A rapid quantification method for sediment deposition in plateau lakes based on hyperspectral unmixing and three-dimensional topography.
By combining hyperspectral unmixing with three-dimensional topography, the problem of spectral interference in sediment deposition areas of high-transparency water bodies was solved, enabling rapid and high-precision quantification of sediment deposition in plateau lakes and providing three-dimensional directional coordinates and data support.
Patent Information
- Application Number
- CN202610989411.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-03
- Publication Date
- 2026-11-10
AI Technical Summary
Existing technologies are unable to effectively eliminate spectral interference from travertine and benthic algae in highly transparent water bodies, resulting in low accuracy in identifying sediment deposition areas. Furthermore, the lack of three-dimensional volume quantification methods makes it difficult to meet the need for rapid and accurate quantification of sediment deposition in plateau lakes.
By combining hyperspectral unmixing with three-dimensional topography, an integrated three-dimensional digital elevation model of land and water is constructed to remove spectral interference from travertine and algae. This model is then macroscopically calibrated using historical scientific research data to achieve rapid and high-precision quantification of sediment deposition.
It enables accurate identification of sediment distribution in highly transparent water bodies, breaking through the limitations of two-dimensional planar distribution to three-dimensional volume quantification, providing direct and accurate three-dimensional directional coordinates and data support, and ensuring the scientific rigor of the quantification results.
Smart Images

Figure CN123071561A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the interdisciplinary fields of ecological environment remote sensing monitoring, hydrogeology and three-dimensional spatial geographic information, and in particular to a rapid quantification method for sediment deposition in plateau lakes based on hyperspectral unmixing and three-dimensional topography. Background Technology
[0002] As a vital component of fragile ecosystems, plateau lakes are increasingly affected by soil erosion and siltation due to extreme rainfall and geological disasters. Timely and accurate quantification of lake siltation is crucial for protecting the landscape value of plateau lakes and developing dredging and restoration plans. Current technologies for monitoring and assessing lake siltation have the following limitations: (1) Spectral interference in highly transparent water bodies is difficult to remove: Karst plateau lakes (such as the Jiuzhaigou Lake Group) have extremely high water transparency, and the lake bottom is widely covered with highly reflective travertine and benthic algae. Conventional remote sensing algorithms are easily affected by spectral interference from the complex bottom material when identifying sediment, which leads to misidentification of native travertine or algae as exogenous sediment, resulting in low identification accuracy and inability to accurately locate sediment deposition areas.
[0003] (2) Lack of physical volume quantification methods: Most studies are still at the stage of "planar analysis" that compares the two-dimensional area changes of lakes using historical images. Even with the introduction of monitoring methods such as drones, there is often a lack of three-dimensional topographic base integrating the surface and underwater, making it difficult to directly convert the "siltation area" observed by remote sensing into "siltation volume" that can be used in engineering.
[0004] (3) The contradiction between the efficiency and accuracy of the computational model: Traditional distributed hydrophysical models (such as InVEST and SWAT) require global grid-based runoff iteration calculations when calculating sediment flux. This consumes a huge amount of computing power when dealing with high-resolution remote sensing data, making it difficult to meet the needs of rapid post-disaster response. At the same time, the lack of long-term measured sediment yield data in high-altitude and cold mountainous areas leads to a lack of objective historical benchmark calibration for the calculation results.
[0005] In summary, although some previous studies have conducted qualitative analyses of lake morphological changes using integrated "space-air-ground" technology, there is still a lack of a rapid assessment method that can effectively eliminate optical interference from the bottom sediment and achieve accurate quantification of three-dimensional volume. Summary of the Invention
[0006] This invention provides a rapid quantification method for sediment deposition in plateau lakes based on hyperspectral unmixing and three-dimensional topography. It aims to integrate hyperspectral endmember unmixing technology with integrated land and water three-dimensional detection technology to remove spectral interference from travertine and algae and accurately map sediment distribution in three-dimensional space. At the same time, it introduces historical scientific research data for macroscopic calibration to achieve efficient and high-precision sediment deposition estimation.
[0007] The method of the present invention is achieved through the following technical solution: This rapid quantification method for sediment deposition in plateau lakes based on hyperspectral unmixing and three-dimensional topography includes the following steps: Step S1: Three-dimensional lake basin base construction: Basic spatial data is obtained by using air-water collaborative detection technology, and a high-precision integrated three-dimensional digital elevation model of water and land is constructed through multi-feature fusion algorithm and nonlinear smoothing model to establish the physical volume reference surface of the lake. Step S2: Hyperspectral feature endmember extraction and sediment spectrum unmixing: Acquire high-resolution hyperspectral images covering the target lake, construct a karst high-permeability water body feature endmember spectral library, use principal component analysis to lock the feature bands that distinguish between sediment and non-sediment, and use a linear spectral mixing model to unmix the high-permeability mixed pixels in the shallow water area one by one to generate a "high-precision two-dimensional distribution map of sediment". Step S3: Three-dimensional spatial mapping and preliminary calculation of sediment physical volume: The extracted "high-precision two-dimensional distribution map of sediment" is used as a mask and vertically projected onto the "integrated three-dimensional lake basin base" constructed in step S1; the elevation profile of the high-abundance sediment-covered area is extracted in three-dimensional space, and combined with the reference water depth of the adjacent non-silting area, the initial sediment deposition volume estimate of the area is calculated by integration. Step S4: True value calibration based on historical scientific research lake capacity: Introduce historical scientific research benchmark data as macro constraints, obtain the actual capacity of the target lake based on historical benchmarks, and obtain the sedimentation volume by subtracting the current actual capacity volume calculated in step S3 from the historical scientific research data.
[0008] Furthermore, in step S4, after the truth value is calibrated, there is also a truth value output step, that is, the total sediment volume is accurately filled into the current three-dimensional lake basin topography according to the "high-precision two-dimensional distribution map of sediment" in step S2, and finally a three-dimensional sediment volume map with accurate geographic coordinates is output.
