Method for monitoring suspended sediment in optical complex water body based on man-machine cooperation

Abstract

The invention provides an optical complex water body suspended sediment monitoring method based on man-machine cooperation, and belongs to the technical field of remote sensing water environment monitoring. The method comprises the following steps: acquiring a multispectral remote sensing image of a target area, preprocessing the multispectral remote sensing image, and performing space-time fusion to obtain a spectral data set with high space-time resolution; carrying out manual labeling on the spectrum data set, constructing a suspended sediment concentration classification data set, and carrying out synthesis expansion on a minority class of samples to generate an equalization training set; training a suspended sediment concentration inversion model by using the equalized training set to obtain an initial suspended sediment concentration predicted value, and dynamically optimizing the spectral feature weight and the classification boundary of the suspended sediment concentration inversion model based on expert knowledge; and locally correcting the initial suspended sediment concentration predicted value and diffusing the initial suspended sediment concentration predicted value to a global range based on a hydrodynamic law and a space consistency constraint to obtain a global suspended sediment concentration predicted value. The method combines remote sensing image analysis, machine learning and expert knowledge to realize high-precision inversion of SSC.

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing water environment monitoring, and particularly relates to a method for monitoring suspended sediment in optically complex waters based on human-machine collaboration. Background Art

[0002] The global coastal suspended sediment concentration (SSC) plays a crucial role in regulating erosion, sedimentation, and geomorphic evolution processes. However, limited by observation means, current large-scale and high-resolution SSC monitoring data are still scarce, which greatly restricts the research on coastal evolution mechanisms, especially quantitative analysis based on empirical models.

[0003] Existing suspended sediment monitoring methods mainly include in-situ sampling, acoustic remote sensing, and optical remote sensing. Although the in-situ sampling method has high accuracy, it is limited by high costs and limited spatial coverage, making it difficult to meet the monitoring requirements for large areas and long time series. Acoustic remote sensing technologies (such as Doppler velocity profilers and multibeam sonars) are affected by sediment scattering and absorption effects in high-sediment-laden waters, resulting in a decrease in measurement accuracy, and the equipment is expensive and the operation is complex. Optical remote sensing technology uses the spectral information of satellite images to invert SSC, which has the advantages of wide spatial coverage and fast data acquisition, and has become an important means for SSC monitoring. However, in optically complex waters (such as high-sediment-laden, organic-rich, or algal waters), traditional optical remote sensing inversion models face the following challenges: 1) The strong scattering and absorption effects of suspended sediment interfere with the water reflectance characteristics, leading to an increase in SSC inversion error. 2) There are significant differences in water optical parameters (such as absorption coefficient and backscattering coefficient) in different sea areas, and traditional global empirical models are difficult to adapt to the environmental characteristics of different regions. 3) SSC inversion methods based on a single satellite sensor or empirical formula are difficult to ensure accuracy in cases of scarce data or complex water conditions, especially when there is a lack of sufficient in-situ data for model calibration.

[0004] In view of the above problems, the adaptability of traditional SSC monitoring methods in large-scale, data-scarce, and optically complex environments is low, and it is difficult to meet the requirements for global SSC dynamic change monitoring. Therefore, this invention is specifically proposed. Summary of the Invention

[0005] Aiming at the problem of scarce large-scale and high-resolution suspended sediment concentration monitoring data, the present invention proposes a method for monitoring suspended sediment in optically complex waters based on human-machine collaboration, which combines remote sensing image analysis, machine learning, and expert knowledge to achieve high-precision inversion of SSC.

[0006] The present invention provides a method for monitoring suspended sediment in optically complex waters based on human-machine collaboration, the method comprising: Collect multi-spectral remote sensing images of the target area and perform preprocessing. Perform spatio-temporal fusion on the preprocessed multi-spectral remote sensing images to obtain a spectral dataset with high spatio-temporal resolution; Manually annotate the spectral dataset to construct a suspended sediment concentration classification dataset, and perform synthetic expansion on minority class samples based on hydrodynamic similarity constraints to generate a balanced training set; Use the balanced training set to train a suspended sediment concentration inversion model to obtain an initial predicted value of the suspended sediment concentration, and dynamically optimize the spectral feature weights and classification boundaries of the suspended sediment concentration inversion model based on expert knowledge, where the suspended sediment concentration inversion model uses a regularization machine learning classifier model; Based on hydrodynamic laws and spatial consistency constraints, locally correct the initial predicted value of the suspended sediment concentration and spread it to the global range to obtain a global predicted value of the suspended sediment concentration.

[0007] In an embodiment of the present invention, performing spatio-temporal fusion on the preprocessed multi-spectral remote sensing images to obtain a spectral dataset with high spatio-temporal resolution specifically includes: Adopt an improved object-based spatio-temporal adaptive reflectance fusion model to fuse the preprocessed multi-spectral remote sensing images to generate a fused image with both high spatial resolution and high temporal resolution. Among them, the calculation formula of the spatio-temporal adaptive reflectance fusion model is: , Among them, represents the fused image; is the high spatial resolution Sentinel image; and respectively represent the surface reflectance of the MODIS image at times t and t−1. The multi-spectral remote sensing images include Sentinel images and MODIS images; Based on the fused image, extract multi-band spectral features to obtain the spectral dataset with high spatio-temporal resolution.

