Remote sensing retrieval method and system for suspended sediment concentration based on TSM-SSC conversion

Abstract

The present application belongs to the technical field of marine remote sensing and water quality monitoring, and discloses a suspended sediment concentration remote sensing inversion method and system based on TSM-SSC conversion. The method is based on surface reflectance data obtained by multi-source remote sensing satellites, selects TSM remote sensing products with consistent acquisition time, and combines with field SSC observation data to construct a paired sample set; by constructing and screening multispectral features, a remote sensing inversion model is trained with TSM as the target, and the model is preferably a random forest model (average R 2 ≈0.75); a TSM-SSC conversion equation is established based on the TSM inversion result and the field SSC observation value, and batch inversion is performed on long-time sequence remote sensing image data to generate suspended sediment concentration spatiotemporal distribution products. The present application improves the stability and applicability of suspended sediment concentration remote sensing inversion, and enhances the inversion capability for complex hydrodynamic conditions and high turbidity water bodies.

Description

Technical Field

[0001] This invention belongs to the field of marine remote sensing and water quality monitoring technology, and discloses a remote sensing inversion method and system for suspended sediment concentration based on TSM-SSC conversion. In particular, it relates to a long-term time-series remote sensing inversion method for suspended sediment concentration based on the conversion of total suspended matter (TSM) to suspended sediment concentration (SSC). Specifically, it is a deep learning model that fits the mapping relationship between remote sensing inversion results and measured data by constructing the model, and obtains spatiotemporal products at the annual and monthly scales by combining the long-term SSC batch inversion results, so as to realize the dynamic monitoring and ecological environment assessment of high turbidity estuaries and nearshore waters. Background Technology

[0002] Suspended sediment concentration (SSC) in coastal and estuarine nearshore waters is a crucial water quality parameter characterizing water turbidity, sediment transport processes, and the state of the ecological environment. It has significant applications in waterway maintenance, coastal engineering design, ecological restoration, and pollution diffusion assessment. Traditional methods for obtaining SSC primarily rely on on-site sampling and laboratory analysis. While these methods offer high accuracy, they suffer from drawbacks such as high sampling costs, limited spatial coverage, and difficulty in acquiring continuous long-term data series, making them unsuitable for meeting the practical needs of large-scale, long-term dynamic monitoring of the water environment.

[0003] To overcome the limitations of in-situ observation, SSC inversion methods based on optical remote sensing imagery have been extensively studied. Among existing technologies, those closest to this invention mainly include: methods that directly invert SSC using multispectral reflectance or its derived indices based on a single remote sensing sensor's spectral empirical or semi-empirical model; and methods that introduce machine learning algorithms to model the nonlinear relationship between remote sensing reflectance and SSC. While these methods can achieve certain inversion accuracy in transparent or low-to-medium turbidity waters, they generally suffer from decreased inversion accuracy, insufficient stability, and poor consistency in high-turbidity waters such as estuaries and nearshore areas with complex optical conditions, making them unsuitable for long-term continuous monitoring applications.

[0004] Analysis reveals that the aforementioned defects in the existing technology stem primarily from the following reasons: (1) Spectral signal saturation and non-uniqueness. Under high turbidity water conditions, the reflectance of red light and near-infrared bands is prone to saturation. SSCs in different concentration ranges may correspond to similar spectral response characteristics, making it difficult for direct inversion models based on a single sensor or a single band combination to accurately distinguish SSC changes; (2) The combined effects of multiple optically active substances. Estuarine and nearshore waters usually contain multiple components such as suspended inorganic particles, phytoplankton and dissolved colored substances. Their combined effects make the correspondence between remote sensing reflectance and SSC more complex, weakening the physical consistency and generalization ability of the direct inversion model; (3) The training samples are sparse in time and space and lack consistency. Field SSC measured data are usually scattered and discontinuous in time, while remote sensing images are affected by factors such as cloud cover, tidal changes and imaging time difference of different sensors, resulting in a limited number of high-quality samples that can be used for modeling. The model is easily affected by uneven distribution of samples and it is difficult to guarantee long-term stability. (4) Limited temporal coverage of a single data source. Although high spatial resolution remote sensing satellites (such as Sentinel-2 and Landsat-7) can provide rich spatial detail information, their revisit period is long. High temporal resolution sensors (such as MODIS and GOCI-II) have certain limitations in terms of spatial resolution or band settings. It is difficult for a single remote sensing data source to simultaneously take into account both long-term temporal continuity and inversion accuracy.

[0005] To address the aforementioned issues, existing research has begun to introduce Total Suspended Solids (TSM) as an intermediate variable, constructing a conversion relationship between TSM and SSC to indirectly estimate SSC. This type of method can leverage TSM products with strong temporal coverage to expand the training sample size, alleviating the problem of insufficient field samples to some extent. However, existing TSM-SSC conversion methods are mostly based on empirical formulas or local fitting relationships from a single data source or within a specific region. Their modeling process typically does not fully consider the differences in spatial resolution, band response, and observation time among multi-source remote sensing data, and lacks systematic constraints on the transmission effect of TSM inversion errors in the SSC conversion process and on physical consistency. Furthermore, existing methods often focus on single-period or short-term timescale analysis, lacking consistency modeling and stability assessment mechanisms for long-term series. This leads to the TSM-SSC conversion relationship easily drifting when there are significant changes in seasons, years, or hydrodynamic conditions, making it difficult to support stable and reliable long-term SSC batch inversion.

