion method for high suspended sediment content based on Sentinel-2 imagery in this embodiment of the invention. Figure 3 This is a schematic diagram showing the order of importance of key features in an embodiment of the present invention; Figure 4 This is a schematic diagram showing the goodness of fit and slope attenuation ratio of key features and measured suspended sediment concentration in an embodiment of the present invention. Figure 5 This is a schematic diagram illustrating the model performance evaluation in an embodiment of the present invention. Detailed Implementation

[0020] Example 1 This embodiment provides a remote sensing inversion method for high suspended sediment content based on Sentinel-2 imagery. Please refer to [link to relevant documentation]. Figure 1 ,include: S1: Collect Sentinel-2 satellite multispectral remote sensing image data of the target water area and the corresponding field measured suspended sediment concentration data, and preprocess the satellite multispectral data.

[0021] Specifically, S1 can be implemented in the following way: S1.1: Perform atmospheric correction, radiometric calibration, band superposition, cloud masking, and water body extraction on the satellite multispectral data, and extract the reflectance data of the target water body in the study area; S1.2: Spatiotemporal matching of the extracted target water body reflectance data with the measured sediment concentration data; S1.3: The isolated forest algorithm is used to detect and clean outliers in the matched dataset to construct a reflectance-sand content sample dataset.

[0022] Specifically, atmospheric correction is performed using an atmospheric radiative transfer model; water body extraction is achieved using a normalized water index combined with a threshold segmentation method.

[0023] In the specific implementation process, Sentinel-2 surface reflectance products were obtained and orthorectification and atmospheric correction were completed; cloud masking and modified normalized water index (MNDWI) were applied for water body extraction; the measured data and image data were spatiotemporally matched, with the spatial deviation controlled within 30 meters and the time window being 1 hour; the isolated forest algorithm was used to remove outlier samples, and the outlier ratio parameter was set to 0.05.

[0024] S2: Construct an initial spectral feature set based on the preprocessed satellite multispectral data, wherein the initial spectral feature set includes a combination of band features in the visible light to shortwave infrared range.

[0025] Specifically, based on the wavelength characteristics of Sentinel-2 multispectral image data, B1-B12 covers single-wavelength spectra of visible light to short-wave infrared, namely visible light (B1-B4), red edge (B5-B7), near-infrared (B8, B8A), and short-wave infrared (B11, B12). Band difference and ratio are combined features obtained by subtracting and dividing the features of different single bands. The combined features of band difference and ratio can cancel the interference of the environment on the spectral reflectance value. Therefore, this scheme sets the band feature combination in the visible light to short-wave infrared range to include single-band reflectance features, band difference features, and band ratio features.

[0026] The initial spectral feature set constructed in this scheme covers sensitive spectral features in different sediment concentration ranges: under low sediment concentration, the visible and near-infrared bands are more sensitive, while under high sediment concentration, the near-infrared, red edge, and short-wave infrared bands are more sensitive. This scheme selects single-band spectral features in the visible-short-wave infrared range, as well as the combined features B4 / B8, B3 / B8, and B3 / B5 dominated by visible light, the combined features B7 / B5, B5 / B6, and B8A / B6 dominated by near-infrared and red edge, and the combined features B12–B2, B11–B8, and B11 / B12 dominated by short-wave infrared features as initial spectral features.

[0027] In the specific implementation process, the initial spectral feature set in S2 has a total of 20 dimensions, including 11 single bands: B1 (443 nm), B2 (496 nm), B3 (560 nm), B4 (664 nm), B5 (703 nm), B6 ​​(740 nm), B7 (782 nm), B8 (835 nm), B8A (864 nm), B11 (1613 nm), and B12 (2202 nm); 2 band differences: B12–B2 and B11–B8; and 7 band ratios: B4 / B8, B3 / B8, B3 / B5, B7 / B5, B5 / B6, B8A / B6, and B11 / B12.

[0028] S3: A two-stage adaptive screening framework is used for feature screening, including: applying a cross-validation recursive feature elimination algorithm to evaluate the feature importance and perform initial screening on the initial spectral feature set, and then performing fine screening based on the feature-suspended mass sensitivity test to obtain the optimal spectral features.