[0008] In an embodiment of the present invention, manually annotating the spectral dataset to construct a suspended sediment concentration classification dataset includes: Through a grid annotation interface, experts visually classify the spectral dataset and divide the suspended sediment concentration patterns into three categories: clear, semi-transparent, and turbid; Combine spectral-spatial features to divide the suspended sediment concentration categories into different groups to construct a suspended sediment concentration classification dataset.

[0009] In an embodiment of the present invention, a variant of SMOTE guided by experts is adopted to synthetically expand minority class samples based on hydrodynamic similarity constraints, including: For each minority class sample in the suspended sediment concentration classification dataset, k nearest neighbor samples to this minority class sample are identified within a hydrodynamically similar region, and the feature difference between this minority class sample and the nearest neighbor samples is calculated, which is expressed as: , where, and respectively represent the feature vectors of minority class sample i and nearest neighbor sample j, and Δd is the interpolation difference.

[0010] An interpolation coefficient is randomly generated, and based on this interpolation coefficient and the interpolation difference, a new sample and the feature vector of this new sample are calculated, where: The new sample is expressed as: i + θ×Δd; The feature vector of the new sample is expressed as: ; where θ is the interpolation coefficient, θ ∈ (0, 1), and is used to control the interpolation position of the feature vector of the new sample between and

[0011] In an embodiment of the present invention, based on expert knowledge, the spectral feature weights and classification boundaries of the suspended sediment concentration inversion model are dynamically optimized, including: Several key spectral features in the balanced training set are screened out through spectral sensitivity analysis, and based on expert knowledge, these several key spectral features are sorted according to importance, and spectral features that may cause misclassification are identified in combination with hydrodynamic laws, and the weights of the corresponding spectral features are adjusted; Expert collaborative annotation is introduced in the water body category boundary region to correct the classification boundary of the suspended sediment concentration inversion model.

[0012] In an embodiment of the present invention, based on hydrodynamic laws and spatial consistency constraints, local correction is performed on the initial suspended sediment concentration prediction value to obtain a local suspended sediment concentration prediction value, including: The initial suspended sediment concentration prediction value is traversed with a sliding window of a given size, and the sediment transport threshold within each window is calculated based on the sediment transport principle, and outliers exceeding this threshold are detected; An expert manually corrects the outliers according to hydrodynamic laws and the measured values of suspended sediment concentration, and adjusts the outliers to conform to hydrodynamic laws; Mark the adjusted data as high-confidence reference data and re-enter it into the suspended sediment concentration inversion model for parameter optimization; ; Introduce spatial consistency constraints into the loss function of the suspended sediment concentration inversion model to control the predicted values of local suspended sediment concentration to satisfy the spatial consistency constraints within the neighborhood range.

[0013] In an embodiment of the present invention, during the process of locally correcting the initial predicted value of the suspended sediment concentration, it further includes: Perform numerical transformation processing on the balanced training set, including dynamically filtering and removing or correcting low-confidence data, adjusting the range of data using normalization or standardization methods, and / or denoising the existing noise; Mark the processed balanced training set as high-confidence reference data and re-enter it into the suspended sediment concentration inversion model for parameter optimization; Introduce residual analysis to correct the error between the predicted value of the local suspended sediment concentration and the measured value of the suspended sediment concentration; Establish a dynamic feedback mechanism to update and correct the high-confidence reference data; Add newly acquired multi-spectral remote sensing images to update the balanced training set.

[0014] In an embodiment of the present invention, the predicted value of the local suspended sediment concentration is expressed as: , wherein, represents the predicted value of the local suspended sediment concentration; α is the weight adjustment coefficient; is the measured value of the suspended sediment concentration; fi(x) is the optimized suspended sediment concentration inversion model, and x is the feature vector of the i-th sample in the balanced training set.

[0015] In an embodiment of the present invention, an adaptive diffusion mechanism is adopted to spread the local correction to the global range, including: Introduce a dynamic balance factor to control the influence weight of the local correction on the global range, and obtain the predicted value of the global suspended sediment concentration, which is expressed as: , wherein, is the predicted value of the global suspended sediment concentration, is the initial predicted value of the suspended sediment concentration , is the predicted value of the local suspended sediment concentration, and P is the dynamic balance factor; Adopt a statistical consistency optimization method. By calculating the mean, standard deviation, and divergence of the data before and after correction, verify the rationality of the data distribution, and adjust the dynamic balance factor in combination with the chi-square test. Introduce global consistency constraints into the model loss function, and dynamically adjust hyperparameters in combination with the gradient descent optimization method to meet the hydrodynamic laws and spatial consistency constraints.

[0016] In an embodiment of the present invention, it further includes: verifying the performance of the suspended sediment concentration inversion model, including: Use Taylor diagrams, scatter plots, and violin plots to visualize the global suspended sediment concentration prediction values, and evaluate the performance of the suspended sediment concentration inversion model. Calculate the NSE efficiency to evaluate the fitting ability of the suspended sediment concentration inversion model. Calculate the MK trend test to evaluate the significant changes in the global suspended sediment concentration prediction values in the time dimension. Calculate the conversion rate to quantify the concentration change rate of the global suspended sediment concentration prediction values between different seasonal changes. Evaluate the change characteristics of the global suspended sediment concentration prediction values with different seasonal changes through grid analysis.