[0006] Therefore, under the existing technical framework based on a single remote sensing data source, direct SSC inversion, or simple TSM-SSC empirical conversion, the aforementioned problems are difficult to solve simultaneously through conventional parameter adjustments or simply changing the model algorithm. The above analysis reveals the following technical contradictions and shortcomings in the existing technology: (1) The contradiction between the need for sample size expansion and the constraints of spatiotemporal consistency. Field SSC measured data is scarce, while the differences in spatial resolution, imaging time and preprocessing process of multi-source remote sensing data make it difficult to construct high-quality training samples with spatiotemporal consistency by simply superimposing data sources, which in turn affects the generalization ability and long-term stability of the model; (2) The contradiction between nonlinear spectral response and insufficient model expressive power under high turbidity conditions. In water environments with significant spectral saturation and mixing of optical components, traditional empirical models or simple machine learning models cannot simultaneously achieve both inversion accuracy and physical rationality; (3) The contradiction between the requirement of temporal continuity and the preservation of the authenticity of spatial structure. Although using TSM as an intermediate variable can improve the temporal coverage capability, if there is a lack of systematic modeling for multi-source differences, nonlinear mapping relationships and physical consistency, it is still difficult to maintain the authenticity and stability of the SSC spatial distribution structure while ensuring long-term temporal continuity.

[0007] Therefore, there is an urgent need for a remote sensing inversion method for suspended sediment concentration that can integrate multi-source remote sensing data, fully explore the intrinsic relationship between TSM and SSC, and simultaneously take into account inversion accuracy, spatial structure rationality and long-term stability under high turbidity and complex hydrodynamic conditions, so as to solve the above-mentioned shortcomings of existing technologies in long-term continuous monitoring of estuaries and nearshore waters. Summary of the Invention

[0008] To overcome the problems of poor long-term temporal continuity, insufficient model generalization ability due to scarce field samples, limitations in resolution / temporal range of single sensors, and decreased inversion accuracy caused by optical signal saturation in high-turbidity water bodies in existing remote sensing inversion methods for suspended sediment concentration (SSC), this invention discloses a method and system for remote sensing inversion of suspended sediment concentration based on TSM-SSC conversion. The technical solution is as follows: This invention is implemented as follows: a remote sensing inversion method for suspended sediment concentration based on TSM-SSC conversion, comprising the following steps: S1. Multi-source remote sensing data preprocessing: Multi-source remote sensing image data is acquired on the Google Earth Engine cloud platform. Clouds and cloud shadows are removed from the multi-source remote sensing images based on the QA band of the remote sensing images. Surface reflectance correction processing is performed on the remote sensing images based on the scale factor and offset parameters in the image metadata. S2. Sample extraction at corresponding locations: Based on the surface reflectance remote sensing data acquired by multi-source satellites, select TSM products acquired on the same day with the same acquisition time; combine the spatial location information of the on-site SSC measurement station to extract the corresponding image data, construct a spatiotemporally consistent paired sample set, and obtain the corresponding sample data of surface reflectance and TSM value for each band. S3. Feature Construction and Screening: Based on the original band data of multispectral remote sensing images, calculate various derived spectral feature indices, correct the normalized water index MNDWI and chromaticity angle, and construct feature vectors for model training. The derived spectral characteristic indices include Normalized Difference Vegetation Index (NDVI), Modified Normalized Difference Water Index (MNDWI), Normalized Difference Index (NDSI), band reflectance ratio, and multi-band total reflectance, etc. The features were screened using correlation analysis and feature importance assessment methods, and those features that contributed significantly to TSM inversion were selected as model inputs.

[0009] S4. TSM Inversion Model Training: Using feature vectors as model input and corresponding TSM values ​​as training targets, various machine learning regression algorithms are employed for model training, and the training is performed based on the evaluation metric R. 2 Select the optimal TSM remote sensing inversion model; S5. TSM-SSC Transformation Modeling: Based on the TSM inversion results obtained from the TSM remote sensing inversion model and the corresponding field SSC observations, a TSM-SSC paired sample set is constructed. Modeling is performed on the paired samples using linear, logarithmic, and power functions, and the model is then evaluated based on the R-index. 2 Determine the optimal TSM-SSC transformation equation; S6. Long-term SSC batch inversion: Using the TSM remote sensing inversion model and the TSM-SSC transformation equation, batch inversion processing is performed on long-term remote sensing image data to generate spatiotemporal distribution products of suspended sediment concentration. The inversion results are statistically analyzed to generate spatiotemporal distribution products of suspended sediment concentration at different time scales (including annual, quarterly, or monthly scales) to characterize the spatial distribution characteristics of SSC and its variation over time.

[0010] In step S1, the multi-source remote sensing data includes surface reflectance data acquired by the MODIS satellite as the main source data, total suspended matter concentration (TSM) products obtained by inversion from the GOCI-II satellite, and surface reflectance remote sensing image data acquired by the Sentinel-2 and Landsat-7 satellites as auxiliary data; all multi-source remote sensing data are obtained through the Google Earth Engine (GEE) cloud computing platform or official data release channels.

[0011] In step S1, the multi-source remote sensing data preprocessing also includes: The QA bitmask information corresponding to different data sources was analyzed to remove clouds, cloud shadows and invalid pixels; remote sensing images with different spatial resolutions were resampled to a unified spatial scale and unified to a projection coordinate system consistent with the measured data of suspended sediment concentration in the field to ensure the spatial consistency of multi-source remote sensing data.

[0012] In step S2, a spatiotemporally consistent paired sample set is constructed, including: Based on the acquisition time of the remote sensing images, TSM data corresponding to the time phase are selected. Based on the spatial location of the measured points of suspended sediment concentration in the field, multi-source remote sensing reflectance data and the TSM data are paired at the pixel level. Statistical analysis methods are used to control the quality of the paired samples, and outliers are identified and removed based on the interquartile range (IQR) criterion of the box plot to obtain a paired sample set for modeling.

