Suspended sediment concentration detection method and device based on hyperspectral remote sensing

Abstract

The invention provides a suspended sediment concentration detection method and device based on hyperspectral remote sensing, and relates to the technical field of water body monitoring. According to the method, a water body type classification mechanism is constructed, and the target water area is divided into different water body types in combination with optical differences presented by water bodies with different components, so that a proper inversion model is called for water body detection; the problem that a single model is poor in applicability in a complex water body environment is avoided, independent detection and evaluation of different water body types are achieved, when the detection effect of a certain water body type is poor, an updating mechanism is triggered, the optimal inversion model of the water body type is retrained by using updated data, and the accuracy of the water body type is improved. While the algorithm precision is improved, interference on inversion detection of other water body types is effectively avoided, so that the system cannot be paralyzed due to failure of an individual inversion model, and the detection capability of the whole system adapting to long-term change of the water body is improved.

Description

Technical Field

[0001] This invention relates to the field of water monitoring technology, specifically to a method and device for detecting suspended sediment concentration based on hyperspectral remote sensing. Background Technology

[0002] Hyperspectral remote sensing technology is an important technique for detecting suspended sediment concentrations in water bodies over large areas, and it is currently a crucial tool for environmental monitoring and hydrodynamic research. However, it faces several limitations in terms of inversion accuracy. For example, water body types vary significantly across different regions, making inversion models incompatible. Furthermore, the distribution of suspended sediment within the same water body is uneven, and other substances in the water can interfere with remote sensing images, affecting feature extraction. Therefore, employing a multi-model fusion strategy is a mainstream development trend in this field.

[0003] In existing technologies, CN120741397A discloses a method for monitoring suspended sediment concentration in water bodies based on FUI water body classification. Its core idea is to use stable indicators such as the Forel-Ule index to classify water bodies into categories like "clear water" and "turbid water" based on a single threshold, and to construct a single empirical or semi-empirical inversion model for each category. However, this method still has certain limitations. For example, classifying water bodies into clear and turbid water solely based on turbidity ignores the complexity of real-world water bodies. Some high-turbidity water bodies, such as the Yellow River, would be classified as entirely turbid under this method, leading to the use of a single model for inversion. Furthermore, this method lacks autonomous evaluation and dynamic optimization design, making it difficult to adapt to long-term changes in the water environment.

[0004] Therefore, existing technologies lack a method for detecting suspended sediment concentration that can achieve detailed identification of water body types, intelligent matching and inversion models, and continuous learning capabilities, in order to solve the problems of insufficient inversion accuracy and weak generalization ability of existing technologies. Summary of the Invention

[0005] The purpose of this invention is to provide a method and apparatus for detecting suspended sediment concentration based on hyperspectral remote sensing, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for detecting suspended sediment concentration based on hyperspectral remote sensing, comprising the following steps: Step 1: Set up measurement points in the target water area to be tested, acquire the hyperspectral reflectance image of the target water area at the set time, and select the pixels that correspond to the measurement points in time and space as sampling pixels. Simultaneously acquire the water body measurement data of each measurement point and associate it with the corresponding sampling pixels. Step 2: Set allocation rules based on the measured water body data, assign a water body type label to each sampled pixel, extract type features from the sampled pixels and use them as input, and use the water body type label of the sampled pixels as output to train the classifier; Step 3: Select the optimal inversion model for each water body type through cross-validation. Extract the inversion features required by the optimal inversion model from the sampled pixels according to the water body type label and use them as input. Use the measured value of the suspended sediment concentration in the water body measurement data as the output to train the optimal inversion model for each water body type. Step 4: Real-time acquisition of the hyperspectral reflectance of the target water area to be detected and label it as the first image. Select the pixels that correspond to the spatiotemporal of the measured points as the detection pixels, and simultaneously acquire the latest measured value of suspended sediment concentration. Identify the water body type of the detection pixels through the classifier, call the corresponding optimal inversion model, and output the predicted value of suspended sediment concentration of the detection pixels. Step 5: For the detected pixels of different water body types, calculate the average error between the predicted value and the latest measured value of suspended sediment concentration. If the error reaches the set threshold, retrain the optimal inversion model. Step 6: Extract the type features of all pixels in the first image and input them into the classifier. Based on the water body type identification results, call the corresponding optimal inversion model and output the predicted values ​​of suspended sediment concentration for all pixels. Integrate the output results to obtain the suspended sediment concentration distribution map of the target water area.

[0007] Furthermore, when setting the measurement points, the target water area to be tested is divided into several sub-regions of equal area, and the center of each sub-region is taken as the measurement point; Multiple time intervals are set, and the time for acquiring the hyperspectral reflectance image is consistent with the water sampling time at each actual measurement point. The measured data of the water body includes the measured values ​​of chlorophyll concentration and suspended sediment concentration.

[0008] Furthermore, based on the measured values ​​of suspended sediment concentration in the sampled pixels, a first classification threshold and a second classification threshold are set, wherein the first classification threshold is greater than the second classification threshold, and a third classification threshold is set based on the chlorophyll concentration values ​​in the sampled pixels. For each sampled pixel, if its measured suspended sediment concentration is greater than the first classification threshold, it is labeled as high turbidity saturated type; if its measured suspended sediment concentration is less than the second classification threshold, it is labeled as clear type. For sampled pixels whose measured suspended sediment concentration is less than or equal to the first classification threshold and greater than or equal to the second classification threshold, if its chlorophyll concentration is greater than the third classification threshold, it is labeled as algae-sediment mixture type; otherwise, it is labeled as sediment-dominant type.

[0009] Furthermore, the type feature is a multidimensional array containing the red band reflectance, near-infrared band reflectance, red edge band ratio, and fluorescence height peak of the pixel.

[0010] Furthermore, multiple candidate models are set up. For each type of water body, the pixels belonging to that type of water body are randomly and evenly divided into multiple groups. In each round, one group is selected as the validation set, and the remaining groups are combined into the training set. A candidate model is called, and the inversion features required by the model are extracted from the pixels in the training set and used as input. The corresponding measured value of suspended sediment concentration is used as the output to train the candidate model. Subsequently, the same inversion features are extracted from the validation set pixels and input into the candidate model. The predicted values ​​of suspended sediment concentration of the validation set pixels are output. The difference between the predicted value and the measured value of each validation set pixel is squared and then summed. The sum is divided by the number of validation set pixels and then the root mean square error is obtained to obtain the root mean square error of the validation set. This is used as the effectiveness index of this round of validation. Repeat the above steps, selecting a new set of pixels as the validation set in each round, and combining the remaining sets into the training set. Train and validate the same candidate model, so that the training effect of the candidate model is validated once for each set of pixels. After the loop is completed, the sum of the effect indices of each round is divided by the number of rounds to obtain the mean root mean square error of the current candidate model. The mean root mean square error of other candidate models for the same water body type is calculated using the same method. The model with the smallest value is selected as the optimal inversion model for this water body type.

[0011] Furthermore, the candidate models include Bayesian linear regression, support vector regression, random forest, and gradient boosting tree models. For each candidate model, the input form of the inversion features is defined according to its algorithm principle.