[0017] As can be seen from the above solutions, the advantages of the present invention are: The method for monitoring suspended sediment in optically complex waters based on human-machine collaboration disclosed by the present invention. This method obtains a spectral dataset with high spatio-temporal resolution through collecting multi-spectral remote sensing images of the target area, preprocessing, and then performing spatio-temporal fusion; manually annotating the spectral dataset to construct a suspended sediment concentration classification dataset, and synthesizing and expanding minority class samples to generate a balanced training set; using the balanced training set to train a suspended sediment concentration inversion model to obtain initial suspended sediment concentration prediction values, and dynamically optimizing the spectral feature weights and classification boundaries of the suspended sediment concentration inversion model based on expert knowledge; based on hydrodynamic laws and spatial consistency constraints, locally correct the initial suspended sediment concentration prediction values and spread them to the global range to obtain global suspended sediment concentration prediction values. This method combines remote sensing image analysis, machine learning, and expert knowledge to achieve high-precision inversion of SSC. And this method iteratively optimizes the inversion results with minimal manual correction, thereby reducing the dependence on traditional on-site measurements, improving the monitoring accuracy and applicability. At the same time, the present invention follows the principle of human-machine interaction, embeds human intelligence in the SSC inversion, enhances the interpretability of the model, and ensures that the SSC inversion conforms to hydrodynamic laws. Description of the Drawings

[0018] Figure 1 is the overall flowchart of the method for monitoring suspended sediment in optically complex waters based on human-machine collaboration provided by the embodiment of the present invention. Figure 2 The figure shows a schematic diagram of the spatial distribution of the measured values and estimated values of the suspended sediment concentration obtained by implementing the method of the present invention in a specific example. Among them, Figure A shows the measured results of the suspended sediment concentration, and Figure B shows the predicted values of the suspended sediment concentration estimated by the inversion model. Figure 3 This is a performance comparison chart obtained by using the suspended sediment concentration inversion models of the present invention and the prior art for this example. Figure A uses a Taylor diagram to compare the model performance, and Figures B - E show the model scatter plots and NSE values. Among them, Figure B is for Model 1, Figure C is for Model 2, Figure D is for Model 3, and Figure E is for Model 4. Figure 4 This is a schematic diagram of the annual suspended sediment concentration distribution estimated by the suspended sediment concentration inversion model for this example from 2017 to 2024. Figure 5 This is a schematic diagram for evaluating the spatial heterogeneity of the inter - annual suspended sediment concentration characteristics for this example. Among them, Figure a shows the Mann - Kendall (MK) trend test Z - scores of the suspended sediment concentration in the coastal waters of China from 2017 to 2024. Figure b is an enlarged view showing the fused satellite images of Sentinel - 2A / MSI and MOD09A1 Terra surface reflectance products at the Yellow River Delta (A, 38°N) and the Yangtze River Estuary (B, 31°N) in 2017 and 2024, and the local Z - scores from 2017 to 2024 are overlaid on the satellite images of the Yellow River Delta and the Yangtze River Estuary in 2024. Figure 6 This is a schematic diagram for evaluating the seasonal heterogeneity of the suspended sediment concentration characteristics for this example. Among them, Figure A is the seasonal distribution of the suspended sediment concentration from 2017 to 2024, and Figure B is the SSC conversion rate between seasons from 2017 to 2024. Detailed implementation manners

[0019] To make the above - mentioned objects, features, and advantages of the present invention more obvious and understandable, the following will describe the detailed implementation manners of the present invention in conjunction with the accompanying drawings. Many specific details are set forth in the following description in order to fully understand the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific implementations disclosed below.

[0020] Aiming at the problem of scarce monitoring data of suspended sediment concentration at large scales and high resolutions, the present invention proposes an optical suspended sediment monitoring method for complex water bodies based on human-machine collaboration, which combines remote sensing image analysis, machine learning, and expert knowledge to achieve high-precision inversion of SSC. This method uses publicly available high-resolution satellite images to estimate based on the relationship between the backscattering coefficient and SSC, and iteratively optimizes the inversion results through minimal manual correction, thereby reducing the dependence on traditional on-site measurements and improving the monitoring accuracy and applicability. This method follows the principle of Human-in-the-Loop (HITL), embeds human intelligence in three key stages of SSC inversion, enhances the interpretability of the model, and ensures that the SSC inversion conforms to hydrodynamic laws. Specifically, please refer to Figure 1 as shown Figure 1 Fig. shows a schematic diagram of the overall process of an optical suspended sediment monitoring method for complex water bodies based on human-machine collaboration provided by an embodiment of the present invention.

[0021] The optical suspended sediment monitoring method for complex water bodies based on human-machine collaboration specifically includes the following steps: Step S1: Collect multi-spectral remote sensing images of the target area and perform preprocessing, and perform spatio-temporal fusion on the preprocessed multi-spectral remote sensing images to obtain a spectral data set with high spatio-temporal resolution.

[0022] In one embodiment, first collect several multi-spectral remote sensing images of the target area, such as Sentinel images and MODIS images. The Sentinel images include multiple bands such as blue, green, red, near-infrared, and short-wave infrared to obtain refined surface feature information; MODIS images, which have high temporal resolution and can be used to monitor short-term water body change trends. Ensure that the selected images cover the spatio-temporal range of the target area to ensure the integrity and continuity of the data. At the same time, collect the measured data of the target area, including the measured values of suspended sediment concentration at different location points, geographical coordinates (including longitude and latitude), etc., and record the sampling time, sampling location, and data quality information.

