A turbidity extraction method and device based on remote sensing data

Abstract

The application provides a turbidity extraction method and device based on remote sensing data, and belongs to the field of satellite remote sensing water quality monitoring. The method comprises the following steps: obtaining the atmosphere-corrected remote sensing data of a target sea area and the to-be-fitted turbidity values of a plurality of sampling points; screening a plurality of target remote sensing bands according to the fitting result of the single-band remote sensing reflectance data in the atmosphere-corrected remote sensing data and the to-be-fitted turbidity values; determining turbidity prediction model samples according to the fitting result of the combination of any two target remote sensing bands and the to-be-fitted turbidity values; modifying the band composition in the turbidity prediction model samples, determining a target turbidity prediction model according to the fitting result of the modified instantiated turbidity prediction model and the to-be-fitted turbidity values; and predicting the real-time turbidity of the target sea area according to the real-time remote sensing data of the target sea area and the target turbidity prediction model. The method and device provided by the application can solve the problems of insufficient single-band inversion accuracy, great influence of interference factors such as aerosols, and limited adaptability of existing models in the near-shore sea area.

Description

Technical Field

[0001] This application relates to the field of satellite remote sensing water quality monitoring technology, and in particular to a method and apparatus for turbidity extraction based on remote sensing data. Background Technology

[0002] Seawater turbidity is a key optical parameter characterizing the scattering and absorption of light by suspended particulate matter in water bodies. Its quantitative inversion is of great significance for marine environmental monitoring, estuarine sediment transport research, aquaculture area planning, and eutrophication assessment. Traditionally, turbidity monitoring mainly relies on shipboard surveys or on-site measurements at fixed stations. Although this method has high accuracy, it is time-consuming, labor-intensive, and costly, and it is difficult to achieve large-scale, synchronous marine monitoring, especially in the environmentally variable nearshore waters, where its spatiotemporal representativeness is limited.

[0003] Satellite remote sensing technology offers advantages such as wide coverage, short revisit cycles, and traceability of historical data, providing an effective means for large-scale, dynamic monitoring of marine turbidity. Currently, turbidity inversion methods based on remote sensing data are mostly semi-analytical methods based on aquatic bio-optical models or empirical algorithms based on statistical regression. Semi-analytical methods based on aquatic bio-optical models achieve parameter inversion by physically modeling the absorption and backscattering characteristics of various water components, but their models are complex, require high accuracy of input parameters, and lack stability in Class II water bodies with complex optical characteristics (especially nearshore waters). Empirical algorithms based on statistical regression directly establish the mathematical relationship between remotely sensed reflectance or its derivatives and measured turbidity values, featuring simplicity and high computational efficiency, and are therefore widely used in practice.

[0004] However, existing statistical regression-based turbidity remote sensing inversion methods still face several challenges when applied to nearshore waters. First, the accuracy of single-band inversion models is generally insufficient. The optical signal of nearshore waters is a mixed signal resulting from the combined effects of multiple components, including phytoplankton, non-algal particles, and colored dissolved organic matter. Relying solely on a single band is insufficient to effectively capture the spectral characteristics of turbidity changes, leading to poor model robustness and limited applicability in different regions or seasons. Second, atmospheric interference factors such as aerosols have a significant impact. The types of aerosols over nearshore waters are complex and vary spatiotemporally, and their scattering contribution is a crucial component of the remote sensing signal. Although L2A-level data has undergone Rayleigh scattering correction, residual aerosol scattering effects, especially in the blue light band, still severely interfere with the extraction of water-leaving radiation signals, thus introducing significant turbidity inversion errors. Finally, the adaptability of existing empirical models is limited. Most turbidity inversion models are constructed based on data from specific sensors in specific regions and time periods, and their band combinations and model coefficients have strong scene dependence. When applied directly to different sensors or different spatiotemporal scenarios of the same sensor, the model performance often drops sharply due to differences in sensor spectral response functions and changes in regional water color and atmospheric conditions, lacking the necessary generalization ability.

[0005] Therefore, it is necessary to study a turbidity inversion method that can achieve high accuracy and high stability in complex optical and atmospheric environments near the coast, and has good adaptability across sensors and time and space. Summary of the Invention

[0006] In view of this, this application provides a turbidity extraction method and apparatus based on remote sensing data, which can solve the problems of insufficient single-band inversion accuracy in nearshore waters, significant influence of interference factors such as aerosols, and limited adaptability of existing models.

[0007] Specifically, this application is implemented through the following technical solution:

[0008] The first aspect of this application provides a turbidity extraction method based on remote sensing data, the method comprising:

[0009] Acquire atmospheric-corrected remote sensing data of the target sea area and turbidity values ​​to be fitted from multiple sampling points;

[0010] Multiple target remote sensing bands are selected based on the first fitting result of the single-band remote sensing reflectance data in the atmospheric corrected remote sensing data and the turbidity value to be fitted.

[0011] A turbidity prediction model example is determined based on the combination of any two target remote sensing bands and the second fitting result of the turbidity value to be fitted.

[0012] Modify the band composition in the turbidity prediction model example, and determine the target turbidity prediction model based on the modified instantiated turbidity prediction model and the second fitting result of the turbidity value to be fitted.

[0013] The real-time turbidity of the target sea area is predicted based on real-time remote sensing data of the target sea area and the target turbidity prediction model.

[0014] A second aspect of this application provides a turbidity extraction device based on remote sensing data, the device comprising an acquisition module, a filtering module, a determination module, and a prediction module;

[0015] The acquisition module is used to acquire atmospheric-corrected remote sensing data of the target sea area and turbidity values ​​to be fitted from multiple sampling points;

[0016] The filtering module is used to filter multiple target remote sensing bands based on the first fitting result of the single-band remote sensing reflectance data in the atmospherically corrected remote sensing data and the turbidity value to be fitted.

[0017] The determining module is used to determine a turbidity prediction model example based on the combination of any two target remote sensing bands and the second fitting result of the turbidity value to be fitted.