[0009] Furthermore, in step S1, the high-precision integrated land and water three-dimensional digital elevation model is obtained by fusing the surface elevation model of the survey area and underwater bathymetry data. The specific steps for obtaining the surface elevation model of the survey area are as follows: (1) Constructing a feature matrix by combining the advantages of multiple sources: Using high-definition orthophotos provided by photogrammetry to obtain multispectral features, these features are mapped onto a lidar point cloud with high-precision three-dimensional coordinates and multi-echo penetration capability to construct a "spectral-geometric" multidimensional feature matrix. For each lidar point P i The three-dimensional coordinates and echo counts were extracted, and the Normalized Difference Vegetation Index (NDVI) was calculated by combining the image weights.
[0010] In the above formula, i represents the i-th pixel currently being calculated; NDVI i The normalized difference vegetation index (NIR) represents the value of the i-th pixel. i R represents the observed reflectance of the i-th pixel in the near-infrared band; i This represents the observed reflectance value of the i-th pixel in the red light band; (2) Vegetation stripping based on joint constraint matrix: Establish a joint decision matrix of "multi-echo-vegetation index" to strip vegetation, and formulate cross-validation rules: if a point satisfies NDVI>T NDVI If it is not the last echo, then the point is forcibly classified as a non-ground point; if the point is the last echo, it satisfies NDVI. <T NDVI If the last echo point NDVI > T, it is directly determined as an exposed surface point; NDVI If so, then substitute it into the local morphological filter for further determination; (3) Adaptive threshold calculation: The optimal vegetation extraction threshold T is dynamically calculated using the maximum inter-class variance method. NDVI To overcome the differences in ambient light and seasonal phenology, the threshold T that maximizes the inter-class variance is found by traversing the histogram. NDVI =arg maxσ 2 (t), where the inter-class variance σ 2 The formula for calculating (t) is:
[0011] The meanings of the parameters in the above formula are as follows: σ 2 (t) represents the inter-class variance, quantifying the degree of difference between the two data parts when the segmentation threshold is set to t; t is the current segmentation threshold, which forcibly divides the entire dataset into "Category 0" and "Category 1"; P0 is the probability or weight of data being classified as "Category 0"; P1 is the probability or weight of data being classified as "Category 1"; P0 + P1 = 1 must be satisfied; μ0 is the mean of data in "Category 0"; μ1 is the mean of data in "Category 1"; μ is the global mean, which is the overall average of all data points participating in the calculation within the survey area. Finally, the clean ground points are used to generate a water elevation model of the survey area through Kriging interpolation.
[0012] Furthermore, in step S1, the underwater depth sounding data is acquired in the following way: The underwater depth sounder was used to perform a full-coverage scan using an unmanned surface vessel equipped with a multibeam echo sounder, and the sound velocity profiler was used to correct the sound wave refraction error. In order to address the satellite signal obstruction in the canyon, a GNSS / IMU dual-system combined navigation and Kalman filtering were used to calculate and compensate the trajectory. Then, an adaptive trend surface filter was used to remove pseudo-random noise such as benthic algae, and the underwater depth sounding data was output.
[0013] Furthermore, in step S1, let the elevation value of the water surface elevation model (DEM) of the survey area at any overlapping grid point (X, Y) within the transition zone be Z. above The elevation value of the underwater depth sounding (DTM) data is Z. under The vertical distance from this point to the land transition boundary is d. land The vertical distance to the underwater transition boundary is d. water ; Nonlinear smoothing fusion of above-water and underwater topographic data includes the following steps: (1) Vertical reference plane unification: After deducting the vertical offset distance (Δh) from the center of the unmanned vessel antenna to the draft line of the still water surface and performing attitude compensation, the instantaneous absolute water surface elevation H is obtained. water = Z top -Δh, thus determining the depth D of the underwater measuring point. sonar Unified conversion to absolute elevation:
[0014] (2) Constructing a spatial transition zone and distance-weighted smoothing: Taking the measured shoreline as the center, extend the distance L1 towards the land and L2 towards the water to construct a spatial transition zone with a total width of L=L1+L2. Use a nonlinear smoothing step-weighted algorithm based on cubic Hermite polynomials to construct normalized distance parameters:
[0015] The value of u ranges from [0,1]. When the point is close to the underwater boundary, u approaches 0, and when it is close to the land boundary, u approaches 1. To ensure the continuity of the first derivative of the terrain at the transition zone boundary, i.e., the slope, a first derivative continuity constraint is introduced to construct nonlinear weighting coefficients. :
[0016] Finally, the formula for calculating the elevation Z of any overlapping point within the transition zone is as follows: .
[0017] Furthermore, in step S2, after acquiring a high-resolution hyperspectral image covering the target lake, the hyperspectral image is preprocessed. The preprocessing steps include, but are not limited to, atmospheric correction and radiometric calibration.
[0018] Furthermore, in step S2, the construction of the characteristic end-member spectral library of karst high-permeability water bodies is achieved by extracting the spectral curves of pure water bodies, typical exposed travertine, typical benthic algae, and pure exogenous sediment input, and using principal component analysis to identify the characteristic bands that distinguish sediment from non-sediment.
[0019] Furthermore, in step S2, a linear spectral mixing model is used to demix each high-transparency mixed pixel in the shallow water area, forcibly stripping and masking the spectral contribution rates of travertine and algae, and extracting the purity abundance value of exogenous sediment, which is between 0 and 1, thereby generating a "high-precision two-dimensional distribution map of sediment".
[0020] The beneficial effects of this invention are as follows: (1) Precisely remove spectral interference from complex substrates in highly permeable water bodies: This invention innovatively introduces hyperspectral feature demixing technology into the quantitative identification of sediment in karst lakes, successfully achieving effective separation of exogenous sediment signals from native substrate signals such as high-reflectivity travertine and benthic algae, overcoming the misjudgment problem that is easily generated by the traditional remote sensing index method in highly permeable water bodies, and greatly improving the reliability and anti-interference ability of sediment spatial distribution extraction.