[0023] In one embodiment, it is further necessary to preprocess the multispectral remote sensing images. First, image quality control and screening are carried out. Cloud detection and removal algorithms are used to eliminate images with severe cloud contamination, ensuring that the quality of the image data meets the analysis requirements. The radiometric consistency of the images is evaluated, and data that does not meet the quality standards is replaced or eliminated. Then, radiometric correction, atmospheric correction, and geometric correction are performed on the remote sensing images to ensure the quality and reliability of the image data. Through radiometric correction, sensor noise is removed, and the actual radiance of the image is restored to ensure accurate and reliable radiometric values. Through the 6S model or other atmospheric correction methods, atmospheric interference factors such as aerosols and molecular scattering are removed to improve the accuracy of the surface reflectance of the image. Through geometric correction, orthorectification is performed based on known geographic coordinates to eliminate geometric distortions caused by sensor position and terrain changes, ensuring that the image is aligned with the actual geographical location.

[0024] Then, spatio-temporal fusion is performed on the preprocessed multispectral remote sensing images to obtain a spectral dataset with high spatio-temporal resolution, specifically including: using an improved object-based spatio-temporal adaptive reflectance fusion model (OS-STARFM) to fuse the preprocessed multispectral remote sensing images to generate a fused image with both high spatial resolution and high temporal resolution, improving the spatio-temporal continuity of surface information. Among them, the calculation formula of the spatio-temporal adaptive reflectance fusion model is: (1), where, represents the fused image; is the high-spatial-resolution Sentinel image; and respectively represent the surface reflectance of the MODIS image at times t and t−1. The multispectral remote sensing images include Sentinel images and MODIS images.

[0025] Based on the fused image, multi-band spectral features are extracted, including bands such as blue, green, red, near-infrared, and short-wave infrared, enhancing the inversion ability of parameters such as suspended sediment concentration (SSC), and constructing a spectral dataset with high spatio-temporal resolution.

[0026] In one embodiment, considering that the spectral bands of Sentinel-2AMSI and MOD09A1Terra are roughly equivalent as shown in Table 1, the Sentinel-2AMSI image shown in Table 1 is selected for the Sentinel image, and the MOD09A1Terra image is selected for the MODIS image.

[0027] Table 1 Step S2: Manually annotate the spectral dataset to construct a suspended sediment concentration classification dataset, and perform synthetic expansion on minority class samples based on hydrodynamic similarity constraints to generate a balanced training set.

[0028] In one embodiment, through a grid annotation interface, experts visually classify the spectral dataset, divide the suspended sediment concentration patterns into three categories: Clear, Translucent, and Turbid, as shown in Equation (2), and combine spectral-spatial features to divide the suspended sediment concentration categories into different groups to construct a suspended sediment concentration classification dataset.

[0029] (2), where C represents the suspended sediment concentration value, with the unit of g / m³.

[0030] In one embodiment, further use an expert-guided SMOTE variant on the suspended sediment concentration classification dataset to perform synthetic expansion on minority class samples based on hydrodynamic similarity constraints to improve data balance, ensure that the interpolation process does not change the spatial boundaries of the original samples, improve the credibility of the generated data, and generate a balanced training set.

[0031] Specifically, for each minority class sample in the suspended sediment concentration classification dataset, identify the k nearest neighbor samples of this minority class sample in the hydrodynamically similar region, and calculate the feature difference between this minority class sample and the nearest neighbor samples. The feature difference is expressed as: ; where and represent the feature vectors of minority class sample i and nearest neighbor sample j respectively, and Δd is the interpolation difference.

[0032] Next, randomly generate an interpolation coefficient, and calculate a new sample and the feature vector of this new sample based on this interpolation coefficient and the interpolation difference. Among them: the new sample is expressed as: i + θ×Δd; the feature vector of the new sample is expressed as: ; where θ is the interpolation coefficient, θ∈(0,1), and is used to control the interpolation position of the feature vector of the new sample between and . This process is repeated until each minority class sample is expanded to a data balanced state.

[0033] Step S3: Use the balanced training set to train a suspended sediment concentration inversion model to obtain an initial predicted value of the suspended sediment concentration, and dynamically optimize the spectral feature weights and classification boundaries of the suspended sediment concentration inversion model based on expert knowledge, where the suspended sediment concentration inversion model uses a regularized machine learning classifier model.

[0034] In one embodiment, a suspended sediment concentration inversion model is constructed based on a regularized machine learning classifier model. In one embodiment, the regularized machine learning classifier model uses a random forest classifier model, and a λ regularization term is introduced during the training process to reduce the risk of overfitting. To ensure the robustness of the model, the hyperparameters of the random forest (RF) classifier are optimized, ensuring the applicability of the suspended sediment concentration inversion model in different water environments and improving the classification ability for complex water bodies. The random forest (RF) classifier is expressed as: (3), where 、 、 are regularization coefficients, m is the number of trees in the random forest, represents the depth of the tree, represents the minimum number of samples for splitting, represents the minimum number of samples in a leaf node, specifies the maximum number of features.

[0035] In one embodiment, based on expert knowledge, the spectral feature weights and classification boundaries of the suspended sediment concentration inversion model are dynamically optimized to improve the accuracy and stability of the model. Through spectral sensitivity analysis and expert-guided optimization, the applicability of the model in different water environments is ensured, and the errors caused by spectral complexity and differences in hydrodynamic characteristics are reduced.