[0018] The determining module is also used to modify the band composition in the turbidity prediction model example, and determine the target turbidity prediction model based on the modified instantiated turbidity prediction model and the second fitting result of the turbidity value to be fitted.

[0019] The prediction module is used to predict the real-time turbidity of the target sea area based on the real-time remote sensing data of the target sea area and the target turbidity prediction model.

[0020] The turbidity extraction method and apparatus based on remote sensing data provided in this application effectively improve the accuracy and stability of turbidity inversion in optically complex nearshore waters by constructing a target turbidity prediction model. Specifically, by selecting the target remote sensing band with the strongest correlation to turbidity from all available bands, interference from noise bands that are significantly affected by aerosols and have weak turbidity responses in nearshore waters is effectively eliminated, enhancing the robustness of the model in complex atmospheric environments. By performing pairwise combinations and mathematical relationship evaluations of the target bands, turbidity prediction model examples with ratio relationships are automatically established, which can effectively capture the spectral morphological characteristics of suspended particulate matter in nearshore waters, overcoming the deficiency of single-band models in representing Class II water bodies. Furthermore, by modifying the band composition within the ratio framework, more explanatory and predictive optimized band combinations and function forms are found, thereby constructing a target turbidity prediction model with optimal performance, enabling it to better adapt to the differences in the optical characteristics of water bodies in different nearshore areas. Attached Figure Description

[0021] Figure 1 A flowchart of Embodiment 1 of the turbidity extraction method based on remote sensing data provided in this application;

[0022] Figure 2 This is a schematic diagram of the structure of Embodiment 2 of the turbidity extraction device based on remote sensing data provided in this application. Detailed Implementation

[0023] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0024] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used herein are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0025] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0026] The following specific embodiments are given to illustrate the technical solution of this application in detail.

[0027] Example 1

[0028] Figure 1 This is a flowchart of an embodiment of the turbidity extraction method based on remote sensing data provided in this application. Please refer to... Figure 1 The method provided in this embodiment may include:

[0029] S101. Obtain atmospheric-corrected remote sensing data of the target sea area and turbidity values ​​to be fitted from multiple sampling points.

[0030] It should be noted that the target sea area refers to a specific coastal area or inland water body that requires turbidity monitoring using remote sensing technology. The selection criteria include, but are not limited to, environmental monitoring needs such as estuaries, ports, and aquaculture areas, hydrological characteristics such as water mixing degree and water depth, and areas with frequent historical turbidity events. For example, it is generally a nearshore sea area that is significantly affected by land-based inputs and has a complex optical composition of water.

[0031] Atmospherically corrected remote sensing data refers to surface reflectance data obtained by atmospheric correction of raw satellite observation data. Raw satellite signals contain scattering and absorption interference from atmospheric molecules and aerosols. The core purpose of atmospheric correction is to eliminate these atmospheric contributions and obtain the water-leaving radiance or remote sensing reflectance that truly represents the characteristics of the water body. Atmospherically corrected remote sensing data is the foundation of all ocean parameter products. Specifically, due to the weak water body signal, only through accurate atmospheric correction can the subsequently established model accurately reflect the true physical relationship between water body components (such as suspended sediment) and spectral characteristics. It should be noted that this atmospherically corrected remote sensing data can originate from various satellite sensors with ocean color observation capabilities, such as HY-1C / D CZI, Landsat OLI, and Sentinel-2 MSI. The data level is typically L2A. This application uses data acquired by the HY-1C / D CZI satellite sensor as an example.

[0032] It should be noted that multiple sampling points refer to geographical locations within the target sea area that are measured synchronously or quasi-synchronously with the satellite transit (usually within ±0.5 hours). The selection of sampling points generally needs to cover different turbidity gradients within the target sea area (e.g., high turbidity areas nearshore, low turbidity areas offshore) and key hydrological regions. The number of sampling points must also meet statistical requirements, typically no less than 30, to ensure the robustness of model training. Furthermore, when selecting sampling points, severe weather events such as rainfall and strong winds should be avoided. Since turbidity identification is also affected by aerosols and high cloud cover, and L2A data only undergoes Rayleigh scattering correction, aerosol scattering is significant in coastal waters. Therefore, to minimize the impact of aerosols and high cloud cover, HY-1C / D CZI satellite data obtained during clear weather should be selected to help increase the use of high-resolution satellite remote sensing images for turbidity observation.

[0033] It should be further noted that aerosol scattering is one of the key factors affecting the accuracy of remote sensing reflectance in nearshore waters. The L2A-level atmospheric-corrected remote sensing data used in this application mainly corrects for Rayleigh scattering and does not completely eliminate the influence of aerosol scattering. The presence of aerosols alters the spectral signal arriving at the sensor, causing a systematic shift in the extracted single-band remote sensing reflectance data. If data severely affected by aerosols is used for modeling or prediction, it will directly introduce fitting and inversion errors into the target turbidity prediction model, leading to inaccuracies in model coefficients or a decrease in the final turbidity prediction accuracy. Therefore, it is necessary to establish a unified data quality standard for all input remote sensing data. That is, priority should be given to satellite imagery acquired under conditions where the aerosol optical thickness is below a preset threshold (wherein the preset threshold is determined based on practical experience, please refer to the description of the relevant technology for details, which will not be repeated here), and the imaging was performed in clear and dry weather (here, clear and dry weather refers to the specified weather type). This standard is based on the understanding that the concentration and type of aerosols mainly vary with meteorological conditions (such as humidity, wind speed, and weather systems). By screening high-quality images with low aerosol load and high atmospheric visibility from the source, this source of error can be controlled to the greatest extent, providing a reliable data foundation for high-precision turbidity inversion.