[0029] Specifically, S3 includes: S3.1: Using the random forest regression model as the base model, calculate the importance score of each feature; S3.2: Set the number of features to be removed in each recursive step, and use a 3-fold cross-validation strategy to evaluate the model performance under different feature combinations, with the coefficient of determination R² as the evaluation index. S3.3: By recursively eliminating features whose importance does not meet the preset conditions, a globally optimal subset of candidate features is obtained; S3.4: Based on the measured sediment concentration data, the suspended sediment sediment concentration samples are divided into low-value intervals and high-value intervals according to the sediment concentration threshold; S3.5: Calculate the goodness of fit and slope decay ratio between the spectral characteristics and suspended sediment concentration, where the goodness of fit is the coefficient of determination, and the slope decay ratio includes the changes in the relative suspended sediment concentration of the spectral characteristics in the high-value range and the changes in the relative suspended sediment concentration of the spectral characteristics in the low-value range. S3.6: Set a goodness-of-fit threshold and a sensitivity threshold. If the goodness-of-fit of a certain feature is lower than the goodness-of-fit threshold and the slope decay ratio is lower than the sensitivity threshold, then remove the feature from the global optimal candidate feature subset.

[0030] In the specific implementation process, the RFECV screening process adopts a two-stage adaptive screening framework. In the first stage (including S3.1 to S3.3), RFECV is used to screen the globally optimal candidate feature set from the perspective of prediction ability. Random forest is used as the base model, and 3-fold cross-validation is set. In each round, the five least important features are removed. The validation set R² is used as the evaluation index to perform the initial screening of the spectral feature library. In the second stage (including S3.4 to S3.6), the goodness of fit and sensitivity between the spectral features screened in the initial stage and the suspended sediment concentration are quantified. After verification and analysis, features with poor goodness of fit and obvious insensitivity of the spectrum are removed.

[0031] Regarding the slope, for the high-value range, it is the relative change value obtained by dividing the relative change value of the spectral characteristics by the suspended sediment content; for the low-value range, it is the relative change value obtained by dividing the relative change value of the spectral characteristics by the suspended sediment content. Regarding the slope attenuation ratio, it is the comparison between the slope of the high-value range and the slope of the low-value range.

[0032] The adaptive filtering framework used in this invention differs from existing feature filtering methods as follows: The existing feature selection method first performs spectral feature sensitivity analysis. By measuring the correlation between spectral features and suspended sediment concentration, some sensitive features are used as initial spectral features. Then, cross-validation is performed to recursively select the optimal subset. However, this method may select highly collinear sensitive features. Therefore, some features in the feature library are actually invalid. The optimal subset combination cannot automatically adapt to different suspended sediment concentration ranges (especially high sediment concentration), which ultimately leads to low accuracy in the correlation with suspended sediment concentration.

[0033] The method of this invention first sets up an initial feature set containing a combination of spectral features sensitive to a wide range of high concentrations. After two-stage screening (recursive deletion for initial screening and sensitivity analysis for fine screening), the optimal feature subset that best fits the suspended sediment concentration range of the river is obtained. The method automatically selects the sensitive feature combination within the range of suspended sediment concentration of the river to be inverted. Sensitivity can reflect whether the spectral features have reached saturation in the global range of suspended sediment concentration, while sensitivity reflects the average level of correlation between the spectral features and the suspended sediment concentration. Therefore, this method can be applied to rivers with high sediment concentration, while existing methods have the problem of high concentration inversion failure in the inversion of rivers with high sediment concentration.

[0034] S4: Using the dataset consisting of the optimal spectral features and the corresponding suspended sediment concentration as input, construct and train a random forest regression model to obtain a trained remote sensing inversion model for suspended sediment concentration.

[0035] During the training of the random forest regression model, the hyperparameters were set as follows: the maximum depth of the decision tree was 15, the minimum number of samples per leaf node was 2, the number of decision trees in the forest was 150, the bootstrap sampling ratio was 0.9, and the parameters were optimized using grid search. The model performance was evaluated using R², RMSE, and MAE.

[0036] S5: Using the trained remote sensing inversion model for suspended sediment concentration, the suspended sediment concentration of the target water body is inverted to generate a time series of sediment concentration.

[0037] Specifically, S5 includes: preprocessing and feature construction of the target image, extracting the optimal feature subset and inputting it into the trained RF model, and outputting the sand content inversion result map with the same spatial resolution as the original image.

[0038] Please see Figure 2 This is a detailed flowchart of the implementation of the remote sensing inversion method for high suspended sediment content based on Sentinel-2 imagery in this embodiment of the invention.