[0036] Specifically, through spectral sensitivity analysis, the influence of different bands on SSC inversion is evaluated, several key spectral features are selected, and an initial suspended sediment concentration (SSC) inversion model is constructed, as shown in Equation (4): (4), where represents several key spectral features of the water body.

[0037] In one embodiment, based on expert knowledge, the several key spectral features are sorted according to their importance to ensure that the explanatory ability of the key bands and features for SSC changes conforms to the hydrodynamic law. Combining the hydrodynamic law, the spectral features that may cause misclassification are identified, and the weights of the corresponding spectral features are adjusted to reduce the model bias.

[0038] Meanwhile, in the water body category boundary region (such as the transition zone from the near shore to the open sea), due to the complex hydrodynamic and optical characteristics, the model is prone to misclassification. To reduce the classification error in this region, expert collaborative annotation is introduced to correct the classification boundary of the suspended sediment concentration inversion model. Through expert collaborative annotation, the classification accuracy can be improved. Combining multi-temporal data, the change trend of SSC is analyzed, and the classification strategy in the boundary region is adjusted to avoid the error caused by single-temporal images.

[0039] Step S4: Based on the hydrodynamic law and spatial consistency constraint, locally correct the initial predicted value of the suspended sediment concentration and spread it to the global range to obtain the global predicted value of the suspended sediment concentration.

[0040] In one embodiment, first, based on the initial predicted value of the suspended sediment concentration, local correction is performed by combining hydrodynamic principles and expert knowledge to obtain the local predicted value of the suspended sediment concentration, and the credibility and accuracy of the model are improved through data optimization. By introducing spatial consistency constraints and optimizing data quality, the inaccuracies caused by model errors and data noise can be effectively reduced, thereby enhancing the stability and reliability of the inversion results.

[0041] Specifically, traverse the initial predicted value of the suspended sediment concentration with a sliding window of a given size (e.g., 3×3), and calculate the sediment transport threshold within each window based on the sediment transport principle. This process ensures that in the water body, the change trend of the suspended sediment concentration is consistent with the actual hydrodynamic process. Specifically, the transport of sediment is affected by factors such as water flow velocity and particle size distribution, and the SSC values in the neighborhood should have a certain spatial continuity. Therefore, by means of a sliding window, calculate the sediment transport threshold within each grid area (see Equation (5)) to determine the reasonable fluctuation range of SSC in this area.

[0042] (5), where, represents the measured value of the suspended sediment concentration, represents the predicted value of the suspended sediment concentration output by the model.

[0043] Then, outliers exceeding this threshold are detected. These outliers often stem from factors such as optical sensor errors, cloud cover, or noise during transmission. These outliers are marked as samples requiring manual correction and are further revised through expert review. This process ensures the spatial consistency of the model results and avoids misclassification or incorrect prediction caused by individual outliers. The experts manually correct the outliers based on the hydrodynamic laws and the measured values of the suspended sediment concentration, adjust the outliers to conform to the sediment transport law of the actual water body, and mark the adjusted data as high-confidence reference data, which is re-input into the suspended sediment concentration inversion model for parameter optimization, thereby improving the spatial consistency of the entire dataset and enhancing the credibility of the inversion results.

[0044] Meanwhile, a spatial consistency constraint is introduced into the loss function of the suspended sediment concentration inversion model to control the local predicted values of the suspended sediment concentration to satisfy the spatial consistency constraint within the neighborhood range. In this way, the model can identify and correct local mutations and inconsistencies, effectively avoiding the influence of non-natural fluctuations on the model inversion results.

[0045] Furthermore, numerical transformation processing is performed on the balanced training set to optimize the training data distribution, specifically including dynamically filtering out or correcting low-confidence data, using normalization or standardization methods to adjust the data range, and / or performing denoising processing on the existing noise, etc. Specifically, to improve the training effect of the model, numerical transformation is performed on the original data to remove the influence of low-confidence data and optimize the data distribution. First, low-confidence data is removed through dynamic filtering. These data may introduce noise due to optical interference, seasonal changes, or the incompleteness of remote sensing images. By comparing the expert-annotated data with the model output, we screen out data points with low credibility and exclude or correct them. To optimize the data distribution, normalization or standardization methods are used to adjust the data range, reducing the differences between eigenvalue s and improving the convergence speed and stability of model training. For the possible noise in remote sensing data, wavelet transform or filtering algorithms are applied for denoising processing to improve the signal-to-noise ratio of the data, enabling the model to learn effective information more accurately and improving the SSC inversion accuracy. At the same time, the processed data is marked as high-confidence reference data and re-input into the suspended sediment concentration inversion model for parameter optimization. The experts further adjust the SSC predicted values that do not conform to physical laws or hydrodynamic models by comparing the remote sensing image inversion results with the actual observation data. These high-confidence reference data provide a reliable reference standard to help the model better adjust the prediction results and reduce errors.

[0046] In addition, in one embodiment, in order to reduce the deviation between the model estimation value and the true value, residual analysis is introduced to analyze the model error and correct the error between the predicted value of the local suspended sediment concentration and the measured value of the suspended sediment concentration. By observing the residual distribution, the estimation process of the model is adjusted to ensure that it is as close as possible to the SSC distribution of the actual water body. This process not only improves the accuracy of the model but also enhances its adaptability in different types of water body environments.