[0034] It should also be noted that turbidity quantitatively characterizes the degree of opacity of water bodies caused by the scattering and absorption of incident light by suspended particulate matter (such as silt, algae, and organic debris), and its unit is usually the nanometer turbidity unit (NTU). Turbidity itself is a comprehensive apparent optical quantity, closely related to the concentration, particle size distribution, composition, and color of suspended particulate matter in water. In environmental monitoring, high turbidity usually indicates a high concentration of suspended sediments, which can affect underwater illumination and primary productivity, and may carry pollutants. In field measurements, specific turbidity readings are obtained at discrete points. This process can be directly measured at each sampling point using on-site instruments (such as turbidimeters). The turbidity value to be fitted here refers to these measured turbidity data, which will be used as the dependent variable in subsequent steps to mathematically fit with remotely sensed reflectance data, which is the independent variable, thereby establishing a quantitative relationship model between the two. The quality of the turbidity value to be fitted directly determines the accuracy and reliability of the final model prediction results.

[0035] In practice, atmospherically corrected remote sensing data can be directly downloaded from the official data platform of the corresponding satellite, based on the target sea area and imaging time. This data has been officially processed and can be used directly.

[0036] Furthermore, after obtaining the atmospherically corrected remote sensing data of the target sea area and the turbidity values ​​to be fitted from multiple sampling points, data matching and association are required. Specifically, the geographic coordinates of each sampling point are spatially registered with the atmospherically corrected remote sensing image, and the multi-band remote sensing reflectance data corresponding to that coordinate location is extracted. Finally, a structured dataset is formed, where each row represents a sampling point, including its latitude and longitude, the measured turbidity value to be fitted, and the reflectance values ​​of multiple bands extracted from the remote sensing image. Through this data matching, the spatially distributed remote sensing image and the discretely measured field sampling are linked.

[0037] S102. Select multiple target remote sensing bands based on the first fitting result of the single-band remote sensing reflectance data in the atmospheric corrected remote sensing data and the turbidity value to be fitted.

[0038] It should be noted that single-band remote sensing reflectance data refers to the reflectance values ​​of ground objects acquired by satellite sensors within a specific spectral range. Each band corresponds to a continuous spectral interval, capable of capturing the reflectance characteristics of water bodies at different wavelengths. The first fitting result refers to the result obtained after correlation analysis at the single-band level, used to screen characteristic bands. Taking the HY-1C / D CZI sensor as an example, Table 1 shows the band parameters of the HY-1C / D CZI sensor as illustrated in the application:

[0039] Table 1. Band parameters of HY-1C / D CZI sensor

[0040]

[0041] Please refer to Table 1. In this embodiment of the application, the single band includes at least the wavelength range data of the blue (B1), green (B2), red (B3), and near-infrared (B4) bands. These bands cover the key spectral range from visible light to near-infrared, providing a complete spectral basis for screening the bands most sensitive to water turbidity response.

[0042] It should be noted that the fitting result here refers to the correlation coefficient, an index that quantifies the strength of the linear relationship between single-band reflectance data and the turbidity value to be fitted, using mathematical methods. It is a value between -1 and 1; the larger the absolute value, the stronger the linear relationship. The target remote sensing bands refer to multiple bands that, after screening, are confirmed to have a significant correlation with water turbidity. Specifically, all bands are sorted from highest to lowest according to their correlation coefficient with turbidity, and a predetermined number (the top n) of bands with correlation coefficients higher than a preset threshold are selected.

[0043] Specifically, based on the fitting results between the single-band remote sensing reflectance data in the atmospherically corrected remote sensing data and the turbidity value to be fitted, multiple target remote sensing bands are selected, including:

[0044] (1) Calculate the correlation coefficient between each single-band remote sensing reflectance data and the turbidity value to be fitted.

[0045] It should be noted that this step operates on each individual band in the dataset. The reflectance value of each band across all sampling points is treated as one variable, and the corresponding turbidity value to be fitted is treated as another variable. The Pearson correlation coefficient between the two is then calculated. The Pearson correlation coefficient measures the strength and direction of the linear relationship between two variables. Its calculation is based on the covariance and standard deviation of the variables; the specific formula can be found in the relevant technical descriptions and will not be repeated here. This calculation allows each band to be assigned a quantified score for subsequent comparisons.

[0046] In addition to calculating the linear Pearson correlation coefficient, distance correlation can also be used. Distance correlation can capture any type of dependency, including nonlinear ones, and a value of 0 indicates that the variables are independent. When a nonlinear but monotonic or non-monotonic dependency between turbidity and reflectance is suspected, distance correlation can serve as an effective supplement or alternative to Pearson correlation, ensuring that important sensitive bands are not overlooked.

[0047] Specifically, Table 2 shows the correlation coefficients of remote sensing reflectance data for each single band as presented in this application:

[0048] Table 2 Correlation coefficients of remote sensing reflectance data for each single band

[0049]

[0050] Please refer to Table 2, which exemplarily illustrates the correlation coefficients between the four bands (B1 to B4) and the turbidity values ​​to be fitted. The data in the table shows that the sensitivity of different bands to turbidity varies significantly. For example, B4 (near-infrared band) and B3 (red band) exhibit high correlations, indicating a stronger linear relationship between the spectral reflectance of these two bands and water turbidity; while B1 (blue band) shows a low correlation, suggesting that it is not suitable as a sole indicator of turbidity.

[0051] (2) For all sampling points, calculate the average correlation coefficient between each single band and the turbidity value to be fitted, and sort the bands from high to low according to the average correlation coefficient.

[0052] It should be noted that the average correlation coefficient here is actually the overall correlation coefficient for each band obtained in the previous step. Therefore, it is only necessary to arrange all bands according to their correlation coefficient with turbidity from high to low. Please refer to Table 2 for the order B4 > B3 > B2 > B1.

[0053] (3) Select the single band that ranks first and whose average correlation coefficient is greater than the preset threshold as the target remote sensing band.

[0054] It should be noted that the preset threshold is a value pre-set based on experience or statistical data. It is used to determine whether the correlation is strong enough, filtering out bands with a weak absolute correlation to turbidity. For example, a preset threshold of 0.5 can be set, then bands with a correlation below 0.5 will be eliminated. Please refer to Table 2. After filtering, the final target remote sensing bands are B4, B3, and B2. This operation ensures that the bands used in the final model have a clear and significant correlation with turbidity, thus guaranteeing the quality of the subsequently built model.