[0012] Furthermore, the inversion features are constructed by calculating and combining the reflectance of pixels in different spectral bands; For the Bayesian linear regression model, the ratio of blue-green band reflectance is used as the inversion feature that needs to be input. For the support vector regression model, a two-dimensional array is constructed, which contains the ratio of red-edge bands and the rate of change of chlorophyll fluorescence peak bands. This two-dimensional array is used as the inversion feature that the support vector regression model needs to input. For the random forest model, a three-dimensional array is constructed, which contains the red band reflectance, the near-infrared band reflectance, and the red edge band ratio. This three-dimensional array is used as the inversion feature that the random forest model needs to input. For the gradient boosting tree model, a multidimensional array is constructed, which includes reflectance of each band, ratio of red-edge band, yellow absorption index, fluorescence height peak, rate of change of chlorophyll fluorescence peak band, and spectral curvature index. This multidimensional array is used as the inversion feature that the gradient boosting tree model needs to input.

[0013] Furthermore, for each detected pixel selected from the first image, its type features are first extracted and input into the trained classifier. Based on its water body type identification result, the corresponding optimal inversion model is called. Then, the inversion features required by the model are extracted and input to obtain the predicted value of suspended sediment concentration for the detected pixel. For each type of water body detected pixel, the sum of the absolute values ​​of the differences between its predicted value of suspended sediment concentration and the latest measured value is calculated and divided by the number of detected pixels of this type of water body to obtain the average error of each water body type. For water body types whose average error reaches a set threshold, the detected pixels and sampled pixels belonging to this type are merged, the corresponding inversion features are extracted as input, and the corresponding measured value of suspended sediment concentration is used as output to retrain the optimal inversion model for this type of water body.

[0014] Furthermore, the method for integrating the suspended sediment concentration distribution map is as follows: the predicted values ​​of suspended sediment concentration of all pixels in the first image are arranged into a two-dimensional matrix, corresponding one-to-one with each pixel; according to the preset concentration-color mapping relationship, the numerical matrix is ​​converted into a color matrix; and based on the color matrix, the suspended sediment concentration distribution map of the target water area is generated.

[0015] The present invention also provides a suspended sediment concentration detection device based on hyperspectral remote sensing, wherein the suspended sediment concentration detection device based on hyperspectral remote sensing is used to perform the above-described suspended sediment concentration detection method based on hyperspectral remote sensing, comprising: Data sampling module: used to set measurement points in the target water area to be detected, acquire hyperspectral reflectance images of the target water area at a set time, and select pixels that correspond to the measurement points in time and space as sampling pixels. Simultaneously acquire water body measurement data of each measurement point and associate them with the corresponding sampling pixels. Classifier training module: It is used to set allocation rules based on the measured water body data, assign a water body type label to each sampled pixel, extract type features from the sampled pixels and use them as input, and use the water body type label of the sampled pixels as output to train a classifier. Inversion Model Training Module: This module is used to select the optimal inversion model for each water body type through cross-validation. It extracts the inversion features required by the optimal inversion model from the sampled pixels based on the water body type label and uses them as input. The output is the measured value of the suspended sediment concentration in the actual water body data. This module trains the optimal inversion model for each water body type. Concentration prediction module: It is used to acquire the hyperspectral reflectance of the target water area to be detected in real time and label it as the first image. It selects the pixels that correspond to the spatiotemporal of the measured point as the detection pixels and simultaneously acquires the latest measured value of suspended sediment concentration. It identifies the water body type of each detection pixel through the classifier, calls the corresponding optimal inversion model, and outputs the predicted value of suspended sediment concentration of each detection pixel. Evaluation and Retraining Module: This module is used to calculate the average error between the predicted value and the latest measured value of suspended sediment concentration for detection pixels of different water body types. If the error reaches a set threshold, the optimal inversion model is retrained. Distribution map integration module: used to extract the type features of all pixels in the first image and input them into the classifier. Based on the water body type identification results, it calls the corresponding inversion model, outputs the predicted values ​​of suspended sediment concentration for all pixels, integrates the output results, and obtains the suspended sediment concentration distribution map of the target water area.

[0016] Compared with the prior art, the beneficial effects of the present invention are: This invention constructs a water body type classification mechanism. By combining the optical differences exhibited by water bodies with different components, the target water area is divided into different water body types. This allows for the use of appropriate inversion models for water body detection, avoiding the problem of poor applicability of a single model in complex water environments.

[0017] Through the core architecture of "classification first, then adaptive inversion", this invention first identifies the water body type for each pixel, and then selects the optimal inversion model for each water body type through multiple rounds of differential verification. This effectively suppresses the measurement error caused by algorithm accuracy issues between different types of water bodies, thereby achieving high-precision inversion detection in complex mixed water areas.

[0018] This invention enables independent detection and evaluation of different water body types. When the detection effect of a certain water body type is poor, an update mechanism is triggered. The updated data is used to retrain the optimal inversion model for this water body type. While improving the accuracy of the algorithm, it effectively avoids interference with the inversion detection of other water body types, so that the system will not be paralyzed due to the failure of individual inversion models, and improves the detection capability of the entire system to adapt to long-term changes in water bodies. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall method flow of the present invention; Figure 2 A schematic diagram of water body type classification provided for embodiments of the present invention; Figure 3 This is a spatial distribution map of suspended sediment concentration provided in an embodiment of the present invention; Figure 4 This is a structural block diagram of the overall device of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0021] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0022] Example: Please see Figures 1 to 3 The present invention provides a technical solution: A method for detecting suspended sediment concentration based on hyperspectral remote sensing, comprising the following steps: Step 1: Set up measurement points in the target water area to be tested, acquire hyperspectral reflectance images of the target water area at the set time, and select pixels that correspond to the measurement points in time and space as sampling pixels. Simultaneously acquire water body measurement data of each measurement point and associate them with the corresponding sampling pixels.

[0023] Furthermore, when setting the measurement points, the target water area to be tested is divided into several sub-regions of equal area, and the center of each sub-region is taken as the measurement point; Multiple time intervals are set, and the time for acquiring the hyperspectral reflectance image is consistent with the water sampling time at each actual measurement point. The measured data of the water body includes the measured values ​​of chlorophyll concentration and suspended sediment concentration.

[0024] In the field of water monitoring technology, in order to establish and verify the quantitative relationship between remote sensing reflectance and water parameters, in addition to large-scale calculations using hyperspectral imagery, a crucial and necessary task is to conduct simultaneous sampling at field measurement points within the target water area. Through laboratory analysis, accurate water data from these measurement points can be obtained, thereby verifying the inversion effect of remote sensing technology. Therefore, for each hyperspectral image, there exists a certain number of pixels with true ground values.

[0025] The selection of measurement points needs to consider the establishment and verification of the "spectral-parameter" data relationship in remote sensing inversion, achieving a balance between spatial coverage, temporal synchronization, and data correlation. Based on this, this embodiment divides the target water area into sub-regions of equal area, using the center of each sub-region as the measurement point—a systematic spatial sampling design. This ensures that the measurement points are approximately uniformly distributed within the target water area, avoiding spatial bias caused by accidental clustering of sampling points in specific regions. The training data obtained in this way can cover as broadly as possible the different optical water body types that may exist within the water area, enabling the subsequently trained classifier and inversion model to learn more comprehensive and representative spectral features, thereby improving the model's generalization ability across the entire water area.

[0026] In this embodiment, the core strategy adopted to construct a highly reliable training dataset is to select all pixels with ground truth values ​​from all historical hyperspectral water reflectance images of the target water area, that is, pixels that are consistent with the time of historical measured points and correspond to spatial coordinates. These pixels are defined as "sampled pixels", and the sampled pixels and their measured water data are used as the basis for subsequent model training and validation.