[0013] The paired samples are matched according to spatial and temporal windows, and the pairing results are filtered in conjunction with tidal information or observation time series information to reduce pairing errors. Specifically, remote sensing image data and field measured data are first initially paired using a preset spatial buffer range and time threshold; then, based on the tidal phase or observation time sequence, paired samples with obvious time series deviations or inconsistent tidal states are eliminated to ensure the consistency and reliability of the paired samples in time and space.

[0014] In step S3, the derived spectral characteristic index includes the following indicators: Normalized Difference Vegetation Index (NDVI): ; Modified Normalized Difference Water Index (MNDWI): ; Normalized Difference Index (NDSI): ; The ratio of red light to near-infrared reflectance: ; Total reflectance (TotalRef): ; In the formula, These are the remote sensing reflectances in the red, green, near-infrared, and near-shortwave infrared bands, respectively. For the first multispectral remote sensing image Reflectivity of each band This represents the total number of bands.

[0015] In step S3, feature selection includes: The correlation between each candidate feature and the target variable was analyzed using the Pearson correlation coefficient. Based on the feature importance evaluation method of the machine learning model, the contribution of each feature to the model prediction results was ranked to obtain the contribution results of each feature vector. Based on the correlation analysis results and feature importance ranking results, features with high correlation and large contribution are selected as the main input features for model training.

[0016] In step S4, the machine learning regression algorithm includes a random forest regression model, a gradient boosting decision tree model, an extreme gradient boosting model (XGBoost), and a lightweight gradient boosting model (LightGBM); based on the evaluation metric R... 2 By comparing and analyzing the models, the preferred model for TSM remote sensing inversion was determined.

[0017] By comparing and analyzing the evaluation indicators of the inversion results of different models, the random forest regression model is selected as the TSM remote sensing inversion model to obtain higher inversion accuracy and model stability.

[0018] In step S5, the physical consistency of the sample pairs formed by the TSM inversion results and the on-site SSC observations is first screened, and sample pairs with TSM values ​​less than the corresponding SSC values ​​are removed. For the screened samples, the TSM-SSC transformation relationship is fitted using various functional forms such as linear functions, logarithmic functions, and power functions, and the evaluation index R is used. 2 The fitting results were analyzed, and the TSM-SSC transformation equation was established.

[0019] In step S6, a spatiotemporal product of SSC is generated, which includes SSC inversion sequence diagrams based on different time scales, including annual, quarterly, or monthly scales; the spatiotemporal product also includes corresponding statistical feature diagrams, which may include mean, standard deviation, etc., to represent the spatiotemporal distribution and trend of SSC.

[0020] Another objective of this invention is to provide a suspended sediment concentration remote sensing inversion system for implementing the aforementioned TSM-SSC conversion-based suspended sediment concentration remote sensing inversion method. This system has a modular structure and includes: The data acquisition and preprocessing module is used to acquire remote sensing image data from multiple remote sensing data sources such as MODIS, GOCI-II, Sentinel-2 and Landsat-7, and to perform quality control and preprocessing operations on the remote sensing image data. The preprocessing operations include cloud and cloud shadow removal, surface reflectance correction, geometric correction and spatial resampling processing to obtain remote sensing reflectance data that meets the requirements of subsequent analysis. The sample pairing and screening module is used to pair multi-source remote sensing reflectance data, TSM product data and on-site measured suspended sediment concentration data based on the acquisition time and spatial location of remote sensing images, and to perform quality control and outlier screening on the paired samples to obtain high-quality sample data that are consistent in time and space. The feature construction and screening module is used to extract original band features based on the remote sensing reflectance data and calculate the derived spectral feature index, and screen the features through correlation analysis and feature importance evaluation to generate feature vectors for model training. The model training and evaluation module is used to train the TSM remote sensing inversion model based on the feature vector and the corresponding TSM sample values, and to evaluate and compare the inversion performance of different models in order to determine the optimal model for TSM batch inversion. The TSM-SSC conversion module is used to construct a conversion sample based on the inversion results of the TSM remote sensing inversion model and the measured data of suspended sediment concentration in the field, and to establish the conversion relationship between TSM and SSC, so as to realize the quantitative conversion of TSM to SSC. The batch inversion and product generation module is used to apply the TSM remote sensing inversion model and TSM-SSC conversion relationship to long-term remote sensing image data, perform batch inversion processing of suspended sediment concentration, and generate spatiotemporal distribution products of suspended sediment concentration at different time scales. The results display and export module is used to visualize the suspended sediment concentration inversion results and output raster products and related statistical analysis results.

[0021] Furthermore, the suspended sediment concentration remote sensing inversion system can be deployed in a cloud computing environment or a local server environment, and the data acquisition and preprocessing module supports automated processing based on a cloud platform to improve the efficiency of long-term remote sensing data processing.

[0022] Furthermore, the model training and evaluation module supports updating and training the TSM remote sensing inversion model after acquiring new field sample data or high-resolution reference data, so as to improve the adaptability and stability of the model in long-term application.

[0023] Combining all the above technical solutions, the beneficial effects of this invention are as follows: First, this invention significantly alleviates the training constraints caused by the scarcity of field SSC samples by introducing remote sensing TSM products as intermediate variables and constructing a TSM-SSC conversion relationship. This allows long-term SSC inversion based on MODIS to maintain high accuracy and temporal continuity even with limited samples, providing a feasible solution for multi-year (2003–2024) continuous monitoring. Through systematic comparison and cross-validation of different machine learning models, the random forest model is adopted as the main model, which improves the robustness of the inversion and the multivariate nonlinear fitting ability (average R²≈0.75). Compared with traditional single empirical models, it significantly improves the overall inversion performance and enhances the applicability to high-turbidity water bodies.