[0055] S103. Determine a turbidity prediction model example based on the combination of any two target remote sensing bands and the second fitting result of the turbidity value to be fitted.

[0056] It should be noted that the combination of any two target remote sensing bands refers to the mathematical combination of any two selected target remote sensing bands mentioned above, forming a new composite variable. This allows for the exploration of the synergistic effect between the spectral information carried by the two bands. The second fitting result refers to the performance evaluation indicators (such as goodness of fit (i.e., coefficient of determination) and error indicators) obtained after fitting at the model level, which differ from the first fitting result.

[0057] Specifically, based on the second fitting result of any combination of two target remote sensing bands and the turbidity value to be fitted, a turbidity prediction model example is determined, including:

[0058] (1) Pair the selected target remote sensing bands together.

[0059] It should be noted that pairing refers to generating all possible unordered band pairs from the selected target remote sensing bands according to the principles of combinatorial mathematics. For example, if the target bands are {B2, B3, B4}, then the pairing results are (B2, B3), (B2, B4), (B3, B4).

[0060] (2) For each pair of bands, construct multiple mathematical relationships to form multiple candidate band combinations. Each mathematical form is used to perform numerical calculations on the two target remote sensing bands in each pair of bands.

[0061] Mathematical relationships refer to the basic calculation methods between reflectance values ​​of two bands. They mainly include ratio relationships, such as B3 / B2, which can enhance the morphological differences of spectral curves of different ground objects; difference relationships, such as B3-B2, which can highlight the absorption or reflection characteristics of a specific spectral range; and sum relationships, such as B3+B2, which reflect the combined reflection energy of the two bands.

[0062] In practice, the previously selected target remote sensing bands are paired up in pairs and iterated through to generate all possible band pairs. Predefined mathematical operations are then applied to each band pair to obtain multiple candidate band combinations that meet the requirements (as shown in the example above). This overcomes the limitation of the limited expressive power of a single band feature and avoids the subjectivity of manually selecting band combinations based on experience.

[0063] (3) Fit each candidate band combination with the turbidity value to be fitted using a model.

[0064] (4) Compare the goodness of fit and error index of all candidate band combinations, and select the mathematical relationship corresponding to the candidate band combination with the highest goodness of fit and the smallest error index as the example of the turbidity prediction model.

[0065] It should be noted that the goodness-of-fit index mainly refers to the coefficient of determination (R²). 2 The value of 0 reflects the model's ability to explain changes in turbidity and ranges from 0 to 1. A larger value indicates a better model fit. The error index mainly refers to the root mean square error (RMSE), which reflects the average deviation between the model's predicted value and the actual value. A smaller value indicates a higher model accuracy.

[0066] Specifically, a regression model with turbidity is established for each candidate band combination, and the performance evaluation index (R²) of each regression model is calculated. 2 (and RMSE). Furthermore, the results can be sorted to select the candidate band combination corresponding to the performance evaluation index with the best overall performance as a turbidity prediction model example.

[0067] Specifically, the goodness of fit and error indices of multiple candidate band combinations are shown in Table 3:

[0068] Table 3. Combinations of multiple candidate bands and their corresponding goodness of fit and error indices

[0069]

[0070] Please refer to Table 3, which exemplarily shows the performance evaluation results of regression fitting of different candidate band combinations with turbidity. By comparing the data in the table, it can be clearly seen that the ratio relationship combination performs better in both the goodness of fit and error index dimensions. Therefore, based on the requirement of the highest goodness of fit and the smallest error index, the turbidity prediction model example can be determined to be the ratio relationship.

[0071] S104. Modify the band composition in the turbidity prediction model example, and determine the target turbidity prediction model based on the modified instantiated turbidity prediction model and the second fitting result of the turbidity value to be fitted.

[0072] It should be noted that band composition refers to the specific bands involved in the calculation (numerator, denominator, or additional term) in the turbidity prediction model example determined in S103. For example, if the turbidity prediction model example is B3 / B2, its band composition is B3 (red light band) as the numerator, B2 (green light band) as the denominator, and 0 as the additional term. Modifying the band composition refers to replacing or complicating the numerator, denominator, or both, while maintaining the ratio relationship of the turbidity prediction model example, and introducing additional terms to enhance the model's expressive power. This includes replacing a single band with another band, replacing a single band with a combination of multiple bands, and simultaneously adjusting the composition of the numerator and denominator. Furthermore, it also includes adding an additional term independent of the ratio structure after adjusting the numerator and denominator. This additional term can be a constant (bias term) or another single-band or multi-band combination reflectance value to compensate for systematic bias or capture additional spectral information. An instantiated turbidity prediction model refers to a specific turbidity prediction model generated after each modification to the band composition. For example, modifying the molecule of B3 / B2 to obtain B4 / B2 is an instantiated turbidity prediction model. The target prediction model is the mathematical model ultimately selected for actual turbidity prediction. It consists of the optimized band combination, the optimal functional form, and its fitting coefficients. It is the model that best fits the second-best turbidity value to the turbidity value being fitted among all tested instantiated turbidity prediction models (i.e., the model with the highest goodness of fit and the smallest error).

[0073] Specifically, the band composition in the turbidity prediction model example, and the determination of the target turbidity prediction model based on the modified instantiated turbidity prediction model and the second fitting result of the turbidity value to be fitted, include:

[0074] (1) Based on the first fitting result, candidate bands are determined, and the numerator, denominator and additional terms of the turbidity prediction model are adjusted according to the candidate bands to construct a subset of instantiated turbidity prediction models.

[0075] It should be noted that the construction methods of the various instantiated turbidity prediction model subsets include, but are not limited to, subsets formed by modifying the denominator while keeping the numerator of the ratio unchanged (e.g., modifying B3 / B2 to B3 / B1), subsets formed by modifying the numerator while keeping the denominator unchanged (e.g., modifying B3 / B2 to B4 / B2), and subsets formed by linearly or nonlinearly combining the output of the ratio with the reflectance of other bands (e.g., generating B3 / B2+B4 or (B3 / B2)*B1), etc.