[0027] Step 2: Set allocation rules based on the measured water body data, assign a water body type label to each sampled pixel, extract type features from the sampled pixels and use them as input, and use the water body type label of the sampled pixels as output to train the classifier.

[0028] In actual aquatic environments, the distribution of suspended sediment is not uniform due to various factors such as hydrodynamics and material sources. Even within the same area of ​​water covered by a single imaging, its optical properties may vary significantly. If a single, global inversion model is used for the entire water area, the model will fit all different types of optical signals, resulting in the final detection results having high reliability only in some areas, while producing systematic biases in other areas.

[0029] To address these issues, this embodiment incorporates expertise in water optics. Based on the main optically active components of water and their dominant influence mechanisms on the spectrum, water optics types are categorized into four major categories with clear physical significance: High turbidity saturation: The concentration of suspended sediment is extremely high, the turbidity of the water is extremely high, and the light transmittance of the water is extremely poor. At this time, the absorption and scattering of light by suspended matter in the water is extremely strong, which causes the spectral reflectance of the water to often show a plateau characteristic in the visible to near-infrared region. That is, the spectral curve no longer increases linearly with the increase of sediment concentration, but approaches a saturation state, and cannot truly reflect the sediment concentration information; Clear type: The concentration of suspended sediment is extremely low, the water is clear, and its optical properties are mainly determined by the properties of the water molecules themselves and the dissolved substances they contain. In the process of hyperspectral imaging, clear water has extremely strong absorption in the near-infrared and short-wave infrared bands, and high reflectivity in the visible blue-green band. Its spectral curve is relatively smooth overall, lacking characteristic absorption peaks and reflection peaks caused by suspended particles or chlorophyll. Algal-sediment mixture type: The concentration of suspended sediment is at a moderate level, while the chlorophyll concentration is high. The water body contains significant concentrations of both suspended sediment and phytoplankton, two optically active substances, and the spectral signal is a superposition of these two. Its spectral reflectance exhibits characteristics of both algae, such as a chlorophyll absorption trough near 675 nm and a chlorophyll fluorescence peak near 700 nm, and characteristics of sediment, such as an overall increase in reflectance, especially in the yellow and red bands.

[0030] Sediment-dominated type: The concentration of suspended sediment is at a moderate level, but its chlorophyll concentration is low. The optical properties of the water body are mainly dominated by inorganic suspended sediment particles, with minimal influence from the absorption of pigments such as chlorophyll. In hyperspectral images, as the sediment concentration increases, the spectral reflectance rises across the entire spectrum from blue to red, and the reflectance peak shifts from the blue-green band to the yellow-red band. The spectral curve typically shows a monotonically increasing trend up to the near-infrared band, lacking obvious algal absorption valleys and chlorophyll fluorescence peaks. The overall spectral morphology usually shows a gradual decrease in reflectance with increasing wavelength.

[0031] Based on the above theoretical foundation, the core idea of ​​this embodiment is to first classify each pixel according to its water body type, and then assign it a suitable inversion model. To this end, it is first necessary to establish the basis and standards for classification and formulate classification rules based on statistical analysis of historical measured data and professional knowledge in the field of water optics. The optical signals of natural water bodies are jointly affected by suspended sediment and phytoplankton. The distribution of suspended sediment is represented by its concentration value, while the distribution of phytoplankton is usually indicated by chlorophyll a concentration. The strong absorption of chlorophyll a in the red light band and its fluorescence emission characteristics in near-red light significantly interfere with the suspended sediment inversion model based on the backscattering principle, and is one of the main sources of error in complex water body inversion.

[0032] Therefore, in this embodiment, chlorophyll a concentration and suspended sediment concentration are used as the core parameters for water body type identification. In the data preparation stage, the measured water body data to be obtained should include at least the chlorophyll a concentration value (Chl) and the suspended sediment concentration value (SSC).

[0033] Furthermore, based on the measured values ​​of suspended sediment concentration in the sampled pixels, a first classification threshold and a second classification threshold are set, wherein the first classification threshold is greater than the second classification threshold, and a third classification threshold is set based on the chlorophyll concentration values ​​in the sampled pixels. For each sampled pixel, if its measured suspended sediment concentration is greater than the first classification threshold, it is labeled as high turbidity saturated type; if its measured suspended sediment concentration is less than the second classification threshold, it is labeled as clear type. For sampled pixels whose measured suspended sediment concentration is less than or equal to the first classification threshold and greater than or equal to the second classification threshold, if its chlorophyll concentration is greater than the third classification threshold, it is labeled as algae-sediment mixture type; otherwise, it is labeled as sediment-dominant type.

[0034] The expression for the allocation rule is: ; in This represents the water body type label for the i-th sampled pixel. This represents the SSC value of the i-th sampled pixel. T1 represents the Chl value of the i-th sampled pixel, T2 is the first classification threshold, T2 is the second classification threshold, and T3 is the third classification threshold; A represents the water body type label "high turbidity saturated type", B represents the water body type label "clear type", C represents the water body type label "algae and sediment mixture type", and D represents the water body type label "sediment-dominant type".

[0035] Table 1 shows some of the SSC-Chl data from this embodiment. Based on complete measured water body data, the SSC values ​​of all sampled pixels were first sorted into an ascending order, and the 95th percentile value was taken as the first classification threshold. Then, the 10th percentile value of the SSC sequence is taken as the second classification threshold, i.e. Next, the Chl values ​​of the sampled pixels are sorted into an ascending sequence, and the 75th percentile value of the Chl sequence is taken as the third classification threshold. .

[0036] The classification threshold set in this embodiment is a refined design based on both the physical significance and statistical robustness of water body optical classification. Its core purpose is to automatically and objectively identify water body groups with significantly different optical characteristics through statistical methods, laying a scientific foundation for subsequent classification and inversion.

[0037] First, the 95th percentile value was chosen as the first classification threshold to accurately capture the "saturation critical point" in the optical effects of water. From an optical principle perspective, when the concentration of suspended sediment is extremely high, the backscattering signal of the water body approaches saturation, and the spectral reflectance no longer increases linearly with increasing concentration. The logic for setting the threshold at the 95th percentile of the data distribution (i.e., the top 5% with the highest concentration) is that these pixels with extremely high concentrations are most likely to exhibit nonlinear saturation optical characteristics. Classifying these pixels separately as "high turbidity saturation type" allows for the matching of specialized nonlinear inversion models, thereby avoiding interference from their abnormal optical characteristics with mainstream linear or quasi-linear models applicable to medium and low concentration water bodies. It also avoids the overall decrease in accuracy caused by forcibly fitting a single model to all concentration ranges.

[0038] Secondly, the 10th percentile was chosen as the second classification threshold to effectively distinguish extremely low-turbidity water bodies, i.e., "clear" water bodies, whose optical characteristics are dominated by water molecules and their dissolved substances. When the concentration of suspended sediment is extremely low, its contribution to the spectral signal becomes weak, and the optical characteristics of the water body are closer to "pure water." Setting the threshold at the 10th percentile (i.e., the lowest 10% of the concentration) aims to identify the "cleanest" pixels in this spectral feature range and model this type of water body separately, because its spectral signal is simple and less affected by other components. This threshold ensures that "clear" water bodies have a sufficiently low and concentrated concentration range, allowing their optical characteristics to be clearly distinguished from other types from the classifier's perspective.