[0021] (2) Achieving a quantitative leap from two-dimensional optical features to three-dimensional physical volume: Breaking through the limitation that traditional water color remote sensing can only output the planar distribution (area) of sediment, through multi-source data fusion, the unmixed two-dimensional sediment abundance is directly mapped to the integrated three-dimensional digital terrain base of water and land, transforming the abstract optical features into the concrete three-dimensional sediment deposition volume, providing direct and accurate three-dimensional directional coordinates and data support for dredging projects and ecological restoration; (3) Constructing global error calibration under the condition of no actual measurement station: In view of the technical bottleneck of the lack of long-term sediment observation data in high-altitude and cold mountainous areas, this invention introduces the actual reservoir capacity of historical scientific research as a macro constraint. This mechanism globally corrects the cumulative error generated in the process of optical inversion and three-dimensional spatial interception, makes up for the defect that the sediment flux in areas without hydrological stations is difficult to objectively verify, and ensures the scientificity and rigor of the final measurement results.
[0022] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description
[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will now be described in further detail with reference to the accompanying drawings, wherein: Figure 1 This is a flowchart of the method of the present invention; Figure 2 A high-precision two-dimensional distribution map of sediment; Figure 3 This is a three-dimensional schematic diagram of a specific embodiment; Figures 4-6This is a comparison diagram of the typical end-member spectral characteristics of highly permeable water bodies in a specific embodiment. Detailed Implementation
[0024] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be understood that the preferred embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.
[0025] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.
[0026] like Figure 1 As shown, the present invention provides a rapid quantification method for sediment deposition in plateau lakes based on hyperspectral unmixing and three-dimensional topography. This method is particularly suitable for the rapid quantification of sediment deposition in plateau lakes with complex, deep canyons and Y-shaped bifurcation systems. The method specifically includes the following steps: Step S1: Three-dimensional lake basin base construction: Basic spatial data is acquired using air-water collaborative detection technology, and a high-precision integrated land-water three-dimensional digital elevation model (DEM) is constructed through multi-feature fusion algorithms and nonlinear smoothing models to establish the physical volume datum of the lake. Specifically, the high-precision integrated land-water three-dimensional digital elevation model is obtained by fusing the surface elevation model of the survey area and underwater bathymetry data. The specific steps for acquiring the surface elevation model of the survey area are as follows: (1) Constructing a feature matrix by combining the advantages of multiple sources: Using high-definition orthophotos provided by photogrammetry to obtain multispectral features, these features are mapped onto a lidar point cloud with high-precision three-dimensional coordinates and multi-echo penetration capability to construct a "spectral-geometric" multidimensional feature matrix. For each lidar point P i The three-dimensional coordinates and echo counts were extracted, and the Normalized Difference Vegetation Index (NDVI) was calculated by combining the image weights.
[0027] Where i represents the i-th pixel currently being calculated; NDVI i The normalized difference vegetation index (NIR) represents the value of the i-th pixel. i R represents the observed reflectance of the i-th pixel in the near-infrared band (typically between 0.7 and 1.1 μm); i : Represents the observed reflectance of the i-th pixel in the red light band (typically between 0.6 and 0.7 μm).
[0028] (2) Vegetation stripping based on joint constraint matrix: Establish a joint decision matrix of "multi-echo-vegetation index" to strip vegetation, and formulate cross-validation rules: if a point satisfies NDVI>T NDVI If it is not the last echo, then the point is forcibly classified as a non-ground point; if the point is the last echo, it satisfies NDVI. <T NDVI If the last echo point NDVI > T, it is directly determined as an exposed surface point; NDVI If so, then substitute it into the local morphological filter for further determination; (3) Adaptive threshold calculation: The optimal vegetation extraction threshold T is dynamically calculated using the maximum inter-class variance method. NDVI To overcome the differences in ambient light and seasonal phenology, the threshold T that maximizes the inter-class variance is found by traversing the histogram. NDVI =arg maxσ 2 (t), where the inter-class variance σ 2 The formula for calculating (t) is:
[0029] The meanings of each parameter are as follows: σ 2 (t) represents the inter-class variance, used to quantify the degree of difference between the two data parts when the splitting threshold is set to t. The core objective of the algorithm is to iterate through all possible values of t and find the value that makes σ = 1 / 2 * t * t. 2 (t) represents the maximum value of t; t is the current segmentation threshold. If analyzing vegetation indices, t is a specific NDVI value that is currently being tested, which forcibly divides the entire dataset into "Category 0" and "Category 1".
[0030] P0 represents the probability (or weight) of data being classified as "Category 0". Below this threshold t, it represents the proportion of pixels or point clouds belonging to Category 0 (e.g., low NDVI water bodies, bare soil, travertine, and other non-vegetated areas) to the total data volume of the survey area. P1 represents the probability (or weight) of data being classified as "Category 1," which is the proportion of data belonging to Category 1 (e.g., high NDVI green vegetation areas on both sides of the river). It necessarily satisfies P0 + P1 = 1.
[0031] μ0 is the data mean of "Category 0", which is the average value of all data points classified as non-vegetation.
[0032] μ1 is the data mean of "Category 1", which is the average value of all data points classified as vegetation.
[0033] μ is the global mean, which is the overall average value of all data points involved in the calculation within the survey area. This parameter can also be directly derived from the probabilities and means of the two classes: μ = P0μ0 + P1μ1.
[0034] Finally, the clean ground points are used to generate a water elevation model of the survey area through Kriging interpolation.