[0076] It should be noted that the purpose of this step is to explore all possible and better-performing band combinations within the framework of ratio relationships.

[0077] (2) Fit all models in the subset of instantiated turbidity prediction models to the turbidity value to be fitted. From all the second fitting results, select the instantiated turbidity prediction model with the highest fitting goodness and the smallest error, and determine the corresponding remote sensing band combination as the optimized band combination.

[0078] Specifically, a uniform evaluation is performed on all models in the subset of instantiated turbidity prediction models, and the R-value of each model is calculated. 2 Using the RMSE index, the instantiated turbidity prediction model with the highest goodness of fit and smallest error is selected, and the corresponding remote sensing band combination is determined as the optimal band combination. This step fits the relationship between the band combination and turbidity.

[0079] (3) The numerical value of the optimized band combination is used as the independent variable and is fitted with a variety of preset function forms respectively. From the fitting results of all function forms, the function form with the highest fitting degree and the smallest error and the corresponding coefficient are selected to form the target turbidity prediction model.

[0080] It should be noted that after determining the optimal band combination (i.e., the input feature variable X of the model, such as X = B3 / B2 + B4), the most accurate mathematical mapping relationship between it and turbidity (Y) is still unknown. Therefore, it is necessary to use the determined band combination value as the independent variable X and attempt to fit it with various preset function forms (such as linear, exponential, power functions, etc.). By exploring the potential nonlinear relationship between the independent variable X and the dependent variable Y in this way, and by selecting the function form and corresponding coefficients with the highest goodness of fit and the smallest error, the target turbidity prediction model is finally constructed.

[0081] Table 4 shows the performance evaluation results of different instantiated turbidity prediction models (i.e., different band combinations X) shown in this application after model fitting with measured turbidity values. Specifically:

[0082] Table 4 Performance evaluation results after fitting

[0083]

[0084] Please refer to Table 4, which aims to compare the performance of different band combinations to select the optimal remote sensing band combination (i.e., independent variable X). The comparison shows that the combination B3 / B2+B4 performs best and is therefore determined as the optimized band combination, i.e., X=B3 / B2+B4. After determining the optimal band combination X, the value of X=B3 / B2+B4 is further used as the independent variable and fitted with various preset function forms (such as linear functions, exponential functions, power functions, etc.). This reveals that the exponential function form achieves the best fit for this band combination.

[0085] Furthermore, the band composition in the turbidity prediction model example, and the determination of the target turbidity prediction model based on the modified instantiated turbidity prediction model and the second fitting result of the turbidity value to be fitted, also include:

[0086] (1) Keeping the numerator of the turbidity prediction model sample unchanged, the bands in the atmospheric corrected remote sensing data are arbitrarily combined to modify the denominator of the turbidity prediction model sample and generate the first instantiated turbidity prediction model subset.

[0087] It should be noted that arbitrary combinations include, but are not limited to, single-band replacement, multi-band linear combinations, and nonlinear combinations. This step is based on the principle of local search optimization, exploring the optimal configuration under the premise of fixing part of the model structure. By keeping the numerator unchanged, it explores all possible combinations of the denominator to ensure that no potential advantageous combinations are overlooked.

[0088] In practical implementation, for a turbidity prediction model example (such as B3 / B2), the denominator is replaced by fixing the numerator B3, including single-band replacements B3 / B1, B3 / B4, and multi-band combinations such as B3 / (B1+B2) and B3 / (B4-B1). This step finds the current optimal denominator by fixing the numerator, for example, obtaining the single-band optimal value.

[0089] (2) Keeping the denominator of the turbidity prediction model sample unchanged, the bands in the atmospheric corrected remote sensing data are arbitrarily combined to modify the numerator of the turbidity prediction model sample and generate a second instantiated turbidity prediction model subset.

[0090] This step forms a symmetrical search with step (1) to ensure the integrity of the numerator optimization. For details, please refer to the description of step (1), which will not be repeated here. This step fixes the current optimal denominator, and determines the optimal numerator that matches the denominator by traversing each band in the candidate set as the numerator and evaluating its fitting effect with turbidity.

[0091] (3) Based on the optimal form of the first instantiated turbidity prediction model subset and the first instantiated turbidity prediction model subset, modify the numerator and denominator terms simultaneously to generate the third instantiated turbidity prediction model subset.

[0092] (4) Fit all models in the first, second and third subsets of instantiated turbidity prediction models to the turbidity value to be fitted. From all the second fitting results, select the instantiated turbidity prediction model with the highest fitting goodness and the smallest error, and determine the corresponding remote sensing band combination as the candidate optimized band combination.

[0093] (5) Add a band as an additional item based on the selected optimized band combination, and fit it with the turbidity value to be fitted. From all the fitting results, select the instantiated turbidity prediction model with the highest fitting goodness and the smallest error, and determine the corresponding remote sensing band combination as the final optimized band combination.

[0094] (6) Fit the values ​​of the optimized band combination to various preset function forms respectively. From the fitting results of all function forms, select the function form with the highest fitting degree and the smallest error and the corresponding coefficient to form the target turbidity prediction model.

[0095] The process of determining the optimal band combination B3 / B2+B4 described above is actually a screening process. First, among many simple mathematical relationships between pairs of bands (as shown in Table 3), the ratio B3 / B2 was initially determined as a high-performing turbidity prediction model example. Next, by modifying the band composition of this example (such as adding an additional term B4 to the ratio), a series of more complex instantiated models, as listed in Table 4, were generated and tested. By comparing the fitting indices of all models in Table 4, based on the principle of highest goodness of fit and lowest error, B3 / B2+B4 was ultimately determined as the optimal band combination.

[0096] Specifically, the calculation of the target turbidity prediction model includes: based on the band composition of the target turbidity prediction model, extracting the corresponding band reflectance from the atmospherically corrected remote sensing data and calculating the combined value; calculating a quadratic polynomial of the combined value, the quadratic polynomial including a quadratic term, a linear term and a constant term; performing a power operation with 10 as the base and the calculation result of the quadratic polynomial as the exponent; and using the result of the power operation as the final turbidity prediction value.