[0039] After screening using the first and second thresholds, it was ensured that the screening of "algae-sediment mixed type" and "sediment-dominant type" water bodies was based on data with significant but non-abnormal sediment signals, making the subsequent classification steps more targeted and practical.

[0040] Choosing the 75th percentile as the third classification threshold is a statistical design based on the optical properties of water bodies and the distribution of measured data. This aims to accurately capture the critical point of chlorophyll interference with the spectrum, ensuring the applicability of the inversion model. The logic is to prioritize the selection of the top 25% of pixels with the highest chlorophyll concentration, representing "algae-silt mixed" water bodies. These pixels have sufficiently high chlorophyll concentrations, making algal pigment absorption and fluorescence emission signals dominant or at least significantly contributing to the spectrum, thus truly constituting a "mixed" signal optically.

[0041] Table 1

[0042] See Figure 2 As shown, based on the above classification rules, all sampled pixels are divided into four water body types: "high turbidity saturated type", "clear type", "algae and sediment mixed type" and "sediment-dominant type".

[0043] Furthermore, the type feature is a multidimensional array containing the red band reflectance, near-infrared band reflectance, red edge band ratio, and fluorescence height peak of the pixel.

[0044] The principle for determining type characteristics is based on the differential responses of water optical components to different spectral bands. The four characteristic parameters mentioned above correspond to different water optical properties, collectively forming a multidimensional classification feature space. Among them: The red band reflectance is the reflectance of the 665nm band. This parameter reflects the combined absorption effect of suspended sediment and chlorophyll. Suspended sediment scatters more strongly in this band, and the reflectance increases with the increase of sediment concentration. Chlorophyll, on the other hand, has a strong absorption valley near the 665nm band, and the reflectance decreases with the increase of chlorophyll concentration. Therefore, water bodies with high sediment concentration and high chlorophyll concentration can be roughly distinguished based on the red band reflectance.

[0045] The near-infrared reflectance is the reflectance of the 780nm band. This parameter reflects the turbidity level and backscattering intensity of the water body. In clean water, water molecules absorb very strongly in the near-infrared band, and the reflectance is close to 0. In turbid water, suspended particles scatter the main component, and the reflectance increases significantly. Therefore, clear water and turbid water can be roughly distinguished based on the near-infrared reflectance.

[0046] The red-edge band ratio is the ratio of the reflectivity of the 700nm band to that of the 670nm band, and the formula is: ; in, The ratio of the red-edge band. The reflectance is for the 700nm and 670nm bands.

[0047] This parameter reflects the chlorophyll fluorescence effect and red-edge shift. The chlorophyll fluorescence peak is located between 685nm and 710nm wavelengths, which is the red-edge region. Water bodies containing more algae will have a significantly increased reflectance near 700nm due to the chlorophyll fluorescence effect, which is called "red-edge uplift". If the red-edge band ratio is greater than 1, it indicates that there is a significant chlorophyll fluorescence signal in the water body. Therefore, the presence or absence of algae can be detected based on the red-edge band ratio, thereby distinguishing between algae-mixed water bodies and sediment-dominated water bodies.

[0048] The peak fluorescence height is calculated from the reflectance of the 665nm, 680nm, and 685nm bands. Specifically: the difference between the reflectance of the 685nm and 665nm bands is calculated and designated as the first difference; the difference between the center wavelengths of the 680nm and 665nm bands is calculated and designated as the second difference; the difference between the center wavelengths of the 685nm and 665nm bands is calculated and designated as the third difference; the product of the first and second differences is divided by the third difference, and then added to the reflectance of the 665nm band. The sum is then subtracted from the reflectance of the 680nm band to obtain the final peak fluorescence height. The calculation formula is: ; in, Indicates the peak value of fluorescence height. This indicates the reflectivity in the 665nm wavelength band. This indicates the reflectivity in the 680nm band. This indicates the reflectivity in the 685nm band.

[0049] This parameter quantifies the height of the chlorophyll fluorescence peak. If the result is positive, it indicates that there is a significant chlorophyll fluorescence signal in the water body, which is highly sensitive to algae in the water body. Therefore, the activity of phytoplankton in the water body can be judged based on the fluorescence height peak.

[0050] The multidimensional array expression composed of the above four parameters is: ; This array serves as a type feature, classifying target pixels into four defined water body types. Based on these type features and the water body type labels of the pixels, a classifier is trained. The spectral response patterns of the four water body types are shown in Table 2.

[0051] Table 2

[0052] Table 2 shows the differences in internal composition and optical properties among the four water body types. The "clear" water body exhibits low red band reflectance and near-zero near-infrared band reflectance, indicating very low levels of suspended sediment and phytoplankton, resulting in high light transmittance. The "high turbidity saturated" water body shows high reflectance and near-saturation in both red and near-infrared bands, indicating extremely high suspended sediment concentration and scattering dominating the optical signal. The "algae-sediment mixture" water body has a significantly higher red-edge band ratio than 1 and a markedly positive fluorescence value, typical of the "absorption-scattering" balance of phytoplankton pigments and the chlorophyll fluorescence effect, indicating that algae and sediment jointly influence the water body's optical properties. The "sediment-dominated" water body has moderate to high reflectance in both red and near-infrared bands, but a red-edge band ratio close to 1 and insignificant fluorescence characteristics, indicating that suspended sediment dominates and algae have a weak influence. Overall, the classification system constructed in this embodiment effectively reveals the different influence mechanisms of water components (such as algae and sediment) on optical properties through multi-band reflectance and fluorescence indicators, providing characteristic basis for water quality monitoring and type identification based on remote sensing data.

[0053] Step 3: Select the optimal inversion model for each water body type through cross-validation. Extract the inversion features required by the optimal inversion model from the sampled pixels according to the water body type label and use them as input. Use the measured value of the suspended sediment concentration in the actual water body data as the output to train the optimal inversion model for each water body type.

[0054] In this embodiment, the sampled pixels with measured values ​​are divided into four different water body types. The ultimate goal is to select an optimal inversion model for each water body type and use the dedicated model to calculate the inversion concentration of a specific type of water body in order to improve the accuracy of the inversion calculation of the entire water body. Therefore, it is necessary to cross-validate different inversion models based on the existing measured water body data, and select the optimal inversion model by comparing the results of multiple rounds of validation.

[0055] Furthermore, the logic for selecting the optimal inversion model for each water body type through cross-validation is as follows: Multiple candidate models are set up. For each type of water body, the pixels belonging to that type of water body are randomly and evenly divided into multiple groups. In each round, one group is selected as the validation set, and the remaining groups are combined into the training set. A candidate model is called, and the inversion features required by the model are extracted from the pixels in the training set and used as input. The corresponding measured value of suspended sediment concentration is used as output to train the candidate model. Subsequently, the same inversion features are extracted from the validation set pixels and input into the candidate model. The predicted values ​​of suspended sediment concentration of the validation set pixels are output. The difference between the predicted value and the measured value of each validation set pixel is squared and then summed. The sum is divided by the number of validation set pixels and then the root mean square error is obtained to obtain the root mean square error of the validation set. This is used as the effectiveness index of this round of validation. Repeat the above steps, selecting a new set of pixels as the validation set in each round, and combining the remaining sets into the training set. Train and validate the same candidate model, so that the training effect of the candidate model is validated once for each set of pixels. After the loop is completed, the sum of the effect indices of each round is divided by the number of rounds to obtain the mean root mean square error of the current candidate model. The mean root mean square error of other candidate models for the same water body type is calculated using the same method. The model with the smallest value is selected as the optimal inversion model for this water body type.