[0024] Secondly, the annual, seasonal, and monthly SSC spatiotemporal distribution products generated by this invention can reveal the spatiotemporal evolution characteristics of regional sediment, such as seasonal patterns of high sediment levels in winter and low levels in summer, as well as interannual fluctuations. This provides long-term, continuous data support for sediment dynamics research, waterway management, and ecological environment assessment. The proposed method can be used in conjunction with high-resolution optical data or hydrodynamic models to complement the spatiotemporal coverage of multi-source data, thereby improving the spatial detail and long-term reliability of sediment monitoring. It has promising engineering application prospects and widespread value. Validation in high-turbidity estuarine environments (such as Hangzhou Bay) shows that the method has stability and high inversion accuracy. Furthermore, multi-source fusion or high-resolution data can be introduced as needed to improve nearshore details, serving as an optional extension scheme for subsequent implementation.

[0025] Third, this invention utilizes a long-term time-series suspended sediment concentration remote sensing inversion method based on TSM-SSC conversion. This method can provide continuous, multi-year SSC monitoring data, offering data support for port and waterway management, estuary sediment regulation, coastal water environment assessment, and ecological protection. Long-term SSC products can be used as a basis for decision-making in hydrodynamic modeling, waterway maintenance planning, and pollution diffusion prediction, thereby reducing manual sampling costs, improving management efficiency, and possessing significant commercialization potential and engineering application value.

[0026] Current long-term SSC remote sensing monitoring in high-turbidity estuaries and nearshore waters, both domestically and internationally, generally suffers from problems such as sample scarcity, optical signal saturation, and difficulties in multi-source data registration, lacking robust inversion methods suitable for multi-year continuous monitoring. This invention introduces TSM products as intermediate variables and combines multi-source remote sensing imagery with machine learning inversion methods to establish a TSM-SSC conversion relationship, achieving long-term continuous SSC inversion under high-turbidity water conditions, filling a technological gap in the field of long-term SSC monitoring of high-turbidity water bodies.

[0027] For a long time, researchers have faced challenges in obtaining high-precision, long-term continuous monitoring results for SSC remote sensing inversion under high-turbidity water bodies and complex hydrodynamic conditions, including spectral non-uniqueness, sample scarcity, sensor resolution, and temporal limitations. This invention, through multi-source data fusion, TSM-SSC transformation modeling, and ensemble learning methods such as random forests, achieves long-term, robust, and high-precision SSC inversion, effectively solving the aforementioned long-standing technical difficulties.

[0028] Traditional SSC remote sensing inversion methods often rely on empirical models or single optical sensors, which typically perform poorly in high-turbidity water bodies and exhibit biased predictions for low-turbidity conditions. This invention, by introducing intermediate variables (TSM), multi-source data, and machine learning models, effectively mitigates the limitations of single spectral information, achieving robust inversion under varying turbidity levels and complex optical conditions. This overcomes the prediction bias present in existing technologies and improves the applicability and reliability of the inversion results. Attached Figure Description

[0029] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure; Figure 1 This is a flowchart of the remote sensing inversion method for suspended sediment concentration based on TSM-SSC conversion provided in this embodiment of the invention; Figure 2 This is a schematic diagram of the remote sensing inversion method for suspended sediment concentration based on TSM-SSC conversion provided in this embodiment of the invention. Figure 3 This is a long-term SSC inversion sequence diagram for 2003-2024 provided in an embodiment of the present invention; Figure 4 This is an SSC inversion result map of the Hangzhou Bay area from 2003 to 2024, provided by an embodiment of the present invention; Figure 5 This is a map showing the SSC inversion results for each quarter in the Hangzhou Bay area provided in this embodiment of the invention; Figure 6 This is a comparison of box plots before and after outlier removal from paired reflectance samples provided in this embodiment of the invention. Figure 7 This is a Pearson correlation coefficient matrix diagram provided in an embodiment of the present invention; Figure 8 This is the feature vector contribution map provided in the embodiments of the present invention; Figure 9 This is a comparison chart of the fitting of TSM and SSC models provided in the embodiments of the present invention. Detailed Implementation

[0030] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0031] The innovation of this invention lies in: 1. Innovative application of intermediate variable TSM: This invention introduces remote sensing TSM products as intermediate variables. By establishing the TSM-SSC conversion relationship, it realizes indirect inversion under the condition of scarce field SSC samples, thereby significantly improving the inversion accuracy and long-term time series continuity of high turbidity water bodies, and providing a feasible solution for multi-year, long-term SSC monitoring from 2003 to 2024.

[0032] 2. Multi-source remote sensing data fusion and high-consistency sample construction: This invention uses MODIS surface reflectance data, GOCI-II TSM products, and multi-source remote sensing images such as Sentinel-2 and Landsat-7. Through radiometric correction, atmospheric correction, cloud and shadow masking, geometric registration, and spatiotemporal matching, a standardized paired sample set between highly consistent multi-source remote sensing features and field SSCs is established, overcoming the limitations of single sensors in spatial resolution, imaging frequency, and band response.

[0033] 3. Multi-type model training and optimization strategy: Based on feature construction, this invention systematically trains empirical regression models, chromatic angle models, ensemble learning models (such as random forests, gradient boosting, etc.) and deep learning models. Through cross-validation, the robustness, generalization ability and inversion accuracy of various models are comprehensively evaluated. Finally, the best performing model is selected as the basic SSC inversion model, which improves the overall inversion stability and accuracy.