[0047] In addition, in one embodiment, as the water body environment changes, new remote sensing data and on-site observation data emerge continuously. Therefore, a dynamic feedback mechanism is established to update the high-confidence reference data. After each data update, the model is retrained and adjusted according to the updated high-confidence reference data to ensure that it can always reflect the latest water body change trend. At the same time, experts will manually correct the new data during each feedback to ensure that it conforms to the hydrodynamic laws and sediment transport characteristics.

[0048] In addition, to further enhance the adaptive ability of the model, the training data is dynamically enhanced by adding newly acquired multi-spectral remote sensing images and site observation data to update the multi-spectral remote sensing images, and then update the balanced training set, so that it can adapt to different environmental and seasonal conditions, improve the accuracy of SSC inversion, and finally achieve stable and reliable suspended sediment concentration inversion.

[0049] After the local correction through the above process, the predicted value of the local suspended sediment concentration is obtained and expressed as: (6), where, represents the predicted value of the local suspended sediment concentration; α is the weight adjustment coefficient; is the measured value of the suspended sediment concentration; x is several key spectral features input, and f i (x) is the optimized suspended sediment concentration (SSC) inversion model, and x is the feature vector of the i-th sample in the balanced training set.

[0050] In one embodiment, in order to ensure the global consistency of the SSC inversion result, an adaptive diffusion mechanism is adopted, a dynamic balance factor is introduced, and the local correction is propagated to the global scope to avoid the cascading error caused by local correction. At the same time, through the Technique for Order Preference by Similarity to an Ideal Solution (TOPSIS) weighted feature space adjustment, the dynamic balance factor is recalibrated to optimize the global predicted value of the suspended sediment concentration and improve the applicability and accuracy of the model in optically complex water body environments.

[0051] Specifically, after introducing local corrections, calculate their impact on the global SSC estimation result to ensure that local adjustments do not amplify errors or generate cumulative biases during subsequent iterations. To this end, introduce a dynamic balance factor to control the influence weight of local corrections on the global scale, and obtain the predicted value of the global suspended sediment concentration, avoiding cascading errors, as shown in Equation (7). The predicted value of the global suspended sediment concentration is expressed as: (7), where, is the predicted value of the global suspended sediment concentration, is the initial predicted value of the suspended sediment concentration prediction value, is the predicted value of the local suspended sediment concentration, and P is the dynamic balance factor.

[0052] In addition, in one embodiment, to prevent local corrections from affecting the overall data distribution, a statistical consistency optimization method is adopted. By calculating the mean, standard deviation, and divergence of the data before and after correction, verify the rationality of the data distribution, and adjust the dynamic balance factor in combination with the chi-square test to ensure that the mean and standard deviation of the SSC estimation values before and after correction are consistent. At the same time, use the Kullback-Leibler divergence to evaluate the change in the data distribution and perform a chi-square test to ensure the rationality of the corrected data distribution. In addition, combine an expert knowledge-driven weight adjustment mechanism to make the corrected data more in line with hydrodynamic laws and improve the credibility of the model.

[0053] In addition, in one embodiment, during the global correction process, for spectral interference that may be caused by optically complex waters (such as eutrophic waters, algal dense areas, turbid estuarine waters), combine multispectral and short-wave infrared (SWIR) to improve the adaptability of the model to highly turbid waters. At the same time, adopt an adaptive spectral normalization method to reduce the interference of water color changes on SSC inversion, and combine hydrodynamic simulation to improve the prediction accuracy. In addition, establish an abnormal area recognition mechanism to dynamically adjust the correction propagation strategy to ensure the stability of SSC inversion in different environments.

[0054] In addition, in one embodiment, after the local correction propagates to the global scale, adjust the loss function of the model, introduce global consistency constraints into the model loss function to ensure the stability of the final SSC inversion result. Combine the gradient descent optimization method to dynamically adjust the hyperparameters of the model to meet the hydrodynamic laws and spatial consistency constraints, and combine long-term observation data for trend analysis to prevent systematic biases in SSC inversion.

[0055] In addition, in one embodiment, during the global correction process, to continuously optimize the SSC inversion model, a dynamic data feedback system is constructed. After each data update, error analysis is performed to adjust the local correction strategy and improve the accuracy of long-term prediction. At the same time, we introduce an automated anomaly detection algorithm, combined with an active learning mechanism, to improve the reliability of the model, and establish a long-term SSC monitoring system based on multi-temporal data to ensure the stability and adaptive ability of the model.

[0056] In addition, in one embodiment, a variety of statistical and visualization methods are further adopted to verify the performance of the suspended sediment concentration inversion model and analyze its spatio-temporal variation trend.

[0057] Specifically, for model performance evaluation, in one embodiment, Taylor diagrams and scatter plots are used to visualize the global suspended sediment concentration prediction values and evaluate the performance of the model. Among them, the Taylor diagram comprehensively displays the correlation coefficient (r), standard deviation (Std.Dev), and centered root mean square error (CRMSE) to measure the closeness of the model to the measured values.

[0058] In one embodiment, the Nash-Sutcliffe efficiency (NSE) is calculated to evaluate the fitting ability of the model. The closer the NSE value is to 1, the more accurate the model estimation is.

[0059] For the spatio-temporal variation trend, in one embodiment, the Mann-Kendall (MK) trend test is used to evaluate the significant changes in SSC in the time dimension. When |Z| > 1.96 and p < 0.05, it indicates a significant trend statistically.