[0056] This embodiment employs the K-fold cross-validation method. A total of N sampling pixels for one type of water body are defined. These N pixels are divided into K groups. In each round, one group is selected as the validation set, and the remaining K-1 groups are merged to form the training set for that round. A candidate model is invoked, and based on its algorithmic properties, the corresponding inversion features are extracted from the pixels in the training set and input into the candidate model. The measured suspended sediment concentration corresponding to each pixel is output, completing the training of the candidate model for that round. The same inversion features are extracted from the selected validation set pixels in this round and input into the trained candidate model to obtain the predicted suspended sediment concentration for each validation set pixel. The difference between the predicted and measured suspended sediment concentration for each validation set pixel is squared, summed, divided by the number of validation set pixels, and then the square root is taken to obtain the root mean square error of the validation set. This root mean square error is used as the effectiveness index for this round of validation. The formula is: ; in, Let be the root mean square error of the k-th round of validation, in mg / L, and n be the number of pixels in the validation set. This represents the measured value of suspended sediment concentration in the j-th pixel of the k-th round of validation sets. Let be the predicted suspended sediment concentration of the j-th pixel in the k-th round of validation set.

[0057] Repeat the above steps, selecting a new set of pixels as the validation set in each round, and merging the remaining sets of pixels as the training set. Train and validate the same candidate model, so that each set of pixels validates the training effect of the candidate model once. After the above loop ends, an effect index is obtained for each round of validation. The sum of the effect indices for each round is divided by the number of rounds K to obtain the mean square root error of the current candidate model. The calculation formula is as follows: ; in, This represents the average root mean square error of candidate model m over K rounds of validation, expressed in mg / L. K represents the number of groups into which all N sampled pixels of a water body type are divided, and also represents the number of rounds of training and validation of model m. This represents the root mean square error of the k-th round of verification.

[0058] Using the same method, calculate the average root mean square error of other candidate models for the same water body type, and select the model with the smallest error as the optimal inversion model for this water body type.

[0059] Furthermore, the candidate models include Bayesian linear regression, support vector regression, random forest, and gradient boosting tree models. For each candidate model, the input form of the inversion features is defined according to its algorithm principle.

[0060] In selecting candidate models, this embodiment considers algorithm diversity, performance complementarity, and adaptability to different characteristics of hyperspectral data. Specifically: In the field of water body detection, the ratio of suspended sediment concentration in clear water to a specific spectral band (such as the blue-green band) has an approximately linear relationship within a certain range. This is based on prior knowledge from water color remote sensing. The regression coefficients of the corresponding Bayesian linear regression model have a clear physical meaning, namely, the change in suspended sediment concentration corresponding to a unit change in reflectance. Therefore, the Bayesian linear regression model can be applied to water bodies where the spectral response and concentration are approximately linearly related.

[0061] In addition, the relationship between water spectra and suspended sediment concentration may be non-linearly separable in the feature space. Support vector regression models can map low-dimensional non-linear problems to high-dimensional linear problems through kernel functions, which can adapt to complex spectral responses and are suitable for situations where there are many hyperspectral data bands but limited measured data.

[0062] The choice of the random forest model is based on the synergistic effect of multiple bands in the suspended sediment inversion design in different water body types. Each band has different importance weights, requiring automatic selection of feature importance. The random forest model can automatically identify key bands through information gain and improve the stability of inversion by combining the algorithm characteristics of multi-tree voting to reduce variance. It is not easily affected by noise signals and can therefore be applied to the general modeling of multiple types of water bodies.

[0063] Finally, hyperspectral data contains dozens or even hundreds of spectral bands and has a large number of spectral features. The relationship between suspended sediment concentration and spectral features may be highly nonlinear and non-monotonic. Therefore, this embodiment also selects the gradient boosting tree model as a high-performance backup model, which can effectively utilize dozens of spectral features without overfitting. Furthermore, it can fully mine spectral information through the input of multi-dimensional features, making it suitable for high-precision inversion calculations of complex water bodies.

[0064] The above four models constitute a model hierarchy from simple to complex. By cross-validating, the optimal algorithm is matched for each type of water body, which can achieve a balance between accuracy and efficiency.

[0065] Furthermore, the inversion features are constructed by calculating and combining the reflectance of pixels in different spectral bands; For the Bayesian linear regression model, the ratio of blue-green band reflectance is used as the inversion feature that needs to be input. For the support vector regression model, a two-dimensional array is constructed, which contains the ratio of red-edge bands and the rate of change of chlorophyll fluorescence peak bands. This two-dimensional array is used as the inversion feature that the support vector regression model needs to input. For the random forest model, a three-dimensional array is constructed, which contains the red band reflectance, the near-infrared band reflectance, and the red edge band ratio. This three-dimensional array is used as the inversion feature that the random forest model needs to input. For the gradient boosting tree model, a multidimensional array is constructed, which includes reflectance of each band, ratio of red-edge band, yellow absorption index, fluorescence height peak, rate of change of chlorophyll fluorescence peak band, and spectral curvature index. This multidimensional array is used as the inversion feature that the gradient boosting tree model needs to input.

[0066] Bayesian linear regression models are essentially parametric models, requiring simple one-dimensional features as input to ensure the stability of the inversion with limited data samples. This embodiment uses the ratio of blue-green reflectance in the spectral data as the inversion feature required for the Bayesian linear regression model. In the field of water body detection, the absorption and scattering effects of suspended sediment on the blue band (490nm band) differ from those on the green band (560nm band). As sediment concentration increases, the reflectance of the blue band typically decreases faster than that of the green band, resulting in a monotonic relationship between the ratio of blue-green reflectance and suspended sediment concentration. Therefore, using the ratio of blue-green reflectance as a one-dimensional scalar compresses the data dimension while retaining key optical information, which aligns with the modeling premise of Bayesian linear regression models. The formula for calculating the blue-green reflectance ratio is: ; in, The ratio of reflectivity in the blue-green bands. The reflectivity is in the 560nm band. The reflectance is measured in the 490nm band. This ratio is used as the inversion feature required as input to the Bayesian linear regression model.

[0067] Support Vector Regression (SVR) models can map input features to a high-dimensional feature space using kernel functions. This embodiment chooses to construct a two-dimensional array as the form of input features for the SVR model. Two-dimensional features are the minimum dimension for constructing complex decisions in the kernel space, avoiding increased computational burden and interference from other irrelevant features. Simultaneously, it refines and extracts the intensity and morphological information of the spectral signal. This embodiment uses the ratio of red-edge bands and the rate of change of chlorophyll fluorescence peak bands in the spectral data as the inversion features required for the SVR model. The ratio of red-edge bands reflects the "red-edge lifting" effect caused by chlorophyll fluorescence, i.e., the fluorescence intensity of algae, and is a direct indicator of algal activity. The rate of change of chlorophyll fluorescence peak bands quantifies the "steepness" of the fluorescence peak and is related to chlorophyll concentration and physiological state.

[0068] The red-edge band ratio is the ratio of the reflectivity of the 700nm band to that of the 670nm band, and the formula is: ; in, The ratio of the red-edge band. The reflectance is for the 700nm and 670nm bands.

[0069] The rate of change of the chlorophyll fluorescence peak is the first derivative of the reflectance at a wavelength of 685 nm, expressed by the formula: ; in, This represents the change in reflectivity. Indicates wavelength interval, This represents the slope of reflectivity as a function of wavelength at a wavelength of 685 nm. The reflectance at a wavelength of 690nm The reflectance is at a wavelength of 680nm.