[0035] In addition, underwater bathymetry (DTM) data is acquired in the following ways: A multibeam echo sounder was used on an unmanned surface vessel (USV) to perform a full-coverage scan, combined with a sound velocity profiler to correct for sound wave refraction errors. To address satellite signal obstruction in the canyon, a GNSS / IMU dual-system integrated navigation system and Kalman filtering were employed to calculate and compensate for the trajectory. Subsequently, adaptive trend surface filtering was used to remove pseudo-random noise such as benthic algae, thereby obtaining underwater depth sounding (DTM) data.
[0036] In this invention, the elevation value of the water surface elevation model (DEM) of the survey area at any overlapping grid point (X, Y) within the transition zone is set to Z. above The elevation value of the underwater depth sounding (DTM) data is Z. under The vertical distance from this point to the land transition boundary is d. land The vertical distance to the underwater transition boundary is d. water The steps of the nonlinear smoothing fusion method for surface and underwater topographic data are as follows: (1) Vertical reference plane unification: After deducting the vertical offset distance (Δh) from the center of the unmanned vessel antenna to the draft line of the still water surface and performing attitude compensation, the instantaneous absolute water surface elevation is obtained: H water = Z top -Δh, and thus the depth of the underwater measuring point (D) sonar Convert to absolute elevation:
[0037] (2) Constructing a spatial transition zone and distance-weighted smoothing: Taking the measured shoreline as the center, extend the distance L1 towards the land side and L2 towards the water side to construct a spatial transition zone with a total width of L = L1 + L2. In this invention, to eliminate the "cliff-like" steps and slope angles at the splicing points, a nonlinear smoothing step-weighted algorithm based on cubic Hermite polynomials is adopted. Constructing normalized distance parameters:
[0038] The value of u ranges from [0,1]. When the point is close to the underwater boundary, u approaches 0, and when it is close to the land boundary, u approaches 1. To ensure the continuity of the first derivative of the terrain at the transition zone boundary, i.e., the slope, a continuous first derivative constraint is introduced to construct a nonlinear weighting coefficient:
[0039] Finally, the formula for calculating the elevation Z of any overlapping point within the transition zone is as follows: .
[0040] Step S2: Hyperspectral Feature Endmember Extraction and Substrate Spectral Unmixing: Acquire high-resolution hyperspectral images covering the target lake. Construct a spectral library of endmember features for karst high-permeability water bodies. Use principal component analysis to identify characteristic bands that distinguish between sediment and non-sediment. Employ a linear spectral mixing model to unmix high-permeability mixed pixels in shallow water areas one by one, generating a "high-precision two-dimensional sediment distribution map." Specifically, this includes the following steps: Acquire high-resolution hyperspectral images covering the target lake, and perform preprocessing operations on the high-resolution hyperspectral images, including but not limited to radiometric calibration and atmospheric correction. Radiometric calibration is the process of converting raw digital quantization (DN) values recorded by a sensor into radiative values with explicit physical meaning (such as radiance and apparent reflectance). Hyperspectral sensors commonly exhibit non-uniform responses, dark current noise, and inter-pixel gain differences along and across orbits. Radiometric calibration, using calibration coefficients obtained through laboratory calibration, on-board calibration, and site-alternate calibration, converts DN values into radiance or apparent reflectance at the top of the atmosphere, eliminating inconsistencies between the sensor's own electronic offset and response, and establishing a unified radiometric benchmark for subsequent quantitative analysis. For hyperspectral imagery of lakes, radiometric calibration must strictly match the center wavelength and spectral response function of each band to ensure the accuracy of radiative values in narrow bands.
[0041] Atmospheric correction aims to remove the effects of gas molecule absorption, aerosol scattering, and proximity effects in the atmosphere, restoring the top-level atmospheric signal received by the sensor to the true surface reflectance or water radiance. Lakes are typical "dark targets" with extremely low reflectance (below 10% in most bands). Atmospheric path radiation accounts for a large proportion of the total signal, and strong absorption in the near-infrared and short-wave infrared bands renders traditional dark target methods ineffective, posing a challenge to aerosol inversion. Atmospheric correction for lakes typically employs radiative transfer models (such as MODTRAN and 6SV) combined with specialized algorithms for water bodies (such as joint iteration in the short-wave infrared band and bio-optical models) to accurately estimate aerosol optical thickness and water vapor content, decouple atmospheric signals from water color signals, and ultimately obtain the remote sensing reflectance (Rrs), providing reliable surface spectral information for the inversion of water quality parameters such as chlorophyll and suspended matter.
[0042] This step involves systematic preprocessing after acquiring high-resolution hyperspectral imagery covering the target lake to establish a quantitative chain from sensor records to ground physical quantities. Core steps include radiometric calibration and atmospheric correction. Radiometric calibration uses sensor absolute calibration coefficients to convert the raw DN values of pixels into top-level atmospheric radiance. Apparent reflectance is further obtained through sun-earth distance correction and solar zenith angle normalization, thus eliminating radiation distortion introduced by differences in response between different detector pixels and imaging time, ensuring a consistent radiometric reference across all image pixels. Atmospheric correction addresses the weak signal and strong absorption characteristics of lake water. It employs a water-specific radiative transfer model coupled with aerosol inversion to estimate and subtract path radiation from atmospheric molecular scattering and aerosol scattering pixel by pixel. Simultaneously, it corrects for absorption by gases such as water vapor and ozone, as well as proximity effects, ultimately converting apparent reflectance into high-precision surface water reflectance or remote sensing reflectance. This preprocessing process not only effectively restores the intrinsic spectral characteristics of ground objects but also provides a rigorous physical foundation and data guarantee for subsequent quantitative modeling of lake water quality parameters, underwater illumination field inversion, and long-term dynamic monitoring.