[0060] In one embodiment, the conversion rate (CR) is calculated to quantify the rate of concentration change between seasonal variations. CR is determined by calculating the percentage change in the global suspended sediment concentration prediction values between the current season and the next season for each grid cell.

[0061] In one embodiment, the variation characteristics of the global suspended sediment concentration prediction values with different seasons are evaluated through grid analysis to reveal the spatial distribution trends in different regions.

[0062] As Figures 2 - 6 shown in Figure 2 Figure 25 shows a schematic diagram of the spatial distribution of the measured values and estimated values of suspended sediment concentration obtained by implementing the method of the present invention in an example. In the figure, the left area A diagram shows the measured results of suspended sediment concentration, and the right area B diagram shows the predicted values of suspended sediment concentration estimated by the inversion model; Figure 3Performance comparison of suspended sediment concentration inversion models. Model 1 is the human-machine collaborative model based on multi-source remote sensing data fusion of the present invention. Model 2 is the existing human-machine collaboration model only. Model 3 is the empirical algorithm model based on multi-source remote sensing data fusion of the prior art. Model 4 is the empirical algorithm model of the prior art. Figure A uses a Taylor diagram to compare the model performance. Figures B - E show the model scatter plots and NSE values. Figure B is for Model 1, Figure C is for Model 2, Figure D is for Model 3, and Figure E is for Model 4; Figure 4 Schematic diagram of the estimated annual suspended sediment concentration distribution from 2017 to 2024 for the model with [example]. The mean figure is the average of the suspended sediment concentration distributions over 8 years; Figure 5 Schematic diagram for the assessment of the spatial heterogeneity of the interannual suspended sediment concentration characteristics of the example. The left figure shows the Mann-Kendall (MK) trend test Z-scores of the suspended sediment concentration in the coastal waters of China from 2017 to 2024, showing a statistically significant trend (|Z|>1.96, p<0.05). Red dots indicate an increase in SSC during this period, and blue dots indicate a decrease in SSC during this period. The enlarged view on the right shows the fused satellite images of Sentinel-2A MSI and MOD09A1 Terra surface reflectance products at the Yellow River Delta (A, 38°N) and the Yangtze River Estuary (B, 31°N) in 2017 and 2024, and the local Z-scores from 2017 to 2024 are overlaid on the satellite images of the Yellow River Delta and the Yangtze River Estuary in 2024, showing that approximately 80% of the pixels in the Yellow River Delta have a significant decrease (p<0.05), while the grid cells with increasing SSC seem to be concentrated in the Yangtze River Estuary and its surrounding waters; Figure 6 Schematic diagram for the assessment of the seasonal heterogeneity of the suspended sediment concentration characteristics of the example. This figure shows the seasonal distribution of the suspended sediment concentration estimated by the model of the present invention from 2017 to 2024, as well as the SSC conversion rate between seasons. It can be seen from this that the method of the present invention can accurately predict the suspended sediment concentration and meet the requirements for monitoring the dynamic changes of global SSC.

[0063] In summary, the optical complex water body suspended sediment monitoring method based on human-machine collaboration disclosed by the present invention combines remote sensing image analysis, machine learning, and expert knowledge to achieve high-precision inversion of SSC. This method uses publicly available high-resolution satellite images, estimates based on the relationship between the backscattering coefficient and SSC, and iteratively optimizes the inversion results through minimal manual correction, thereby reducing the dependence on traditional in-situ measurements, improving the monitoring accuracy and applicability. This method follows the principle of human-machine interaction, embeds human intelligence in the SSC inversion, enhances the interpretability of the model, and ensures that the SSC inversion conforms to the hydrodynamic laws.

[0064] The above specific embodiments do not constitute a limitation to the protection scope of the present invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions and improvements made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. An optical complex water body suspended sediment monitoring method based on human-machine collaboration, characterized in that, The method includes: Collecting multi-spectral remote sensing images of the target area and performing preprocessing, performing spatio-temporal fusion on the preprocessed multi-spectral remote sensing images to obtain a spectral dataset with high spatio-temporal resolution; Manually annotating the spectral dataset to construct a suspended sediment concentration classification dataset, and synthetically expanding minority class samples based on hydrodynamic similarity constraints to generate a balanced training set; Training a suspended sediment concentration inversion model using the balanced training set to obtain an initial predicted value of the suspended sediment concentration, and dynamically optimizing the spectral feature weights and classification boundaries of the suspended sediment concentration inversion model based on expert knowledge, where the suspended sediment concentration inversion model uses a regularization machine learning classifier model; Based on hydrodynamic laws and spatial consistency constraints, locally correcting the initial predicted value of the suspended sediment concentration and spreading it to the global range to obtain a global predicted value of the suspended sediment concentration.

2. The method according to claim 1, wherein Performing spatio-temporal fusion on the preprocessed multi-spectral remote sensing images to obtain a spectral dataset with high spatio-temporal resolution, specifically including: Using an improved object-based spatio-temporal adaptive reflectance fusion model to fuse the preprocessed multi-spectral remote sensing images to generate a fused image with both high spatial resolution and high temporal resolution, where the calculation formula of the spatio-temporal adaptive reflectance fusion model is: ， Among them, represents the fused image; is the Sentinel image with high spatial resolution; and respectively represent the surface reflectance of the MODIS image at times t and t−1, and the multispectral remote sensing image includes the Sentinel image and the MODIS image; Based on the fused image, extracting multi-band spectral features to obtain the spectral dataset with high spatio-temporal resolution.