[0070] The two-dimensional array expression consisting of the ratio of the red-edge bands and the rate of change of the chlorophyll fluorescence peak bands is as follows: ; This two-dimensional array is used as the inverse feature that the support vector regression model needs to input.

[0071] Random forest models are tree-based ensemble models that excel at handling medium-dimensional features with interactions. This embodiment chooses to construct a three-dimensional array as the input feature for the random forest model. This covers the core optical features of water bodies while avoiding increased computational complexity and improving the speed of training and inversion. In this embodiment, red band reflectance, near-infrared band reflectance, and the red-edge band ratio from the spectral data are used as the inversion features required by the random forest model. These three parameters correspond to three key optical processes in suspended sediment inversion and exhibit interpretable interactions. Red band reflectance reflects the suspended sediment signal in the water body, near-infrared band reflectance reflects the turbidity of the water body, and the red-edge band ratio reflects the algal signal.

[0072] The red band reflectance is the reflectance of the 665nm band, the near-infrared band reflectance is the reflectance of the 780nm band, and the red edge band reflectance is the ratio of the reflectance of the 700nm band to that of the 670nm band, expressed by the following formula: ; in, The ratio of the red-edge band. The reflectance is for the 700nm and 670nm bands.

[0073] The three-dimensional array expression consisting of red band reflectance, near-infrared band reflectance, and the ratio of the red edge band is as follows: ; This three-dimensional array is used as the inversion feature that the random forest model needs to input.

[0074] Gradient boosting tree models are high-performance ensemble models capable of effectively handling high-dimensional heterogeneous features and automatically selecting key variables by ranking feature importance. Therefore, this embodiment chooses to construct a multidimensional array as the input feature form for the gradient boosting tree model, containing more and more comprehensive spectral information, enabling the model to integrate multiple optical indicators and learn global spectral features. The multidimensional array constructed in this embodiment includes reflectance for each band, red-edge band ratio, yellow absorption index, fluorescence height peak, rate of change of chlorophyll fluorescence peak band, and spectral curvature index. Among them: The reflectance of each band is the reflectance of each band from 490nm wavelength to 1020nm wavelength, with each band spaced 10nm apart.

[0075] The red-edge band ratio is the ratio of the reflectivity of the 700nm band to that of the 670nm band, and the formula is: ; in, The ratio of the red-edge band. The reflectance is for the 700nm and 670nm bands.

[0076] The yellow absorption index is the ratio of the reflectance at the 560nm wavelength to that at the 490nm wavelength, and the formula is: ; Where YAI represents the yellow absorption index, The reflectance is for the 560nm and 490nm bands.

[0077] The peak fluorescence height is calculated from the reflectance of the 665nm, 680nm, and 685nm bands. Specifically: the difference between the reflectance of the 685nm and 665nm bands is calculated and designated as the first difference; the difference between the center wavelengths of the 680nm and 665nm bands is calculated and designated as the second difference; the difference between the center wavelengths of the 685nm and 665nm bands is calculated and designated as the third difference; the product of the first and second differences is divided by the third difference, and then added to the reflectance of the 665nm band. The sum is then subtracted from the reflectance of the 680nm band to obtain the final peak fluorescence height. The calculation formula is: ; in, Indicates the peak value of fluorescence height. This indicates the reflectivity in the 665nm wavelength band. This indicates the reflectivity in the 680nm band. This indicates the reflectivity in the 685nm band.

[0078] The rate of change of the chlorophyll fluorescence peak is the first derivative of the reflectance at a wavelength of 685 nm, expressed by the formula: ; in, This represents the change in reflectivity. Indicates wavelength interval, This represents the slope of reflectivity as a function of wavelength at a wavelength of 685 nm. The reflectance at a wavelength of 690nm The reflectance is at a wavelength of 680nm.

[0079] The spectral curvature index is calculated by subtracting the reflectance of the 750nm band from the average reflectance of the 700nm and 800nm ​​bands, using the following formula:

[0080] SCI stands for Spectral Curvature Index. The reflectance values ​​are for the 700nm, 800nm, and 750nm bands, respectively.

[0081] The multidimensional array expression formed by the above features is as follows: ; in, The reflectance of each band within the wavelength range of 490nm to 1020nm is used as the inversion feature required as input to the gradient boosting tree model.

[0082] Step 4: Acquire the hyperspectral reflectance of the target water area to be detected in real time and label it as the first image. Select the pixels that correspond to the spatiotemporal of the measured points as the detection pixels. Simultaneously acquire the latest measured value of suspended sediment concentration. Identify the water body type of the detection pixels through the classifier, call the corresponding optimal inversion model, and output the predicted value of suspended sediment concentration of the detection pixels.

[0083] Step 5: For the detection pixels of different water body types, calculate the average error between the predicted value and the latest measured value of suspended sediment concentration. If the error reaches the set threshold, retrain the optimal inversion model.

[0084] Furthermore, for each detected pixel selected from the first image, its type features are first extracted and input into the trained classifier. Based on its water body type identification result, the corresponding optimal inversion model is called. Then, the inversion features required by the model are extracted and input to obtain the predicted value of suspended sediment concentration for the detected pixel. For each type of water body detected pixel, the sum of the absolute values ​​of the differences between its predicted value of suspended sediment concentration and the latest measured value is calculated and divided by the number of detected pixels of this type of water body to obtain the average error of each water body type. For water body types whose average error reaches a set threshold, the detected pixels and sampled pixels belonging to this type are merged, the corresponding inversion features are extracted as input, and the corresponding measured value of suspended sediment concentration is used as output to retrain the optimal inversion model for this type of water body.

[0085] After training each model, the target water area is subjected to the latest remote sensing detection to acquire its hyperspectral image data, which is then labeled as the first image. In the latest detection, pixels that spatiotemporally correspond to the measured points set in this study are selected as detection pixels to evaluate the training effectiveness of each model.

[0086] First, extract the type features of each detected pixel, as described above. The pre-trained classifier is input to identify the water body type to which the detected pixel belongs. The corresponding optimal inversion model is called according to different water body types. The inversion features required for the optimal inversion model are extracted from the detected pixel. The predicted value of suspended sediment concentration of each detected pixel is output through the optimal inversion model. Each detected pixel has a predicted value of suspended sediment concentration and a measured value of suspended sediment concentration.

[0087] The number of detected pixels for each water body type is counted. The absolute values ​​of the differences between the predicted and measured values ​​for each type of detected pixel are summed, and then divided by the number of detected pixels for each water body type to obtain the average error for each water body type. The calculation formula is as follows: ; in, This represents the average error for water body type C, expressed in mg / L. This indicates the number of detected pixels belonging to water body type C. This represents the measured value of suspended sediment concentration in the i-th detected pixel of water body type C. This represents the predicted suspended sediment concentration of the i-th detected pixel in water body type C.

[0088] This embodiment uses a historical performance reference method to set the retraining threshold. During the optimal model matching phase, cross-validation was performed on each inversion model for each water body type, selecting the model with the smallest average root mean square error. The smallest model is taken as the optimal inversion model. It reflects the average inversion performance index of an inversion model based on sampled pixels of a certain water body type, after K rounds of training and validation. The optimal inversion model selected for each water body type based on this index is the one that minimizes the inversion error of pixels of that water body type. Therefore, it can be used as an indicator to evaluate the subsequent inversion performance. Moreover, as the inversion data is updated, this index will also change dynamically, giving the entire system a good self-updating capability.