[0039] SSC represents suspended sediment concentration data. Points are extracted from Sentinel-2 multispectral images to obtain multispectral data of the measurement points. Pure water bodies are extracted from the Sentinel-2 multispectral images to obtain multispectral images of pure water bodies in the river section.

[0040] In one implementation, after S5, the method further includes a model accuracy verification step S6, which involves dividing the reflectance-suspended sediment concentration sample dataset into a training set and a validation set using a stratified sampling method at a ratio of 7:3; using the validation set to calculate the coefficient of determination R², root mean square error, and mean absolute error of the model inversion results; and if the coefficient of determination R², root mean square error, and mean absolute error all meet preset conditions, then the model accuracy is determined to meet the inversion requirements.

[0041] The following section uses the Wubu Station and Shizuishan Station on the Yellow River as examples to describe the specific implementation of this invention: 1. Data Collection and Preprocessing: Daily average sediment concentration data and concurrent Sentinel-2 images from Wubu and Shizuishan stations were acquired from 2020 to 2024. After cloud masking, MNDWI water body extraction, spatiotemporal matching, and outlier removal, 256 and 267 valid samples were obtained from Wubu and Shizuishan stations, respectively, with sediment concentration data ranging from 0.07 to 44.5 kg / m³. 3 0.21-8.28 kg / m 3 .

[0042] 2. Feature Construction: Constructing 20-dimensional initial features based on 11 bands; 3. Feature Screening: First, the data samples were divided into categories based on their weight (below 2.5 kg / m³). 3and above 2.5 kg / m 3 The interval; then a two-stage feature selection is performed: (1) Using random forest as the base model (parameters as before), features were recursively eliminated through 3-fold cross-validation to obtain the candidate feature subset with the highest cross-validation score. Five features were selected for Wubu station (B7 / B5, B5 / B6, B8A, B11 / B12, B4 / B8), and seven features were selected for Shizuishan station (B8-B11, B3 / B8, B4 / B8, B7 / B5, B8A, B3 / B5, B8A / B6). Due to the higher sand concentration at Wubu station, the number of features selected was less due to spectral saturation. However, the B8A band was an important spectral feature at both stations (see Figure 3 ); (2) The key features were fitted to the measured suspended sediment concentration using power function, exponential function, and polynomial methods, respectively, and the goodness of fit R was used. 2 The optimal model is used to represent the relationship between characteristics and suspended sediment concentration, and the slope attenuation ratio (SDR) is calculated (see [reference]). Figure 4 ), where R 2 Features with a SDR < 0.5 and an SDR < 0.01 were removed. Finally, features B3 / B5 and B4 / B8 were removed from Shizuishan Station, and combined features B8-B11, B3 / B8, B7 / B5, B8A, and B8A / B6 were used. Features B5 / B6 and B4 / B8 were removed from Wubu Station, and combined features B7 / B5, B8A, and B11 / B12 were used. 4. Model Training and Validation: The training and validation sets were divided in a 7:3 ratio to train the RF model. The results are as follows:

[0043] Therefore, the model constructed in this invention is applicable to sediment concentration inversion within different concentration ranges, and the inversion R² can reach above 0.8 in the range of 0-45 kg / m³ (see [reference]). Figure 5 ).

[0044] 5. Sediment concentration mapping: The trained model is used to generate time series maps of sediment concentration in the river sections near Shizuishan Station and Wubu Station, which accurately identifies periods of high sediment concentration and is more effective than traditional methods.

[0045] Example 2 Based on the same inventive concept, this embodiment discloses a remote sensing inversion device for high suspended sediment content based on Sentinel-2 imagery, comprising: The data collection and preprocessing module is used to collect Sentinel-2 satellite multispectral remote sensing image data of the target water area and the corresponding field measured suspended sediment concentration data, and to preprocess the satellite multispectral data. An initial spectral feature set construction module is used to construct an initial spectral feature set based on preprocessed satellite multispectral data, wherein the initial spectral feature set includes a combination of band features in the visible light to shortwave infrared range; The feature selection module is used to perform feature selection using a two-stage adaptive selection framework, including: applying a cross-validation recursive feature elimination algorithm to evaluate the feature importance and perform initial screening on the initial spectral feature set, and then performing fine screening based on the feature-suspended mass sensitivity test to obtain the optimal spectral features; The model training module is used to take the dataset consisting of the optimal spectral features and the corresponding suspended sediment concentration as input, construct and train a random forest regression model, and obtain a trained remote sensing inversion model of suspended sediment concentration. The sediment concentration inversion module is used to invert the suspended sediment concentration of a target water body using a trained remote sensing inversion model for suspended sediment concentration, and generate a sediment concentration time series.