3. The method according to claim 1, characterized in that Manually annotating the spectral dataset to construct a suspended sediment concentration classification dataset, including: Through a grid annotation interface, experts visually classify the spectral dataset, and divide the suspended sediment concentration patterns into three categories: clear, semi-transparent, and turbid; Combining spectral-spatial features, dividing the suspended sediment concentration categories into different groups to construct a suspended sediment concentration classification dataset.

4. The method according to claim 3, wherein Using an expert-guided SMOTE variant to synthetically expand minority class samples based on hydrodynamic similarity constraints, including: For each minority class sample in the suspended sediment concentration classification dataset, identifying k nearest neighbor samples of the minority class sample in a hydrodynamically similar area, and calculating the feature difference between the minority class sample and the nearest neighbor samples, which is expressed as: Among them, and respectively represent the feature vectors of the minority-class sample i and the nearest neighbor sample j, and Δd is the interpolation difference; Randomly generating an interpolation coefficient, and calculating a new sample and the feature vector of the new sample based on the interpolation coefficient and the interpolation difference, where: The new sample is expressed as: i + θ×Δd; The feature vector of the new sample is represented as: ; where θ is the interpolation coefficient, θ ∈ (0, 1), which is used to control the feature vector of the new sample at and the interpolation position between them.

5. The method according to claim 4, wherein Based on expert knowledge, the suspended mud Spectrally optimizing the spectral feature weights and classification boundaries of the sand concentration inversion model, including: Screening out several key spectral features in the balanced training set through spectral sensitivity analysis, sorting the several key spectral features according to importance based on expert knowledge, and identifying spectral features that may cause misclassification in combination with hydrodynamic laws, and adjusting the weights of the corresponding spectral features; Introducing expert collaborative annotation in the water body category boundary area to correct the classification boundary of the suspended sediment concentration inversion model.

6. The method according to claim 5, characterized in that, Based on hydrodynamic laws and spatial consistency constraints, locally correct the initial predicted value of suspended sediment concentration to obtain a locally predicted value of suspended sediment concentration, including: Traverse the initial predicted value of suspended sediment concentration with a sliding window of a given size, calculate the sediment transport threshold within each window based on the sediment transport principle, and detect outliers exceeding this threshold; Manually correct the outliers by experts according to hydrodynamic laws and measured values of suspended sediment concentration, and adjust the outliers to conform to hydrodynamic laws; Mark the adjusted data as high-confidence reference data and re-enter it into the suspended sediment concentration inversion model for parameter optimization; Introduce spatial consistency constraints into the loss function of the suspended sediment concentration inversion model to control the locally predicted value of suspended sediment concentration to satisfy spatial consistency constraints within the neighborhood range.

7. The method according to claim 6, characterized in that During the process of locally correcting the initial predicted value of suspended sediment concentration, it further includes: Perform numerical transformation processing on the balanced training set, including dynamically filtering and removing or correcting low-confidence data, adjusting the range of data using normalization or standardization methods, and / or denoising the existing noise; Mark the processed balanced training set as high-confidence reference data and re-enter it into the suspended sediment concentration inversion model for parameter optimization; Introduce residual analysis to correct the error between the locally predicted value of suspended sediment concentration and the measured value of suspended sediment concentration; Establish a dynamic feedback mechanism to update and correct the high-confidence reference data; Add newly acquired multi-spectral remote sensing images to update the balanced training set. The locally predicted value of suspended sediment concentration 8. The method according to claim 7, characterized in that is expressed as: Adopt an adaptive diffusion mechanism to propagate the ， Among them, represents the predicted value of the local suspended sediment concentration; α is the weight adjustment coefficient; is the measured value of the suspended sediment concentration; fi(x) is the optimized inversion model of the suspended sediment concentration, and x is the feature vector of the i-th sample in the balanced training set.

9. The method according to claim 8, wherein local correction to the global scope, including: Introduce a dynamic balance factor to control the influence weight of local correction on the global scope, and obtain the globally predicted value of suspended sediment concentration, which is expressed as: Adopt a statistical consistency optimization method to verify the rationality of data distribution by calculating the mean, standard deviation and divergence of the data before and after correction, and adjust the dynamic balance factor in combination with the chi-square test; ， Among them, is the predicted value of the global suspended sediment concentration, is the initial suspended sediment concentration Predicted value, is the predicted value of the local suspended sediment concentration, and P is the dynamic equilibrium factor; Introduce global consistency constraints into the model loss function, and dynamically adjust the hyperparameters in combination with the gradient descent optimization method to satisfy hydrodynamic laws and spatial consistency constraints. It also includes: verifying the performance of the suspended sediment concentration inversion model, including:

10. The method according to claim 1, characterized in that, Visualize the globally predicted value of suspended sediment concentration using Taylor diagrams, scatter plots and violin plots to evaluate the performance of the suspended sediment concentration inversion model; Calculate the NSE efficiency to evaluate the fitting ability of the suspended sediment concentration inversion model; Calculate the MK trend test to evaluate the significant change of the globally predicted value of suspended sediment concentration in the time dimension; Calculate the conversion rate to quantify the concentration change rate of the globally predicted value of suspended sediment concentration between different seasonal changes; Evaluate the change characteristics of the globally predicted value of suspended sediment concentration with different seasonal changes through grid analysis. ​

Images