[0089] Based on the above theory, the threshold for retraining the optimal model is set as follows: ; in, The retraining threshold for water body type C. Let be the root mean square error of the optimal inversion model for water body type C during training and validation. This is the tolerance coefficient, with a value ranging from 1.5 to 2.0.

[0090] like This indicates that the optimal inversion model for water body type C needs to be retrained.

[0091] At this point, in addition to the historical sampled pixels and their measured values, the newly acquired detected pixels also have ground truth values, and therefore can also be used as training and validation data. Sampled pixels and detected pixels of the same type are merged, and the inversion features required for this type of water body are extracted. The measured values ​​of suspended sediment concentration corresponding to each pixel are used as the output, and the optimal inversion model for this type of water body is retrained.

[0092] Step 6: Extract the type features of all pixels in the first image and input them into the classifier. Based on the water body type identification results, call the corresponding optimal inversion model and output the predicted values ​​of suspended sediment concentration for all pixels. Integrate the output results to obtain the suspended sediment concentration distribution map of the target water area.

[0093] Furthermore, the method for integrating the suspended sediment concentration distribution map is as follows: the predicted values ​​of suspended sediment concentration of all pixels in the first image are arranged into a two-dimensional matrix, corresponding one-to-one with each pixel; according to the preset concentration-color mapping relationship, the numerical matrix is ​​converted into a color matrix; and based on the color matrix, the suspended sediment concentration distribution map of the target water area is generated.

[0094] The specific steps for integrating the suspended sediment concentration distribution map are as follows: Let the first image have a total of There are 100 pixels, and the coordinates of each pixel are 100 pixels. ,in ; Extract the type feature for each pixel: ; Input the type features into the classifier to identify the water body type of each pixel and call the corresponding optimal inversion model; Extract the inversion features required as input for each optimal inversion model and output the predicted suspended sediment concentration for each pixel; The obtained predicted values ​​are arranged according to the spatial location of the pixels. Two-dimensional matrix: ; The suspended sediment concentration was divided into four concentration ranges, and a starting and ending color was assigned to each range: The low concentration range is 0-20 mg / L, and the color of the range is a gradient from light blue to cyan; The medium and low concentration range is 20-100 mg / L, and the color of the range is a gradual change from green to yellow; The medium to high concentration range is 100-300 mg / L, and the color of the range is a gradual change from orange to red; The high concentration range is above 300 mg / L, and the color of the range is dark purple with an RGB value of (139, 0, 139).

[0095] For the first three concentration ranges, the RGB value of each pixel is calculated using linear interpolation: First, based on the predicted suspended sediment concentration *c* of the pixel and the concentration range [a, b] to which this concentration belongs, the proportionality coefficient *t* is calculated using the following formula: ; Linear interpolation calculations were performed on the RGB values ​​of the starting and ending colors of this concentration range: ; ; ; in , , These represent the component values ​​of a pixel with a concentration value of c in the red, green, and blue color channels, respectively. , , These represent the component values ​​of the starting color in the red, green, and blue color channels of the density range to which the pixel belongs; , , These represent the component values ​​of the endpoint color of the density range to which the pixel belongs in the red, green, and blue color channels, respectively. The RGB value of a pixel with a composite concentration value of c is then... .

[0096] Arrange all calculated RGB values ​​according to pixel position as follows Color matrix: ; See Figure 3 As shown, based on the color matrix, the image is rendered, and cartographic elements such as legend, scale bar, and compass are added to generate the final suspended sediment concentration distribution map.

[0097] Please see Figure 4 The present invention also provides a suspended sediment concentration detection device based on hyperspectral remote sensing, which is used to perform the above-described suspended sediment concentration detection method based on hyperspectral remote sensing, including: Data sampling module: used to set measurement points in the target water area to be detected, acquire hyperspectral reflectance images of the target water area at a set time, and select pixels that correspond to the measurement points in time and space as sampling pixels. Simultaneously acquire water body measurement data of each measurement point and associate them with the corresponding sampling pixels. Classifier training module: It is used to set allocation rules based on the measured water body data, assign a water body type label to each sampled pixel, extract type features from the sampled pixels and use them as input, and use the water body type label of the sampled pixels as output to train a classifier. Inversion Model Training Module: This module is used to select the optimal inversion model for each water body type through cross-validation. It extracts the inversion features required by the optimal inversion model from the sampled pixels based on the water body type label and uses them as input. The output is the measured value of the suspended sediment concentration in the actual water body data. This module trains the optimal inversion model for each water body type. Concentration prediction module: It is used to acquire the hyperspectral reflectance of the target water area to be detected in real time and label it as the first image. It selects the pixels that correspond to the spatiotemporal of the measured point as the detection pixels and simultaneously acquires the latest measured value of suspended sediment concentration. It identifies the water body type of each detection pixel through the classifier, calls the corresponding optimal inversion model, and outputs the predicted value of suspended sediment concentration of each detection pixel. Evaluation and Retraining Module: This module is used to calculate the average error between the predicted value and the latest measured value of suspended sediment concentration for detection pixels of different water body types. If the error reaches a set threshold, the optimal inversion model is retrained. Distribution map integration module: used to extract the type features of all pixels in the first image and input them into the classifier. Based on the water body type identification results, it calls the corresponding inversion model, outputs the predicted values ​​of suspended sediment concentration for all pixels, integrates the output results, and obtains the suspended sediment concentration distribution map of the target water area.

[0098] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0099] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0100] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0101] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for detecting suspended sediment concentration based on hyperspectral remote sensing, characterized in that, The specific steps include: Set up measurement points in the target water area to be tested, acquire hyperspectral reflectance images of the target water area at set times, and select pixels that correspond to the measurement points in time and space as sampling pixels. Simultaneously acquire water body measurement data of each measurement point and associate them with the corresponding sampling pixels. Based on the measured water body data, an allocation rule is set, and a water body type label is assigned to each sampled pixel. Type features are extracted from the sampled pixels and used as input. The water body type label of the sampled pixel is used as output to train the classifier. The optimal inversion model is selected for each water body type through cross-validation. The inversion features required by the optimal inversion model are extracted from the sampled pixels according to the water body type label and used as input. The measured value of suspended sediment concentration in the water body measurement data is used as output to train the optimal inversion model for each water body type. The hyperspectral reflectance of the target water area to be detected is acquired in real time and labeled as the first image. Pixels corresponding to the spatiotemporal measurement points are selected as detection pixels. The latest measured value of suspended sediment concentration is acquired simultaneously. The water body type of the detection pixels is identified by the classifier. The corresponding optimal inversion model is called and the predicted value of suspended sediment concentration of the detection pixels is output. For the detection pixels of different water body types, the average error between the predicted value and the latest measured value of suspended sediment concentration is calculated. If the error reaches the set threshold, the optimal inversion model is retrained. Extract the type features of all pixels in the first image and input them into the classifier. Based on the water body type identification results, call the corresponding optimal inversion model and output the predicted values ​​of suspended sediment concentration for all pixels. Integrate the output results to obtain the suspended sediment concentration distribution map of the target water area.