[0046] Since the apparatus in Embodiment 2 of this invention is the same apparatus used in the remote sensing inversion method for high suspended sediment content based on Sentinel-2 imagery in Embodiment 1, those skilled in the art can understand the specific structure and variations of this apparatus based on the method described in Embodiment 1 of this invention, and therefore will not be repeated here. All apparatuses used in the method of Embodiment 1 of this invention fall within the scope of protection of this invention.

[0047] Example 3 Based on the same inventive concept, the present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the remote sensing inversion method for high suspended sediment content based on Sentinel-2 imagery of Embodiment 1.

[0048] Since the computer-readable storage medium described in Embodiment 3 of this invention is the same computer-readable storage medium used in implementing the remote sensing inversion method for high suspended sediment content based on Sentinel-2 imagery in Embodiment 1 of this invention, those skilled in the art can understand the specific structure and variations of this computer-readable storage medium based on the method described in Embodiment 1 of this invention, and therefore will not be repeated here. All computer-readable storage media used in the method of Embodiment 1 of this invention fall within the scope of protection of this invention.

[0049] Example 4 Based on the same inventive concept, the present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in Embodiment 1.

[0050] Since the computer equipment described in Embodiment 4 of this invention is the same computer equipment used in implementing the remote sensing inversion method for high suspended sediment content based on Sentinel-2 imagery in Embodiment 1 of this invention, those skilled in the art can understand the specific structure and variations of this computer equipment based on the method described in Embodiment 1 of this invention, and therefore will not be described again here. All computer equipment used in the method of Embodiment 1 of this invention falls within the scope of protection of this invention.

[0051] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0052] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0053] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various modifications and variations to the embodiments of the invention without departing from the spirit and scope of the invention. Thus, if these modifications and variations of the embodiments of the invention fall within the scope of the claims of the invention and their equivalents, the invention also intends to include these modifications and variations.

Claims

1. A remote sensing inversion method for high suspended sediment content based on Sentinel-2 imagery, characterized in that, include: Collect Sentinel-2 satellite multispectral remote sensing image data of the target water area and on-site measured suspended sediment concentration data for the corresponding period, and preprocess the satellite multispectral data; An initial spectral feature set is constructed based on preprocessed satellite multispectral data, wherein the initial spectral feature set includes a combination of band features in the visible light to shortwave infrared range; A two-stage adaptive screening framework is used for feature screening, including: applying a cross-validation recursive feature elimination algorithm to evaluate the feature importance and perform initial screening on the initial spectral feature set, and then performing fine screening based on the feature-suspended mass sensitivity test to obtain the optimal spectral features; The dataset consisting of the optimal spectral features and the corresponding suspended sediment concentration is used as input to construct and train a random forest regression model, thereby obtaining a trained remote sensing inversion model for suspended sediment concentration. Using a trained remote sensing inversion model for suspended sediment concentration, the suspended sediment concentration of the target water body is inverted to generate a time series of sediment concentration.

2. The remote sensing inversion method for high sediment content in suspended sediment based on Sentinel-2 imagery as described in claim 1, characterized in that, Preprocessing of the satellite multispectral data includes: Atmospheric correction, radiometric calibration, band superposition, cloud masking, and water body extraction were performed on the satellite multispectral data, and the reflectance data of the target water body was extracted in the study area. The extracted target water body reflectance data is spatiotemporally matched with the measured sediment concentration data; The isolated forest algorithm was used to detect and clean outliers in the matched dataset, and a reflectance-sand content sample dataset was constructed.

3. The remote sensing inversion method for high suspended sediment content based on Sentinel-2 imagery as described in claim 2, characterized in that, The atmospheric correction is performed using an atmospheric radiative transfer model; the water body extraction is achieved using a normalized water body index combined with a threshold segmentation method.