[0097] It should be noted that, based on the preceding description, the final determined model band structure is B3 / B2+B4. Three bands, B2, B3, and B4, are extracted from the remote sensing data, and the independent variable X is calculated according to this relationship. The independent variable X = reflectance of band B3 ÷ reflectance of band B2 + reflectance of band B4. Here, the quadratic polynomial refers to ax... 2 The mathematical expression for +bx+c is given, where a, b, and c are coefficients determined through fitting, corresponding to the quadratic term coefficient, the linear term coefficient, and the constant term, respectively. Further, the turbidity value T can be obtained. Assuming a = 8.9342, b = -12.2559, and c = 4.3776, then T = .

[0098] S105. Predict the real-time turbidity of the target sea area based on the real-time remote sensing data of the target sea area and the target turbidity prediction model.

[0099] Specifically, predicting the real-time turbidity of the target sea area based on real-time remote sensing data of the target sea area and the target turbidity prediction model includes:

[0100] (1) Select a sub-region in the target sea area whose aerosol optical thickness distribution is consistent with the aerosol distribution characteristics during the sampling period when the turbidity value to be fitted is obtained.

[0101] It should be noted that aerosol optical thickness is a key parameter characterizing the light attenuation capability of aerosol particles, and its spatial distribution and temporal variation are one of the main sources of interference affecting the stability of remote sensing reflectance signals in nearshore waters. Since the target turbidity prediction model constructed in this application is trained based on sample data collected under specific aerosol conditions, significant differences between the aerosol environment during application and the modeling process will introduce uncontrollable systematic errors, leading to a decrease in prediction accuracy. Therefore, to ensure the reliability of the model application, it is necessary to select regions with similar aerosol conditions for prediction.

[0102] Specifically, aerosol optical thickness data for the target sea area on the modeling sampling date and the real-time prediction date can be obtained from satellite products. Sub-regions can be delineated by calculating the similarity of spatial distribution patterns (such as correlation coefficients) or setting the fluctuation range of AOD values.

[0103] (2) Extract the real-time remote sensing data of the sub-region, input the real-time remote sensing data into the target turbidity prediction model, and obtain the real-time turbidity prediction result of the sub-region.

[0104] It should be noted that after determining the sub-region with consistent aerosol conditions, the band reflectance data corresponding to each pixel in the sub-region and required by the target turbidity prediction model are extracted from the real-time atmospheric corrected remote sensing data. Then, the data is input into the target turbidity prediction model to obtain the real-time turbidity prediction result of the sub-region.

[0105] It should also be noted that the method further includes applying the target turbidity prediction model to remote sensing data from other satellites, specifically including:

[0106] (1) Acquire the sensor characteristics of the target satellite and the atmospheric-corrected remote sensing data of the target satellite, wherein the sensor characteristics include at least the spectral response function.

[0107] It should be noted that, to broaden the applicability and promotion value of the target turbidity prediction model constructed in this application, a cross-sensor application process was also designed. Here, the target satellite refers to a satellite platform planned for turbidity monitoring with similar visible to near-infrared wavelengths, such as Landsat OLI and Sentinel-2 MSI. Sensor characteristics are a set of parameters describing the photoelectric performance of a sensor, where the spectral response function is the sensor's relative response efficiency to different wavelengths of light within each nominal band. Due to differences in the spectral response functions of different sensors, even when observing the same ground object, the reflectance values ​​obtained within the same band range will exhibit systematic deviations. Therefore, obtaining the spectral response function of the target satellite is a prerequisite for model transfer applications. Simultaneously, the atmospherically corrected remote sensing data from the target satellite must be at the same level as the modeling data (e.g., L2A level) to ensure data comparability.

[0108] (2) Based on the difference between the spectral response function of the target satellite and the spectral response function of the HY-1C / D CZI sensor, the function coefficients of the target turbidity prediction model are corrected.

[0109] Due to differences in spectral response functions, directly applying a target turbidity prediction model trained on HY-1C / D CZI data to target satellite data will result in inaccurate predictions due to systematic spectral biases. Therefore, it is necessary to correct the model's function coefficients. Specific correction methods can include: band simulation based on the spectral response function, using a hyperspectral ground cover spectral library or synchronously measured hyperspectral data, combined with the spectral response functions of HY-1C / D CZI and the target satellite, to simulate and calculate the reflectance values ​​of both in the same band, establishing a conversion relationship between them; or a refitting strategy, collecting a new set of measured turbidity data synchronously with the target satellite's transit in the target sea area, using this dataset and the target satellite's L2A data, and refitting only the coefficients within the fixed framework of the band combinations and function forms determined by the HY-1C / D CZI model, thus obtaining corrected model coefficients suitable for the target satellite. This step, through adaptive adjustment of the coefficient domain, effectively compensates for the spectral mismatch caused by differences in sensor hardware.

[0110] (3) Apply the corrected target turbidity prediction model to the atmospheric corrected remote sensing data of the target satellite to calculate the real-time turbidity of the target sea area.

[0111] It should be noted that after completing the sensor-specific calibration of the model coefficients, the calibrated target turbidity prediction model can be formally applied to the atmospheric-corrected remote sensing data acquired by the target satellite. The application process is completely consistent with that described in S105. This design expands the practicality of the method in this application, enabling continuous and coordinated monitoring of turbidity in the target sea area using data from other satellites when HY-1C / D CZI data coverage is insufficient, thereby improving the temporal resolution of monitoring and data availability.