[0034] 4. TSM-SSC Conversion Modeling and Optimization: This invention utilizes optimized TSM inversion results and field SSC observation data to construct various TSM-SSC conversion models, including linear, logarithmic, and power function models. The optimal conversion equation is determined through fitting accuracy evaluation, forming a stable and reliable TSM-SSC inversion system, and realizing quantitative conversion from TSM to SSC.

[0035] 5. Long-term batch inversion and spatiotemporal product generation: Based on the established TSM-SSC conversion model, this invention performs batch inversion on long-term MODIS remote sensing image sequences to generate SSC spatiotemporal distribution products at the annual, quarterly, and monthly scales. The spatiotemporal variation characteristics and inversion accuracy are systematically analyzed and verified, providing continuous and reliable data support for sediment dynamics research, waterway management, and ecological environment assessment.

[0036] This invention effectively solves the problem of insufficient generalization ability of traditional single models in complex marine systems by constructing a multi-source remote sensing-multi-model fusion SSC inversion technology system, realizing long-term, highly stable suspended sediment concentration inversion and spatiotemporal variation analysis, and providing reliable data support for marine sediment transport monitoring and ecological environment assessment.

[0037] Example 1, as Figure 1 As shown in the embodiments of the present invention, the remote sensing inversion method for suspended sediment concentration based on TSM-SSC conversion consists of data acquisition and preprocessing, feature extraction and variable construction, multi-model comparison and optimal model selection, TSM-SSC conversion establishment, long-term batch inversion and product generation, specifically including the following steps: S1. Multi-source remote sensing data preprocessing: Multi-source remote sensing image data is acquired on the Google Earth Engine cloud platform. Clouds and cloud shadows are removed from the multi-source remote sensing images based on the QA band of the remote sensing images. Surface reflectance correction processing is performed on the remote sensing images based on the scale factor and offset parameters in the image metadata. This invention first acquires MODIS surface reflectance time-series products (MOD09GA, MYD09GA, MOD09A1) as the primary source data for long-term SSC inversion. Simultaneously, it integrates Sentinel-2 and Landsat-7 auxiliary optical imagery and GOCI-II TSM products through the Google Earth Engine (GEE) platform to expand the training sample and compensate for insufficient field observations. Furthermore, it collects and organizes field SSC measurement data (examples are gravity-filtered measurements from 13 sampling stations), performing unified projection, atmospheric correction, cloud and shadow masking, negative and outlier removal, scale matching, and spatiotemporal registration on various remote sensing and field data. Through these processes, 947 spatial-temporal paired samples are initially generated. After quality screening and outlier removal, 602 high-quality samples are obtained for subsequent model training and validation. This step, through multi-source data fusion and a rigorous preprocessing procedure, ensures the physical consistency and temporal and spatial coverage of the training samples, laying a data foundation for high-precision SSC inversion, while also solving the spatiotemporal limitations of traditional single-source remote sensing data in long-term monitoring.

[0038] The multi-source remote sensing data includes surface reflectance data acquired by the MODIS satellite as the primary source data, and TSM products retrieved from the GOCI-II satellite and remote sensing image data acquired by the Sentinel-2 and Landsat-7 satellites as auxiliary data. The products and acquisition methods of the satellite data are shown in Table 1.

[0039] Table 1 Satellite Products and Acquisition Methods S2. Sample extraction at corresponding locations: Based on the surface reflectance remote sensing data acquired by multi-source satellites, select TSM products acquired on the same day with the same acquisition time; combine the spatial location information of the on-site SSC measurement station to extract the corresponding image data, construct a spatiotemporally consistent paired sample set, and obtain the corresponding sample data of surface reflectance and TSM value for each band. After data preprocessing, this invention constructs spectral features for remote sensing image pixels and paired samples. Specifically, this includes calculating derived indices such as the original multispectral bands, chromaticity angle α (based on the CIE-XYZ model), normalized difference vegetation index (NDVI), modified normalized water index (MNDWI), normalized difference index (NDTI), band reflectance ratio, and multi-band total reflectance. Subsequently, features are screened through correlation analysis with TSM / SSC and importance ranking of random forest variables. Abnormal samples are removed using box plots and outlier detection. Temporal and meteorological factors (such as tides and wind speed) can be optionally introduced to enhance the model's adaptability to short-term disturbances. The high-response, low-redundancy feature matrix constructed through this step significantly improves the model's generalization ability and effectively mitigates the impact of the non-uniqueness of the spectra of high-turbidity water bodies on SSC inversion.

[0040] The establishment of a high-quality paired sample set also includes: selecting TSM data corresponding to the acquisition time of the remote sensing images, and performing pixel-level pairing of multi-source remote sensing reflectance data with the TSM data based on the spatial location of the measured points of suspended sediment concentration in the field; performing quality control of the paired samples using statistical analysis methods, and identifying and removing outliers based on the interquartile range (IQR) criterion of the box plot, to obtain a high-quality paired sample set for modeling; such as Figure 6 As shown.