4. The remote sensing inversion method for high sediment content in suspended sediment based on Sentinel-2 imagery as described in claim 1, characterized in that, The combination of band characteristics in the visible-shortwave infrared range includes single-band reflectance characteristics, band difference characteristics, and band ratio characteristics. Among them, the single-band reflectance characteristics include the surface reflectance of the B1, B2, B3, B4, B5, B6, B7, B8, B8A, B11, and B12 bands of the Sentinel-2 satellite. Band difference characteristics include: B12–B2, B11–B8; The band ratio characteristics include: B4 / B8, B3 / B8, B3 / B5, B7 / B5, B5 / B6, B8A / B6, and B11 / B12.

5. The remote sensing inversion method for high sediment content in suspended sediment based on Sentinel-2 imagery as described in claim 1, characterized in that, A cross-validation recursive feature elimination algorithm is used to evaluate and initially screen the initial spectral feature set, and then a fine screening is performed based on the feature-suspended mass sensitivity test to obtain the optimal spectral features, including: Using the random forest regression model as the base model, the importance scores of each feature are calculated; The number of features to be removed in each recursive step is set, and the performance of the model under different feature combinations is evaluated using a 3-fold cross-validation strategy. The coefficient of determination R² is used as the evaluation index. By recursively eliminating features whose importance does not meet the preset conditions, a globally optimal subset of candidate features is obtained. Based on the measured sediment concentration data, the suspended sediment sediment concentration samples were divided into low-value and high-value ranges according to the sediment concentration threshold. The goodness of fit and slope decay ratio between spectral characteristics and suspended sediment concentration are calculated. The goodness of fit is the coefficient of determination, and the slope decay ratio includes the changes in spectral characteristics relative to suspended sediment concentration in the high-value range and the changes in spectral characteristics relative to suspended sediment concentration in the low-value range. Set a goodness-of-fit threshold and a sensitivity threshold. If the goodness-of-fit of a certain feature is lower than the goodness-of-fit threshold and the slope decay ratio is lower than the sensitivity threshold, then remove the feature from the global optimal candidate feature subset.

6. The remote sensing inversion method for high suspended sediment content based on Sentinel-2 imagery as described in claim 5, characterized in that, During the training of the random forest regression model, the hyperparameters were set as follows: the maximum depth of the decision tree was 15, the minimum number of samples per leaf node was 2, the number of decision trees in the forest was 150, and the bootstrap sampling ratio was 0.

9.

7. The remote sensing inversion method for high sediment content in suspended sediment based on Sentinel-2 imagery as described in claim 2, characterized in that, After constructing and training a random forest regression model by using the dataset consisting of the optimal spectral features and the corresponding suspended sediment concentration as input, the method further includes a model accuracy verification step. Specifically, the reflectance-suspended sediment concentration sample dataset is divided into a training set and a validation set in a 7:3 ratio using a stratified sampling method. The validation set is used to calculate the coefficient of determination R², root mean square error, and mean absolute error of the model inversion results. If the coefficient of determination R², root mean square error, and mean absolute error all meet the preset conditions, the model accuracy is determined to meet the inversion requirements.

8. A remote sensing inversion device for high suspended sediment content based on Sentinel-2 imagery, characterized in that, include: The data collection and preprocessing module is used to collect Sentinel-2 satellite multispectral remote sensing image data of the target water area and the corresponding field measured suspended sediment concentration data, and to preprocess the satellite multispectral data. An initial spectral feature set construction module is used to construct an initial spectral feature set based on preprocessed satellite multispectral data, wherein the initial spectral feature set includes a combination of band features in the visible light to shortwave infrared range; The feature selection module is used to perform feature selection using a two-stage adaptive selection framework, including: applying a cross-validation recursive feature elimination algorithm to evaluate the feature importance and perform initial screening on the initial spectral feature set, and then performing fine screening based on the feature-suspended mass sensitivity test to obtain the optimal spectral features; The model training module is used to take the dataset consisting of the optimal spectral features and the corresponding suspended sediment concentration as input, construct and train a random forest regression model, and obtain a trained remote sensing inversion model of suspended sediment concentration. The sediment concentration inversion module is used to invert the suspended sediment concentration of a target water body using a trained remote sensing inversion model for suspended sediment concentration, and generate a sediment concentration time series.

9. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the remote sensing inversion method for high suspended sediment content based on Sentinel-2 imagery as described in any one of claims 1 to 7.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the remote sensing inversion method for suspended high-sand content based on Sentinel-2 imagery as described in any one of claims 1 to 7.

Images