[0043] Construct a characteristic end-member spectral library of karst high-permeability water bodies: extract spectral curves of pure water bodies, typical exposed travertine, typical benthic algae, and pure exogenous sediment input, and use principal component analysis (PCA) to identify characteristic bands that distinguish sediment from non-sediment. Specifically, after completing the preprocessing of hyperspectral images and obtaining the true reflectance of ground features, a spectral library needs to be established for key ground feature categories in the lake environment to support the separation of sediment and non-sediment signals. Based on the optical shallow water characteristics and material composition of the target lake, high-quality spectral curves of pure water, typical exposed travertine, typical benthic algae, and pure exogenous sediment are extracted. Pure water samples are selected from deep water areas in the center of the lake or shaded areas to ensure sufficient water depth and zero visibility of the bottom sediment. Their spectra highlight the low reflectivity of pure water in the visible light band and the strong absorption characteristics in the near-infrared band. Exposed travertine samples are taken from exposed travertine beaches or reefs on the lake shore. They have extremely high overall reflectivity and exhibit unique waveforms at specific mineral absorption valleys. Benthic algae samples need to be extracted from areas with clear, optically shallow water. Generally, the water depth is required to be less than the transparency depth. The spectrum is also mixed with water column signals. However, by selecting algal mats close to the water surface or just submerged, the chlorophyll absorption valleys and near-infrared "red edge" effects of algae can be highlighted to the greatest extent. Exogenous sediment samples come from turbid plumes in estuaries, water bodies with high sediment content, or exposed alluvial sediment on the lake shore. Their spectra often show that the reflectivity in the visible light band increases rapidly with increasing wavelength, lacks obvious absorption characteristics, and still maintains high reflectivity in the near-infrared to short-wave infrared band. Each type of sample underwent abnormal pixel removal and spectral smoothing and noise reduction to construct a representative spectral set that covers both sediment and non-sediment types.
[0044] After obtaining the spectral curves of the typical ground features, principal component analysis (PCA) was used to identify the characteristic bands that distinguish between sediment and non-sediment. First, each group of spectral samples was categorized as sediment (pure exogenous sediment input) and non-sediment (pure water, exposed travertine, benthic algae), and these categories, along with all bands, formed a standardized sample matrix. Principal component transformation was applied to this matrix, extracting the top principal components with the largest cumulative variance contribution. By plotting the scatter plot of the principal components, the degree of separation between sediment and non-sediment samples in the high-dimensional spectral space could be visually examined—typically, the first principal component mainly carries overall brightness information, while the second and third principal components reveal differences in spectral shape. Further analysis of the loading vectors of each principal component revealed that the bands with the highest absolute loading values were the spectral locations that contributed the most to the corresponding principal component. Because sediment exhibits a monotonically increasing, non-absorption continuous spectrum in the visible-near-infrared range, while non-sedimentary landforms show strong absorption in water bodies, mineral absorption in travertine, and chlorophyll fluorescence peaks and red-edge structures in benthic algae, these differences generate significant spectral variance in specific spectral bands and are highlighted on the loading maps of one or more discriminant principal components. Therefore, based on the peak and trough positions of the loadings and the relative differences in reflectance between sediment and non-sedimentary landforms in this band, a set of characteristic bands can be identified. These bands, verified by PCA, demonstrate the highest separability between sediment and non-sedimentary signals, providing a quantitative basis for subsequent construction of sediment indices, band ratios, or classification decision trees.
[0045] A linear spectral mixing model (LSMM) was used to demix high-transparency mixed pixels in shallow water areas one by one, forcibly stripping and masking the spectral contribution rates of travertine and algae, and extracting the purity abundance value (between 0 and 1) of exogenous sediment to generate a "high-precision two-dimensional sediment distribution map". A typical image is shown below. Figure 2 As shown.
[0046] Traditional linear spectral mixing (LSMM) is unsuitable for the physical scenario of highly permeable shallow lakes in this application. Using a traditional LSMM model to unmix highly permeable mixed pixels in shallow water suffers from poor accuracy and low efficiency. To meet the applicable scenario of this application, the traditional two-dimensional planar unmixing model is upgraded to a formulaic model coupled with the previous "three-dimensional water depth data." This is mainly because in highly permeable karst water bodies, the spectral signals of the substrate (such as travertine and sediment) are attenuated by exponential absorption and scattering during transmission. Traditional linear spectral mixing (LSMM) models suffer from severe distortion in abundance inversion due to water depth variations. Therefore, this application proposes a deep-coupled spectral mixing model based on water column attenuation compensation.
[0047] The model directly calls the pixel-by-pixel absolute water depth values provided by the "integrated land and water 3D digital elevation model" constructed in step S1, and its unmixing equation is defined as follows:
[0048] The parameters are defined as follows: Represents the mixed apparent reflectance of pixels received by the drone's hyperspectral sensor; This represents the actual physical abundance (0 ≤ f) of the i-th endmember (such as sediment, travertine, algae) within that pixel. i ≤1 and ∑f i = 1); R i (λ) represents the pure reflectance of the i-th endmember; K d (λ) represents the diffuse attenuation coefficient of the target water body at wavelength λ, which is used to characterize the optical absorption attenuation characteristics of the water body itself; H represents the absolute water depth at the location corresponding to the mixed pixel. This parameter is dynamically provided pixel-by-pixel by the underwater 3D DTM generated in step S1. ε is the residual term.
[0049] This invention overcomes the limitation of conventional remote sensing unmixing algorithms that do not consider hydrological parameters, and innovatively embeds underwater three-dimensional physical morphology (water depth H) directly as a dynamic compensation factor into the spectral characteristic equation. This study provides a precise mathematical model of the two-way water column attenuation process of light, involving "downward transmission - underwater reflection - upward emission." This mechanism effectively filters out spectral variation interference caused by lake basin topographic undulations, achieving high-precision and interference-resistant stripping of the true physical abundance of exogenous sediment in complex underwater non-homogeneous media.
[0050] It should be noted that "forced stripping" means that when calculating sediment abundance, the contribution of travertine abundance and benthic algae abundance to the final sediment map is set to zero—that is, the spectral responses of these two types of substrates are subtracted from the mixed signal, retaining only the effective abundance of endmembers of pure water and pure exogenous sediment. For pixels with travertine abundance or algae abundance higher than a certain threshold (e.g., >60%), they are identified as non-sediment-dominant pixels and directly masked as background values to avoid sediment abundance artifacts caused by model residuals in such areas.