[0112] Optionally, applying the target turbidity prediction model to remote sensing data from other satellites further includes: acquiring atmospherically corrected remote sensing data of the target satellite and the turbidity value to be fitted for the target sea area; selecting target remote sensing bands based on the fitting results of the single-band remote sensing reflectance data of the target satellite and the turbidity value to be fitted; determining a turbidity prediction model example based on the fitting results of any combination of two target remote sensing bands and the turbidity value to be fitted; modifying the band composition in the turbidity prediction model example, and determining a target turbidity prediction model suitable for the target satellite based on the fitting results of the modified instantiated turbidity prediction model and the turbidity value to be fitted. It should be noted that this optional implementation allows the turbidity extraction method described above to be applied to remote sensing data from other satellites, that is, the systematic modeling methodology provided in this application is directly applied to the remote sensing data of the target satellite and the synchronously measured dataset. Specifically, when facing different satellite sensors (such as Landsat OLI, Sentinel-2 MSI), it does not rely on adjusting the coefficients of the original HY-1C / D CZI model, but rather re-initializes the entire modeling process. Starting with the selection of sensitive bands, the process involves optimizing band combinations, modifying the model structure, and determining the function form, ultimately resulting in a target turbidity prediction model that is completely retrained based on the spectral characteristics and data features of the target sensor.

[0113] This approach, together with the previously mentioned coefficient correction method, constitutes a solution to the problem of limited model adaptability. The coefficient correction method focuses on efficient transfer, rapidly adapting the model through mathematical correction given known differences in sensor spectral response functions. This is suitable for scenarios with high efficiency requirements and sensor calibration knowledge. The current remodeling method, however, focuses on precise adaptation. It does not pre-determine correction relationships but allows the model to relearn from the data. Theoretically, this can more thoroughly absorb the combined influence of the target sensor's spectral response, radiometric characteristics, etc., and can pursue optimal inversion accuracy when supported by sufficient synchronous measured data. These two methods break through the dependence on a single data source and can be flexibly and effectively deployed on diverse satellite remote sensing platforms.

[0114] The method provided in this embodiment firstly identifies the target sea area as a complex nearshore optical region, emphasizing that sampling points must cover different turbidity gradients and avoid severe weather processes. Simultaneously, L2A-level atmospheric-corrected remote sensing data is used to ensure the representativeness and reliability of the modeling data, providing high-quality input for subsequent model construction and reducing the impact of data errors on the inversion results from the source. Secondly, in the band selection stage, the correlation coefficient between a single band and turbidity is calculated, and a threshold is set to select target bands. This precisely focuses on spectral regions sensitive to turbidity, eliminating interference from low-correlation bands while retaining the necessary spectral information for combined modeling, thus solving the problems of blind and redundant band selection in existing technologies. First, through various mathematical combinations of pairwise target bands and turbidity fitting, the optimal combination form is determined as a model example. Then, by modifying the band composition, multiple sets of instantiated models are generated, and the band combinations are selected and optimized. Finally, the target model is determined by combining various functional forms, fully exploring the potential of multi-band synergistic turbidity representation. Compared with single-band models, this significantly improves the goodness of fit and effectively addresses the shortcomings of insufficient single-band inversion accuracy. Furthermore, in the turbidity prediction stage, predictions were made for sub-regions whose aerosol distribution characteristics were consistent with those during the sampling period. This targeted reduction of the impact of aerosols, a key interfering factor, avoided prediction errors caused by differences in the atmospheric environment, and improved the stability of the model application. It was also demonstrated that the model can achieve large-scale turbidity prediction based on real-time remote sensing data, significantly improving monitoring efficiency and spatial coverage compared to traditional field measurements. Finally, the model's adaptability was further expanded. By correcting the model coefficients for differences in the spectral response functions of different satellite sensors, the target turbidity prediction model can be transferred to other satellite data, overcoming the limitations of existing models that rely on specific sensors and enhancing the method's universality.

[0115] Example 2

[0116] Corresponding to the aforementioned embodiment of a turbidity extraction method based on remote sensing data, this application also provides an embodiment of a turbidity extraction device based on remote sensing data.

[0117] Figure 2 This is a schematic diagram of the structure of Embodiment 2 of the turbidity extraction device based on remote sensing data provided in this application. Please refer to... Figure 2 The apparatus provided in this embodiment includes an acquisition module 210, a filtering module 220, a determination module 230, and a prediction module 240;

[0118] The acquisition module 210 is used to acquire atmospheric-corrected remote sensing data of the target sea area and turbidity values ​​to be fitted from multiple sampling points;

[0119] The filtering module 220 is used to filter multiple target remote sensing bands based on the first fitting result of the single-band remote sensing reflectance data in the atmospheric corrected remote sensing data and the turbidity value to be fitted.

[0120] The determining module 230 is used to determine a turbidity prediction model example based on the fitting result of the combination of any two target remote sensing bands and the second turbidity value to be fitted.

[0121] The determining module 230 is also used to modify the band composition in the turbidity prediction model example, and determine the target turbidity prediction model based on the modified instantiated turbidity prediction model and the second fitting result of the turbidity value to be fitted.

[0122] The prediction module 240 is used to predict the real-time turbidity of the target sea area based on the real-time remote sensing data of the target sea area and the target turbidity prediction model.

[0123] The apparatus of this embodiment can be used to perform... Figure 1 The steps of the method embodiment shown are similar in principle and process, and will not be repeated here.