[0041] S3. Feature Construction and Screening: Based on the original band data of multispectral remote sensing images, calculate various derived spectral feature indices, correct the normalized water index MNDWI and chromaticity angle, and construct feature vectors for model training. After feature construction, this invention systematically trains and compares the performance of the SSC inversion model. Specific methods include constructing empirical statistical models (linear / exponential / power functions), chromaticity angle models, various machine learning models (decision trees, random forests, gradient boosting (GTB / XGBoost / LightGBM), and deep learning models (multilayer perceptrons, MLP). Five-fold cross-validation and training / validation set partitioning are employed to comprehensively evaluate the model's robustness, generalization ability, and inversion accuracy, using metrics including the coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE). Example results show that the random forest model, with an average R²≈0.75 under cross-validation, outperforms other candidate models and is therefore selected as the primary model for subsequent TSM inversion and SSC estimation. This step, through multi-model comparison and cross-validation strategies, significantly improves the stability of the inversion model and its applicability in complex aquatic systems, overcoming the problem of insufficient generalization ability of traditional single models.

[0042] The feature selection method includes: analyzing the correlation between each candidate feature and the target variable using the Pearson correlation coefficient, as shown in the correlation coefficient matrix. Figure 7 As shown; a feature importance evaluation method based on machine learning models ranks the contribution of each feature to the model's prediction results, and the contribution results of each feature vector are as follows. Figure 8 As shown; based on the correlation analysis results and feature importance ranking results, features with high correlation and large contribution are selected as the main input features for model training.

[0043] S4. TSM Inversion Model Training: Using feature vectors as model input and corresponding TSM values ​​as training targets, various machine learning regression algorithms are employed for model training, and the training is performed based on the evaluation metric R. 2 Select the optimal TSM remote sensing inversion model; To alleviate the scarcity of SSC samples in the field, this invention introduces remote sensing TSM products as intermediate variables. The TSM results obtained from the optimized inversion model are paired with observed SSC data to construct a TSM-SSC conversion sample set. Based on this, linear, logarithmic, and power function models are fitted, and the optimal conversion model is selected using indicators such as R² and RMSE (the example uses a linear model, R²≈0.80, RMSE≈12.10 mg / L). Abnormal samples that do not meet physical constraints (such as TSM being much smaller than SSC) are removed. This step achieves quantitative conversion from TSM to SSC, alleviating the problem of insufficient field samples, while ensuring the reliability and physical consistency of the inversion results under high turbidity water bodies and complex hydrodynamic conditions.

[0044] The machine learning regression algorithms include random forest regression model, gradient boosting decision tree model (GBDT), extreme gradient boosting model (XGBoost), and lightweight gradient boosting model (LightGBM); and are based on evaluation metrics (R... 2 The models are compared and analyzed to determine the preferred model for TSM remote sensing inversion.

[0045] Table 2 Statistical results of machine model fitting S5. TSM-SSC Transformation Modeling: Based on the TSM inversion results obtained from the TSM remote sensing inversion model and the corresponding field SSC observations, a TSM-SSC paired sample set is constructed. Modeling is performed on the paired samples using linear, logarithmic, and power functions, and the model is then evaluated based on the R-index. 2 Determine the optimal TSM-SSC transformation equation; S6. Long-term SSC batch inversion: Using the TSM remote sensing inversion model and the TSM-SSC transformation equation, batch inversion processing is performed on long-term remote sensing image data to generate spatiotemporal distribution products of suspended sediment concentration.

[0046] The trained and validated SSC inversion model and TSM-SSC conversion equation were applied in batches to MODIS full-time series data, generating SSC spatiotemporal products at annual, quarterly, and monthly scales from 2003 to 2024. Error statistics, spatial consistency analysis, and temporal series characteristic analysis were performed on the inversion results, including interannual mean, standard deviation, and seasonal patterns. Exemplary accuracy verification was conducted to evaluate the model's robustness and applicability under different tidal, storm, and seasonal conditions. Output products include annual average SSC distribution maps, quarterly and monthly SSC distribution sequences, inversion error statistics tables, and spatial difference analysis results. These can be directly used for regional sediment transport model research, ecological environment assessment, and engineering planning support, and can also be exported to common geographic information formats for further applications. This step achieved standardized batch processing of long-term SSC inversion and spatiotemporal products, solving the problem of traditional methods' difficulty in achieving long-term continuous monitoring and significantly improving the dynamic monitoring capability of suspended sediment in marine areas.

[0047] First, the physical consistency of sample pairs consisting of TSM inversion results and on-site SSC observations is screened, eliminating sample pairs where the TSM value is less than the corresponding SSC value. For the screened samples, the TSM-SSC transformation relationship is fitted using various functional forms, including linear, logarithmic, and power functions, and evaluated using evaluation indicators. R 2 The fitting results are analyzed to establish the TSM-SSC transformation equation; for example... Figure 9 As shown.

[0048] Example 2, the suspended sediment concentration remote sensing inversion system based on TSM-SSC conversion provided in this embodiment of the invention includes: The data preparation module acquires MODIS surface reflectance, GOCI-II TSM product and field SSC data, and performs radiometric correction, atmospheric correction, cloud and shadow masking, geometric registration and spatiotemporal matching on the data to form a standardized dataset. The feature extraction and sample construction module extracts original bands, chromaticity angles and multiple types of suspended matter index features from multi-source pixels, completes outlier removal, normalization and feature selection, and constructs training sample pairs. The model training and optimization module trains multiple SSC inversion models based on the sample set, uses cross-validation to evaluate stability and generalization ability, and selects the optimal model. The TSM-SSC conversion module establishes a TSM-SSC conversion function based on remote sensing TSM results and field SSC data to achieve concentration inversion. The long-term inversion module applies the optimized model to the MODIS image series from 2003 to 2024 to generate long-term SSC spatial distribution products.