[0051] The sediment abundance values after stripping are normalized to 0-1 for each pixel. 0 indicates that the pixel contains no sediment in the actual sediment composition, while 1 indicates that the sediment is covered by pure exogenous sediment. This abundance value has a clear physical meaning and can be directly interpreted as the coverage ratio or volume percentage of sediment within the field of view (depending on the endmember definition method). The final result is a spatially continuous, finely detailed "high-precision two-dimensional sediment distribution map." The grayscale values on the map transition continuously from 0 to 1, accurately depicting the input path, diffusion gradient, deposition hotspots, and fine structure of sediment plume edges in the shallow water area of the lake. This provides high-resolution spatial explicit data support for identifying the source of exogenous sediment, estimating sediment transport flux, and assessing the evolution of lake sediment deposition.
[0052] Step S3: Three-dimensional spatial mapping and preliminary calculation of sediment physical volume: The extracted "high-precision two-dimensional distribution map of sediment" is used as a mask and vertically projected onto the "integrated three-dimensional lake basin base" constructed in step S1; the elevation profile of the high-abundance sediment-covered area is extracted in three-dimensional space, and combined with the reference water depth of the adjacent non-silting area, the initial sediment deposition volume estimate of the area is calculated by integration. Step S4: True Value Calibration Based on Historical Scientific Expedition Lake Capacity: Historical scientific expedition benchmark data is introduced as a macro-constraint to obtain the actual historical benchmark capacity of the target lake. The difference between the current actual capacity volume calculated in Step S3 and the historical scientific expedition data is used to obtain the sedimentation volume. Furthermore, after true value calibration, a true value output step can be selected. That is, the total sedimentation volume is accurately filled into the current three-dimensional lake basin topography according to the "high-precision two-dimensional sediment distribution map" from Step S2, ultimately outputting a three-dimensional sedimentation volume map with precise geographic coordinates.
[0053] The technical solution of the invention described above will be further illustrated by a specific embodiment below.
[0054] This embodiment selects a typical high-altitude karst lake within the core scenic area of Jiuzhaigou as the study area. The water system in this area is distributed in a "Y" shape, with a significant drop in elevation from south to north. The water in this lake is extremely transparent year-round, with a large area of highly reflective travertine deposits and abundant benthic algae on the lakebed, and a lack of long-term measured sediment hydrological data. The specific implementation steps of this embodiment are as follows: Step S1: Constructing an integrated three-dimensional lake basin base The terrain data was acquired by using a drone equipped with a PM1500 long-range lidar system and a high-precision five-lens tilt camera, while the underwater portion used an iBoat BS15 unmanned surface vessel equipped with a multibeam echo sounder.
[0055] (1) Refined processing and vegetation stripping of the DEM on the water: In the Hi-LiDAR software environment, the original laser point cloud and orthophoto (DOM) were imported. In order to solve the problem of misjudgment where it is difficult to distinguish between low shrubs on both banks and real undulating ground, the software automatically performed the registration of the point cloud and the image, assigned the spectral bands of the DOM to the laser points and calculated the NDVI value. In this embodiment, the fixed threshold was abandoned, and the NDVI of the entire survey area was statistically analyzed to construct a histogram. The variance σ between vegetation and non-vegetation classes was obtained by traversing the Otsu algorithm. 2 (t) Maximum adaptive threshold T NDVI If a point is the last echo of the laser beam and its NDVI... <T NDVI This allows for precise identification of real ground points. This fusion extraction technology effectively penetrates the forest understory to obtain a high-fidelity DEM (Dual Image Model) over water.
[0056] (2) Integrated data fusion of surface and underwater systems: To address the problem of geometric kinks that easily occur when splicing two sets of data, this embodiment uses a cubic Hermite polynomial distance-weighted smoothing algorithm for professional fusion. First, a unified vertical reference surface is established based on the measured elevation of GNSS / RTK. Then, a spatial transition zone is established, with the landward extension distance L1 set as the width of the drawdown zone between the historical highest water level and the current water level (values range from 5 to 15 m), and the waterward extension distance L2 crossing the multipath effect blind zone of multibeam ultra-shallow water (values range from 10 to 20 m). Within the transition zone, the normalized distance parameter u is calculated, and nonlinear weighting coefficients are assigned. (ω=3u 2 -2u 3 This method eliminates abrupt slope changes through adaptive spatial weight allocation, generating a high-precision, integrated three-dimensional lake basin base with continuous physical morphology and an accuracy better than 5cm. A specific schematic diagram is shown below. Figure 3 As shown.
[0057] Step S2: Spectral unmixing and sediment extraction of highly permeable water body substrate (1) In this embodiment, high-resolution UAV hyperspectral images covering the target karst lake are used, with a band range of 400nm-1000nm. Atmospheric correction, radiometric calibration and other preprocessing are performed on them.
[0058] (2) Constructing a characteristic endmember spectral library: Pure pixels (ROIs) of pure water bodies, typical exposed travertine, typical benthic algae, and pure exogenous sediment were selected from the images. A comparison of the typical bottom sediment endmember spectral characteristics of highly permeable water bodies in this embodiment is shown in the figure below. Figures 4-6 As shown in Figure 4, the exogenous sediment end-members exhibit a gradual increase in reflectance from the visible to near-infrared bands, with no bio-absorption characteristics. Figure 5 As exposed travertine end-members, they exhibit high reflectance with broad peaks in the blue-green light band, and their absorption by water shows a significant decreasing trend after 600nm. Figure 6It is a benthic algal terminus with a significant chlorophyll absorption valley around 670 nm and exhibits a typical red edge effect.