[0124] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0125] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for turbidity extraction based on remote sensing data, characterized in that, The method includes: Acquire atmospheric-corrected remote sensing data of the target sea area and turbidity values ​​to be fitted from multiple sampling points; Multiple target remote sensing bands are selected based on the first fitting result of the single-band remote sensing reflectance data in the atmospheric corrected remote sensing data and the turbidity value to be fitted. A turbidity prediction model example is determined based on the second fitting result of any two target remote sensing bands and the turbidity value to be fitted. Modify the band composition in the turbidity prediction model example, and determine the target turbidity prediction model based on the modified instantiated turbidity prediction model and the second fitting result of the turbidity value to be fitted. The real-time turbidity of the target sea area is predicted based on the real-time remote sensing data of the target sea area and the target turbidity prediction model. The step of determining turbidity prediction model examples based on the second fitting result of the combination of any two target remote sensing bands and the turbidity value to be fitted includes: The selected target remote sensing bands are then paired up in pairs. For each pair of bands, multiple mathematical relationships are constructed to form multiple candidate band combinations. Each mathematical form is used to perform numerical calculations on the two target remote sensing bands in each pair of bands. Each candidate band combination is model-fitted with the turbidity value to be fitted. Compare the goodness of fit and error index of all candidate band combinations, and select the mathematical relationship corresponding to the candidate band combination with the highest goodness of fit and the smallest error index as the example of the turbidity prediction model. The turbidity prediction model example is a ratio. Modifying the band composition in the turbidity prediction model example, and determining the target turbidity prediction model based on the modified instantiated turbidity prediction model and the second fitting result of the turbidity value to be fitted, includes: Candidate bands are determined based on the first fitting result. The numerator, denominator, and additional terms of the turbidity prediction model are adjusted according to the candidate bands to construct a subset of instantiated turbidity prediction models. All models in the subset of instantiated turbidity prediction models are fitted to the turbidity value to be fitted. From all the second fitting results, the instantiated turbidity prediction model with the highest fitting goodness and the smallest error is selected, and the corresponding remote sensing band combination is determined as the optimized band combination. The numerical values ​​of the optimized band combination are used as independent variables, and function fitting is performed with various preset function forms respectively. From the fitting results of all function forms, the function form with the highest goodness of fit and the smallest error and its corresponding coefficients are selected to form the target turbidity prediction model.

2. The method according to claim 1, characterized in that, The step of predicting the real-time turbidity of the target sea area based on real-time remote sensing data of the target sea area and the target turbidity prediction model includes: Select a sub-region in the target sea area whose aerosol optical thickness distribution is consistent with the aerosol distribution characteristics during the sampling period when the turbidity value to be fitted is obtained; Real-time remote sensing data of the sub-region is extracted and input into the target turbidity prediction model to obtain the real-time turbidity prediction result of the sub-region.

3. The method according to claim 1, characterized in that, The step of selecting multiple target remote sensing bands based on the first fitting result of the single-band remote sensing reflectance data in the atmospherically corrected remote sensing data and the turbidity value to be fitted includes: Calculate the correlation coefficient between each single-band remote sensing reflectance data and the turbidity value to be fitted; For all sampling points, calculate the average correlation coefficient between each single band and the turbidity value to be fitted, and sort the bands from high to low according to the average correlation coefficient; The single bands that are ranked high and whose average correlation coefficient is greater than a preset threshold are selected as the target remote sensing bands.

4. The method according to claim 1, characterized in that, Based on the band composition of the target turbidity prediction model, the corresponding band reflectance is extracted from the atmospherically corrected remote sensing data and the combined value is calculated. Calculate the quadratic polynomial of the combined numerical values, wherein the quadratic polynomial comprises a quadratic term, a linear term, and a constant term; Using 10 as the base and the result of the quadratic polynomial as the exponent, perform exponentiation; The result of the exponentiation is used as the final turbidity prediction value.

5. The method according to claim 1, characterized in that, The method also includes applying the target turbidity prediction model to remote sensing data from other satellites, specifically including: Acquire the sensor characteristics of the target satellite and the atmospheric-corrected remote sensing data of the target satellite, wherein the sensor characteristics include at least the spectral response function; Based on the difference between the spectral response function of the target satellite and the spectral response function of the HY-1C / D CZI sensor, the function coefficients of the target turbidity prediction model are corrected. The corrected target turbidity prediction model is applied to the atmospheric-corrected remote sensing data of the target satellite to calculate the real-time turbidity of the target sea area.

6. The method according to claim 1, characterized in that, The single-band remote sensing reflectance data varies with aerosol concentration. Atmospheric-corrected remote sensing data with a specified weather type and aerosol optical thickness below a preset threshold are selected as the real-time remote sensing data.

7. A turbidity extraction device based on remote sensing data, characterized in that, The device includes an acquisition module, a filtering module, a determination module, and a prediction module; The acquisition module is used to acquire atmospheric-corrected remote sensing data of the target sea area and turbidity values ​​to be fitted from multiple sampling points; The filtering module is used to filter multiple target remote sensing bands based on the first fitting result of the single-band remote sensing reflectance data in the atmospherically corrected remote sensing data and the turbidity value to be fitted. The determining module is used to determine a turbidity prediction model example based on the combination of any two target remote sensing bands and the second fitting result of the turbidity value to be fitted. The determining module is also used to modify the band composition in the turbidity prediction model example, and determine the target turbidity prediction model based on the modified instantiated turbidity prediction model and the second fitting result of the turbidity value to be fitted. The prediction module is used to predict the real-time turbidity of the target sea area based on the real-time remote sensing data of the target sea area and the target turbidity prediction model. The step of determining turbidity prediction model examples based on the second fitting result of the combination of any two target remote sensing bands and the turbidity value to be fitted includes: The selected target remote sensing bands are then paired up in pairs. For each pair of bands, multiple mathematical relationships are constructed to form multiple candidate band combinations. Each mathematical form is used to perform numerical calculations on the two target remote sensing bands in each pair of bands. Each candidate band combination is model-fitted with the turbidity value to be fitted. Compare the goodness of fit and error index of all candidate band combinations, and select the mathematical relationship corresponding to the candidate band combination with the highest goodness of fit and the smallest error index as the example of the turbidity prediction model. The turbidity prediction model example is a ratio. Modifying the band composition in the turbidity prediction model example, and determining the target turbidity prediction model based on the modified instantiated turbidity prediction model and the second fitting result of the turbidity value to be fitted, includes: Candidate bands are determined based on the first fitting result. The numerator, denominator, and additional terms of the turbidity prediction model are adjusted according to the candidate bands to construct a subset of instantiated turbidity prediction models. All models in the subset of instantiated turbidity prediction models are fitted to the turbidity value to be fitted. From all the second fitting results, the instantiated turbidity prediction model with the highest fitting goodness and the smallest error is selected, and the corresponding remote sensing band combination is determined as the optimized band combination. The numerical values ​​of the optimized band combination are used as independent variables, and function fitting is performed with various preset function forms respectively. From the fitting results of all function forms, the function form with the highest goodness of fit and the smallest error and its corresponding coefficients are selected to form the target turbidity prediction model.

Images