[0049] like Figure 2 As shown in the figure, this invention provides a method for inverting long-term suspended sediment concentration based on multi-source remote sensing and multi-model fusion. The study area is Hangzhou Bay. The method includes the following steps: Data preparation: MODIS surface reflectance data, GOCI-II TSM products, and concurrent field SSC observation data were acquired. Radiometric correction, atmospheric correction, cloud and shadow masking, geometric registration, and spatiotemporal pairing were performed on each type of remote sensing data sequentially to construct high-quality training and validation sample sets with one-to-one correspondence between multi-source features and field SSCs. This step, through multi-source data fusion, addresses the issues of insufficient coverage and temporal discontinuity from a single sensor, providing standardized input data for subsequent model training. Feature extraction and sample construction: Original multispectral bands, chromaticity angles, and various suspended sediment sensitivity indices are extracted from pixels that have completed spatiotemporal registration. Outlier removal, standardization, and variable selection are performed on these features to construct an input feature matrix and target SSC paired samples for model training. Through feature selection, this invention significantly improves the stability and generalization ability of the high-turbidity water body inversion. Model Training and Multi-Model Comparison: Based on the constructed sample set, an empirical regression model, a chromaticity angle model, an ensemble learning model, and a deep learning model were trained respectively. The stability and accuracy of each model were evaluated using five-fold cross-validation. Experimental results show that the random forest model achieved an average R²≈0.75 and RMSE≈12.1 mg / L on the validation set, outperforming the traditional empirical model (average R²≈0.58) and the MLP model (average R²≈0.72), fully demonstrating the superior performance of the method of this invention under complex water conditions. TSM-SSC Conversion and Accuracy Verification: Linear, logarithmic, and power function conversion models were constructed using remote sensing TSM inversion results and field SSC observation data. Finally, the linear model was selected as the quantitative estimation formula for SSC, which ensured the physical consistency of the conversion and alleviated the limitation of scarce field samples. Long-term SSC inversion output: The constructed inversion model and TSM-SSC transformation equation were applied to MODIS full-time series imagery to generate annual, seasonal, and monthly SSC spatiotemporal products from 2003 to 2024. The inversion results show that the SSC in Hangzhou Bay exhibits a seasonal variation, with higher values ​​in winter and lower values ​​in summer, and shows a year-on-year trend. Error statistics and spatial consistency analysis indicate that the method of this invention maintains high stability and accuracy under long-term, different tidal, and seasonal conditions, providing reliable data support for sediment dynamic monitoring, ecological environment assessment, and engineering applications.

[0050] Among them, the long-term SSC inversion series for the Hangzhou Bay area from 2003 to 2024 is as follows: Figure 3 As shown, the SSC inversion results for each month are as follows: Figure 4 As shown, the SSC inversion results for each quarter are as follows: Figure 5 As shown.

[0051] To verify the system performance, a typical sea area in Hangzhou Bay was used as the experimental area. Multi-source remote sensing and field SSC paired samples were constructed, and the accuracy and stability of the empirical regression model, chromatic angle model, ensemble learning model and deep learning model were compared. The results are shown in Tables 3 and 4.

[0052] Table 3 Statistical results of mathematical model fitting Table 4 Statistical results of machine model fitting To evaluate the performance of the TSM-SSC conversion model, linear, logarithmic, and power function models were constructed respectively, and their fitting accuracy was compared. The statistical results are shown in Table 5.

[0053] Table 5. Statistical results of the accuracy of TSM and SSC fitting models (unit: mg / L) The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention and within the spirit and principles of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A remote sensing inversion method for suspended sediment concentration based on TSM-SSC conversion, characterized in that, The method includes the following steps: S1. Multi-source remote sensing data preprocessing: Multi-source remote sensing image data is acquired on the Google Earth Engine cloud platform. Clouds and cloud shadows are removed from the multi-source remote sensing images based on the QA band of the remote sensing images. Surface reflectance correction processing is performed on the remote sensing images based on the scale factor and offset parameters in the image metadata. S2. Sample extraction at corresponding locations: Based on the surface reflectance remote sensing data acquired by multi-source satellites, select TSM products acquired on the same day with the same acquisition time; combine the spatial location information of the on-site SSC measurement station to extract the corresponding image data, construct a spatiotemporally consistent paired sample set, and obtain the corresponding sample data of surface reflectance and TSM value for each band. S3. Feature Construction and Screening: Based on the original band data of multispectral remote sensing images, calculate various derived spectral feature indices, correct the normalized water index MNDWI and chromaticity angle, and construct feature vectors for model training. S4. TSM Inversion Model Training: Using feature vectors as model input and corresponding TSM values ​​as training targets, various machine learning regression algorithms are employed for model training, and the training is performed based on the evaluation metric R. 2 Select the optimal TSM remote sensing inversion model; S5. TSM-SSC Transformation Modeling: Based on the TSM inversion results obtained from the TSM remote sensing inversion model and the corresponding field SSC observations, a TSM-SSC paired sample set is constructed. Modeling is performed on the paired samples using linear, logarithmic, and power functions, and the model is then evaluated based on the R-index. 2 Determine the optimal TSM-SSC transformation equation; S6. Long-term SSC batch inversion: Using the TSM remote sensing inversion model and the TSM-SSC transformation equation, batch inversion processing is performed on long-term remote sensing image data to generate spatiotemporal distribution products of suspended sediment concentration.

2. The method for remote sensing inversion of suspended sediment concentration based on TSM-SSC conversion according to claim 1, characterized in that, In step S1, the multi-source remote sensing data includes surface reflectance data acquired by the MODIS satellite as the main source data, and total suspended matter concentration (TSM) products obtained by inversion from the GOCI-II satellite, combined with surface reflectance remote sensing image data acquired by the Sentinel-2 and Landsat-7 satellites as auxiliary data; all multi-source remote sensing data are obtained through the Google Earth Engine cloud computing platform or official data release channels.