[0059] (3) Substrate spectral unmixing: The extracted travertine, algae, and sediment end-member spectra ( Figure 4 As shown, the FCLS demixing tool uses the endmember matrix as input and enforces the physical constraints of ANC (non-negative, ensuring abundance is greater than 0) and ASC (summing to 1). It calculates the abundance value of each sediment element pixel by pixel and extracts the abundance matrix (abundance range 0-1) representing the true coverage of exogenous sediment. This step precisely eliminates optical interference from the native sediment, outputting a high-purity "two-dimensional sediment abundance distribution map".
[0060] Step S3: Three-dimensional spatial mapping and preliminary calculation of sediment physical volume (1) Use the “two-dimensional sediment abundance distribution map” output in step S2 as a spatial mask and project it vertically onto the surface of the “three-dimensional lake basin base with integrated water and land” constructed in step 1 to accurately locate the three-dimensional spatial boundary of sediment deposition. (2) Within the overlapping projection area, the elevation difference of the target pixel is extracted by combining the baseline elevation of the surrounding non-silted area (pure travertine area). The elevation difference is multiplied by the pixel area, and then multiplied by the hyperspectral sediment abundance weight of the pixel. The global spatial integral is then summed to calculate the estimated initial sediment deposition volume in the three-dimensional space.
[0061] Step S4: Global calibration and verification based on historical relative sedimentation rates (1) Calculate the total volume: The total volume of the lake before siltation, based on historical scientific research records, is 900,000 m³. 3 Simultaneously, using the three-dimensional topographic data from step 1, the current measured capacity of the lake was calculated to be 700,000 cubic meters. 3 Subtracting the two directly (original reservoir capacity - current reservoir capacity), the total absolute increase in sediment deposition volume of the lake over the years is calculated to be 200,000 m³. 3 .
[0062] (2) Weighting: Extract the "sediment abundance distribution map" obtained from hyperspectral unmixing in step S2. The higher the abundance value of a pixel, the thicker the sediment in that area. This two-dimensional abundance map will be used as the "sediment spatial distribution weight map".
[0063] (3) Volume Division: The "total sediment volume" calculated in step (1) is accurately filled into the current three-dimensional lake basin topography (i.e., according to the "spatial allocation weight" in step (2)). Figure 3 (As shown in the schematic diagram). The final output is a 3D sediment volume map with accurate geographic coordinates.
[0064] This map not only presents the total siltation volume of the entire lake but also specifies the exact cubic meters of siltation accumulated in each local depression, providing direct guidance for dredging projects. In practical applications, the map can be presented interactively, supporting rotation, cross-sectional cutting, and arbitrary point queries. The total siltation volume of the entire lake is clearly marked on a dynamic statistics panel. This total volume is calculated by integrating the siltation volume of all valid pixels, intuitively reflecting the overall scale of sediment accumulation in the lake. More importantly, the map provides detailed breakdown and annotation of the siltation volume for each local depression, pool, ancient river channel, and harbor basin depression—users can select any area or click on a single point in a low-lying area on the lakebed, and the system will instantly provide the siltation volume at that location, accurate to the cubic meter level. The engineering dredging guidance thus obtains quantitative basis: based on the volume and spatial distribution of silt accumulation in the pits and depressions shown on the map, targeted zonal dredging plans can be designed, the workload of cutter suction dredgers and grab bucket dredgers in each zone can be verified, the routes of dredging vessels and the layout of sludge discharge pipes can be planned, and precise dredging operations of "fixed point, fixed quantity, and fixed depth" can be implemented. While thoroughly restoring the lake's storage capacity and flood discharge capacity, over-excavation can be avoided to prevent damage to the natural impermeable layer or ancient sedimentary structure of the lake bottom, thus achieving a scientific, economical and ecologically sound approach to dredging projects.
[0065] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.
Claims
1. A rapid quantification method for sediment deposition in plateau lakes based on hyperspectral unmixing and three-dimensional topography, characterized by: The method includes: Step S1: Three-dimensional lake basin base construction: Basic spatial data is obtained by using air-water collaborative detection technology, and a high-precision integrated three-dimensional digital elevation model of water and land is constructed through multi-feature fusion algorithm and nonlinear smoothing model to establish the physical volume reference surface of the lake. Step S2: Hyperspectral feature endmember extraction and sediment spectrum unmixing: Acquire high-resolution hyperspectral images covering the target lake, construct a karst high-permeability water body feature endmember spectral library, use principal component analysis to lock the feature bands that distinguish between sediment and non-sediment, and use a linear spectral mixing model to unmix the high-permeability mixed pixels in the shallow water area one by one to generate a "high-precision two-dimensional distribution map of sediment". Step S3: Three-dimensional spatial mapping and preliminary calculation of sediment physical volume: The extracted "high-precision two-dimensional distribution map of sediment" is used as a mask and vertically projected onto the "integrated three-dimensional lake basin base" constructed in step S1; the elevation profile of the high-abundance sediment-covered area is extracted in three-dimensional space, and combined with the reference water depth of the adjacent non-silting area, the initial sediment deposition volume estimate of the area is calculated by integration. Step S4: True value calibration based on historical scientific research lake capacity: Introduce historical scientific research benchmark data as macro constraints, obtain the actual capacity of the target lake based on historical benchmarks, and obtain the sedimentation volume by subtracting the current actual capacity volume calculated in step S3 from the historical scientific research data.
2. The rapid quantification method for sediment deposition in plateau lakes based on hyperspectral unmixing and three-dimensional topography according to claim 1, characterized in that: In step S4, after the true value is calibrated, the true value output step is also included. That is, the total volume of sedimentation is accurately filled into the current three-dimensional lake basin topography according to the "high-precision two-dimensional distribution map of sediment" in step S2, and finally a three-dimensional sedimentation volume map with accurate geographic coordinates is output.