2. The method for detecting suspended sediment concentration based on hyperspectral remote sensing according to claim 1, characterized in that: When setting the measurement points, the target water area to be tested is divided into several sub-regions of equal area, and the center of each sub-region is taken as the measurement point; Multiple time intervals are set, and the time for acquiring the hyperspectral reflectance image is consistent with the water sampling time at each actual measurement point. The measured data of the water body includes the measured values ​​of chlorophyll concentration and suspended sediment concentration.

3. The method for detecting suspended sediment concentration based on hyperspectral remote sensing according to claim 2, characterized in that: The specific method for setting allocation rules based on actual water body measurement data is as follows: Based on the measured values ​​of suspended sediment concentration in the sampled pixels, a first classification threshold and a second classification threshold are set, wherein the first classification threshold is greater than the second classification threshold. Based on the chlorophyll concentration values ​​in the sampled pixels, a third classification threshold is set. For each sampled pixel, if its measured suspended sediment concentration is greater than the first classification threshold, it is labeled as high turbidity saturated type; if its measured suspended sediment concentration is less than the second classification threshold, it is labeled as clear type. For sampled pixels whose measured suspended sediment concentration is less than or equal to the first classification threshold and greater than or equal to the second classification threshold, if its chlorophyll concentration is greater than the third classification threshold, it is labeled as algae-sediment mixture type; otherwise, it is labeled as sediment-dominant type.

4. The method for detecting suspended sediment concentration based on hyperspectral remote sensing according to claim 3, characterized in that: The type features are multidimensional arrays that include red band reflectance, near-infrared band reflectance, red edge band ratio, and fluorescence height peak values ​​on the pixel spectral image.

5. The method for detecting suspended sediment concentration based on hyperspectral remote sensing according to claim 4, characterized in that: The logic for selecting the optimal inversion model for each water body type through cross-validation is as follows: Multiple candidate models are set up. For each type of water body, the pixels belonging to that type of water body are randomly and evenly divided into multiple groups. In each round, one group is selected as the validation set, and the remaining groups are combined into the training set. A candidate model is called, and the inversion features required by the model are extracted from the pixels in the training set and used as input. The corresponding measured value of suspended sediment concentration is used as output to train the candidate model. Subsequently, the same inversion features are extracted from the validation set pixels and input into the candidate model. The predicted values ​​of suspended sediment concentration of the validation set pixels are output. The difference between the predicted value and the measured value of each validation set pixel is squared and then summed. The sum is divided by the number of validation set pixels and then the root mean square error is obtained to obtain the root mean square error of the validation set. This is used as the effectiveness index of this round of validation. Repeat the above steps, selecting a new set of pixels as the validation set in each round, and combining the remaining sets into the training set. Train and validate the same candidate model, so that the training effect of the candidate model is validated once for each set of pixels. After the loop is completed, the sum of the effect indices of each round is divided by the number of rounds to obtain the mean root mean square error of the current candidate model. The mean root mean square error of other candidate models for the same water body type is calculated using the same method. The model with the smallest value is selected as the optimal inversion model for this water body type.

6. The method for detecting suspended sediment concentration based on hyperspectral remote sensing according to claim 5, characterized in that: The candidate models include Bayesian linear regression, support vector regression, random forest, and gradient boosting tree. For each candidate model, the input form of the inversion features is defined according to its algorithm principle.

7. The method for detecting suspended sediment concentration based on hyperspectral remote sensing according to claim 6, characterized in that: The inversion features are constructed by calculating and combining the reflectance of pixels in different spectral bands; For the Bayesian linear regression model, the ratio of blue-green band reflectance is used as the inversion feature that needs to be input. For the support vector regression model, a two-dimensional array is constructed, which contains the ratio of red-edge bands and the rate of change of chlorophyll fluorescence peak bands. This two-dimensional array is used as the inversion feature that the support vector regression model needs to input. For the random forest model, a three-dimensional array is constructed, which contains the red band reflectance, the near-infrared band reflectance, and the red edge band ratio. This three-dimensional array is used as the inversion feature that the random forest model needs to input. For the gradient boosting tree model, a multidimensional array is constructed, which includes reflectance of each band, ratio of red-edge band, yellow absorption index, fluorescence height peak, rate of change of chlorophyll fluorescence peak band, and spectral curvature index. This multidimensional array is used as the inversion feature that the gradient boosting tree model needs to input.

8. The method for detecting suspended sediment concentration based on hyperspectral remote sensing according to claim 7, characterized in that: For each detected pixel selected from the first image, its type features are first extracted and input into the trained classifier. Based on the water body type identification result, the corresponding optimal inversion model is called. Then, the inversion features required by the model are extracted and input to obtain the predicted value of suspended sediment concentration for the detected pixel. For each type of water body detected pixel, the sum of the absolute values ​​of the differences between the predicted value and the latest measured value of suspended sediment concentration is calculated and divided by the number of detected pixels of this type of water body to obtain the average error of each water body type. For water body types whose average error reaches a set threshold, the detected pixels and sampled pixels belonging to this type are merged, the corresponding inversion features are extracted as input, and the corresponding measured value of suspended sediment concentration is used as output to retrain the optimal inversion model for this type of water body.

9. The method for detecting suspended sediment concentration based on hyperspectral remote sensing according to claim 8, characterized in that: The method for integrating the suspended sediment concentration distribution map is as follows: the predicted values ​​of suspended sediment concentration of all pixels in the first image are arranged into a two-dimensional matrix, corresponding one-to-one with each pixel. According to the preset concentration-color mapping relationship, the numerical matrix is ​​converted into a color matrix. Based on the color matrix, the suspended sediment concentration distribution map of the target water area is generated.

10. A device for detecting suspended sediment concentration based on hyperspectral remote sensing, characterized in that: The suspended sediment concentration detection device based on hyperspectral remote sensing is used to perform the suspended sediment concentration detection method based on hyperspectral remote sensing as described in any one of claims 1-9, including: Data sampling module: used to set measurement points in the target water area to be detected, acquire hyperspectral reflectance images of the target water area at a set time, and select pixels that correspond to the measurement points in time and space as sampling pixels. Simultaneously acquire water body measurement data of each measurement point and associate them with the corresponding sampling pixels. Classifier training module: It is used to set allocation rules based on the measured water body data, assign a water body type label to each sampled pixel, extract type features from the sampled pixels and use them as input, and use the water body type label of the sampled pixels as output to train a classifier. Inversion Model Training Module: This module is used to select the optimal inversion model for each water body type through cross-validation. It extracts the inversion features required by the optimal inversion model from the sampled pixels based on the water body type label and uses them as input. The output is the measured value of the suspended sediment concentration in the actual water body data. This module trains the optimal inversion model for each water body type. Concentration prediction module: It is used to acquire the hyperspectral reflectance of the target water area to be detected in real time and label it as the first image. It selects the pixels that correspond to the spatiotemporal of the measured point as the detection pixels and simultaneously acquires the latest measured value of suspended sediment concentration. It identifies the water body type of each detection pixel through the classifier, calls the corresponding optimal inversion model, and outputs the predicted value of suspended sediment concentration of each detection pixel. Evaluation and Retraining Module: This module is used to calculate the average error between the predicted value and the latest measured value of suspended sediment concentration for detection pixels of different water body types. If the error reaches a set threshold, the optimal inversion model is retrained. Distribution map integration module: used to extract the type features of all pixels in the first image and input them into the classifier. Based on the water body type identification results, it calls the corresponding inversion model, outputs the predicted values ​​of suspended sediment concentration for all pixels, integrates the output results, and obtains the suspended sediment concentration distribution map of the target water area.

Images