3. The method for remote sensing inversion of suspended sediment concentration based on TSM-SSC conversion according to claim 1, characterized in that, In step S1, the multi-source remote sensing data preprocessing also includes: The QA bitmask information corresponding to different data sources was analyzed to remove clouds, cloud shadows and invalid pixels; remote sensing images with different spatial resolutions were resampled to a unified spatial scale and unified to a projection coordinate system consistent with the measured data of suspended sediment concentration in the field to ensure the spatial consistency of multi-source remote sensing data.

4. The method for remote sensing inversion of suspended sediment concentration based on TSM-SSC conversion according to claim 1, characterized in that, In step S2, a spatiotemporally consistent paired sample set is constructed, including: Based on the acquisition time of the remote sensing images, TSM data corresponding to the time phase are selected. Based on the spatial location of the measured points of suspended sediment concentration in the field, multi-source remote sensing reflectance data and the TSM data are paired at the pixel level. Statistical analysis methods are used to control the quality of the paired samples, and outliers are identified and removed based on the interquartile range (IQR) criterion of the box plot to obtain a paired sample set for modeling.

5. The method for remote sensing inversion of suspended sediment concentration based on TSM-SSC conversion according to claim 1, characterized in that, In step S3, the derived spectral characteristic index includes the following indicators: Normalized Difference Vegetation Index (NDVI): ； Modified Normalized Difference Water Index (MNDWI): ； Normalized Difference Index (NDSI): ； The ratio of red light to near-infrared reflectance: ； Total reflectance (TotalRef): ； In the formula, These are the remote sensing reflectances in the red, green, near-infrared, and near-shortwave infrared bands, respectively. For the first multispectral remote sensing image Reflectivity of each band This represents the total number of bands.

6. The method for remote sensing inversion of suspended sediment concentration based on TSM-SSC conversion according to claim 1, characterized in that, In step S3, feature selection includes: The correlation between each candidate feature and the target variable was analyzed using the Pearson correlation coefficient. Based on the feature importance evaluation method of the machine learning model, the contribution of each feature to the model prediction results was ranked to obtain the contribution results of each feature vector. Based on the correlation analysis results and feature importance ranking results, features with high correlation and large contribution are selected as the main input features for model training.

7. The method for remote sensing inversion of suspended sediment concentration based on TSM-SSC conversion according to claim 1, characterized in that, In step S4, the machine learning regression algorithm includes a random forest regression model, a gradient boosting decision tree model, an extreme gradient boosting model, and a lightweight gradient boosting model; based on the evaluation metric R... 2 By comparing and analyzing the models, the preferred model for TSM remote sensing inversion was determined.

8. The method for remote sensing inversion of suspended sediment concentration based on TSM-SSC conversion according to claim 1, characterized in that, In step S5, the physical consistency of the sample pairs formed by the TSM inversion results and the on-site SSC observations is first screened, and sample pairs with TSM values ​​less than the corresponding SSC values ​​are removed. For the screened samples, the TSM-SSC transformation relationship is fitted using various functional forms such as linear functions, logarithmic functions, and power functions, and the evaluation index R is used. 2 The fitting results were analyzed, and the TSM-SSC transformation equation was established.

9. The method for remote sensing inversion of suspended sediment concentration based on TSM-SSC conversion according to claim 1, characterized in that, In step S6, a spatiotemporal product of SSC is generated, which includes SSC inversion sequence diagrams based on different time scales, including annual, quarterly, or monthly scales; the spatiotemporal product also includes corresponding statistical feature diagrams, which include the mean and standard deviation, used to represent the spatiotemporal distribution and trend of SSC.

10. A remote sensing inversion system for suspended sediment concentration based on TSM-SSC conversion, characterized in that, The system is implemented using the remote sensing inversion method for suspended sediment concentration based on TSM-SSC conversion as described in any one of claims 1-9. The system comprises: The data acquisition and preprocessing module is used to acquire remote sensing image data from multiple remote sensing data sources such as MODIS, GOCI-II, Sentinel-2 and Landsat-7, and to perform quality control and preprocessing operations on the remote sensing image data. The preprocessing operations include cloud and cloud shadow removal, surface reflectance correction, geometric correction and spatial resampling processing to obtain remote sensing reflectance data that meets the requirements of subsequent analysis. The sample pairing and screening module is used to pair multi-source remote sensing reflectance data, TSM product data and on-site measured suspended sediment concentration data based on the acquisition time and spatial location of remote sensing images, and to perform quality control and outlier screening on the paired samples to obtain high-quality sample data that are consistent in time and space. The feature construction and screening module is used to extract original band features based on the remote sensing reflectance data and calculate the derived spectral feature index, and screen the features through correlation analysis and feature importance evaluation to generate feature vectors for model training. The model training and evaluation module is used to train the TSM remote sensing inversion model based on the feature vector and the corresponding TSM sample values, and to evaluate and compare the inversion performance of different models in order to determine the optimal model for TSM batch inversion. The TSM-SSC conversion module is used to construct a conversion sample based on the inversion results of the TSM remote sensing inversion model and the measured data of suspended sediment concentration in the field, and to establish the conversion relationship between TSM and SSC, so as to realize the quantitative conversion of TSM to SSC. The batch inversion and product generation module is used to apply the TSM remote sensing inversion model and TSM-SSC conversion relationship to long-term remote sensing image data, perform batch inversion processing of suspended sediment concentration, and generate spatiotemporal distribution products of suspended sediment concentration at different time scales. The results display and export module is used to visualize the suspended sediment concentration inversion results and output raster products and related statistical analysis results.

Images