3. The rapid quantification method for sediment deposition in plateau lakes based on hyperspectral unmixing and three-dimensional topography according to claim 1 or 2, characterized in that: In step S1, the high-precision integrated land and water three-dimensional digital elevation model is obtained by fusing the surface elevation model of the survey area and underwater bathymetry data. The specific steps for obtaining the surface elevation model of the survey area are as follows: (1) Constructing a feature matrix by combining the advantages of multiple sources: Using high-definition orthophotos provided by photogrammetry to obtain multispectral features, these features are mapped onto a lidar point cloud with high-precision three-dimensional coordinates and multi-echo penetration capability to construct a "spectral-geometric" multidimensional feature matrix. For each lidar point P i The three-dimensional coordinates and echo counts were extracted, and the Normalized Difference Vegetation Index (NDVI) was calculated by combining the image weights. In the above formula, i represents the i-th pixel currently being calculated; NDVI i The normalized difference vegetation index (NIR) represents the value of the i-th pixel. i R represents the observed reflectance of the i-th pixel in the near-infrared band; i This represents the observed reflectance value of the i-th pixel in the red light band; (2) Vegetation stripping based on joint constraint matrix: Establish a joint decision matrix of "multi-echo-vegetation index" to strip vegetation, and formulate cross-validation rules: if a point satisfies NDVI>T NDVI If it is not the last echo, then the point is forcibly classified as a non-ground point; if the point is the last echo, it satisfies NDVI. <T NDVI If the last echo point NDVI > T, it is directly determined as an exposed surface point; NDVI If so, then substitute it into the local morphological filter for further determination; (3) Adaptive threshold calculation: The optimal vegetation extraction threshold T is dynamically calculated using the maximum inter-class variance method. NDVI To overcome the differences in ambient light and seasonal phenology, the threshold T that maximizes the inter-class variance is found by traversing the histogram. NDVI =arg maxσ 2 (t), where the inter-class variance σ 2 The formula for calculating (t) is: The meanings of the parameters in the above formula are as follows: σ 2 (t) represents the inter-class variance, quantifying the degree of difference between the two data parts when the segmentation threshold is set to t; t is the current segmentation threshold, which forcibly divides the entire dataset into "Category 0" and "Category 1"; P0 is the probability or weight of data being classified as "Category 0"; P1 is the probability or weight of data being classified as "Category 1", which necessarily satisfies P0 + P1 = 1; μ0 is the mean of data in "Category 0"; μ1 is the mean of data in "Category 1"; μ is the global mean, which is the overall average of all data points participating in the calculation within the survey area. Finally, the clean ground points are used to generate a water elevation model of the survey area through Kriging interpolation.
4. The rapid quantification method for sediment deposition in plateau lakes based on hyperspectral unmixing and three-dimensional topography according to claim 3, characterized in that: In step S1, the underwater depth sounding data is acquired in the following way: The underwater depth sounder was used to perform a full-coverage scan using an unmanned surface vessel equipped with a multibeam echo sounder, and the sound velocity profiler was used to correct the sound wave refraction error. In order to address the satellite signal obstruction in the canyon, a GNSS / IMU dual-system combined navigation and Kalman filtering were used to calculate and compensate the trajectory. Then, an adaptive trend surface filter was used to remove pseudo-random noise such as benthic algae, and the underwater depth sounding data was output.
5. The rapid quantification method for sediment deposition in plateau lakes based on hyperspectral unmixing and three-dimensional topography according to claim 4, characterized in that: In step S1, let the elevation value of the water surface elevation model (DEM) of the survey area at any overlapping grid point (X, Y) within the transition zone be Z. above The elevation value of the underwater depth sounding (DTM) data is Z. under The vertical distance from this point to the land transition boundary is d. land The vertical distance to the underwater transition boundary is d. water ; Nonlinear smoothing fusion of above-water and underwater topographic data includes the following steps: (1) Vertical reference plane unification: After deducting the vertical offset distance (Δh) from the center of the unmanned vessel antenna to the draft line of the still water surface and performing attitude compensation, the instantaneous absolute water surface elevation H is obtained. water = Z top -Δh, thus determining the depth D of the underwater measuring point. sonar Unified conversion to absolute elevation: (2) Constructing a spatial transition zone and distance-weighted smoothing: Taking the measured shoreline as the center, extend the distance L1 towards the land and L2 towards the water to construct a spatial transition zone with a total width of L=L1+L2. Use a nonlinear smoothing step-weighted algorithm based on cubic Hermite polynomials to construct normalized distance parameters: The value of u ranges from [0,1]. When the point is close to the underwater boundary, u approaches 0, and when it is close to the land boundary, u approaches 1. To ensure the continuity of the first derivative of the terrain at the transition zone boundary, i.e., the slope, a first derivative continuity constraint is introduced to construct nonlinear weighting coefficients. : Finally, the formula for calculating the elevation Z of any overlapping point within the transition zone is as follows: 。 6. The rapid quantification method for sediment deposition in plateau lakes based on hyperspectral unmixing and three-dimensional topography according to claim 1, characterized in that: In step S2, after acquiring a high-resolution hyperspectral image covering the target lake, the hyperspectral image is preprocessed. The preprocessing steps include, but are not limited to, atmospheric correction and radiometric calibration.
7. The rapid quantification method for sediment deposition in plateau lakes based on hyperspectral unmixing and three-dimensional topography according to claim 1, characterized in that: In step S2, the construction of the characteristic end-member spectral library of karst high-permeability water bodies is achieved by extracting the spectral curves of pure water bodies, typical exposed travertine, typical benthic algae, and pure exogenous sediment input, and using principal component analysis to identify the characteristic bands that distinguish sediment from non-sediment.
8. The rapid quantification method for sediment deposition in plateau lakes based on hyperspectral unmixing and three-dimensional topography according to claim 1, characterized in that: In step S2, a linear spectral mixing model is used to demix each high-transparency mixed pixel in the shallow water area, forcibly stripping and masking the spectral contribution rates of travertine and algae, and extracting the purity abundance value of exogenous sediment, which is between 0 and 1, thereby generating a "high-precision two-dimensional distribution map of sediment".
