A water depth inversion method and system based on multispectral remote sensing images

Abstract

The application relates to the technical field of surveying, and discloses a water depth inversion method and system based on multispectral remote sensing images, which uses an independent verification sample set to perform hierarchical precision evaluation, and analyzes model performance differences in dimensions such as water depth range and bottom type. Full-chain uncertainty analysis quantifies error propagation in links such as atmospheric correction and water level correction, and an uncertainty quantification model of ensemble learning is constructed to generate a pixel-level precision distribution map. Therefore, the technical scheme forms a closed-loop dynamic adaptation framework by systematically integrating multi-temporal data dynamic modeling and uncertainty quantification mechanisms, embeds a time and space change compensation mechanism throughout data acquisition to result verification, effectively deals with the interference of the time and space heterogeneity of water body optical properties on water depth inversion, and solves the problems of parameter mismatch and precision reduction of a water depth inversion model caused by the dynamic changes of the optical properties of nearshore water bodies in seasonal and tidal scales.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of surveying, in particular to a water depth inversion method and system based on multi-spectral remote sensing images. BACKGROUND

[0002] Remote sensing science and technology, as an important part of spatial information science, has been widely applied in the field of resource and environment monitoring and spatial data acquisition. As an important branch of remote sensing technology, marine remote sensing focuses on the inversion of key parameters such as ocean color, water temperature and water depth from multi-spectral and hyperspectral image data obtained by satellite or aerial platforms. In shallow topographic mapping and seabed geomorphology research, the water depth inversion technology based on multi-spectral remote sensing images analyzes the water body transmission characteristics of blue and green bands, and establishes an empirical or semi-analytical model of spectral radiation and water depth. The implementation of this technology relies on the accurate modeling of the water body radiation transfer process, in which the inherent optical properties of the water body, including the total suspended matter backscattering coefficient, phytoplankton chlorophyll absorption coefficient and the exponential decay absorption coefficient of colored dissolved organic matter (CDOM), directly determine the light attenuation law in the water body and the transmission efficiency of the bottom reflectance signal.

[0003] The current mainstream multi-spectral water depth inversion method has significant limitations in dealing with the spatial and temporal heterogeneity of water body optical properties. Most of the inversion algorithms used in business applications use a single time-phase set of water body optical parameters for global modeling, ignoring the systematic drift of nearshore water body optical properties on the inter-monthly and seasonal scales caused by land-based material input and biogeochemical processes. For example, a large amount of land-based suspended sediment carried by runoff in the rainy season can increase the attenuation coefficient of coastal water by 2-5 times in a short period of time; the outbreak of diatom blooms in spring can cause the chlorophyll concentration to jump from the winter baseline level (<1mg / m³) to 10-30mg / m³; and the CDOM concentration can differ by several times on both sides of the estuary salinity front. These dynamic changes are systematically ignored or averaged in the existing static modeling framework, resulting in long-term mismatch between model parameters and actual water body conditions.

[0004] The contradiction between the spatio-temporal variation of the water body optical characteristics and the static assumption of the model causes multiple precision degradation phenomena in practical applications. In the monsoon climate dominated estuary and bay area, when the training model in the dry season is applied in the wet season, the increased light attenuation caused by high suspended matter concentration will be misjudged as an increase in water depth, resulting in a systematic positive deviation of 1-3 meters; in the eutrophic lake and reservoir, the strong absorption of chlorophyll to the blue band during the algal bloom in spring and summer causes the spectral response of the same water depth to be significantly reduced, resulting in a "false shallowing" phenomenon in the inversion result. More seriously, in the intertidal zone with strong tidal power, the periodic fluctuation of suspended matter concentration (up to 5-10 times) caused by the rising and falling tides causes the inversion water depth at the same location to differ by 40%-60% at different tides, completely losing the reliability of topographic monitoring, and seriously restricting the practical value of the technology in dynamic water environment.

[0005] Therefore, a water depth inversion method and system based on multispectral remote sensing images are proposed to solve the above problems. SUMMARY

[0006] The purpose of the present application is to provide a water depth inversion method and system based on multispectral remote sensing images to solve the problem that most of the inversion algorithms for business applications ignore the systematic drift of the optical characteristics of nearshore water bodies caused by land-based material input and biogeochemical processes on the inter-monthly and seasonal scales.

[0007] To achieve the above purpose, the present application provides the following technical scheme: a water depth inversion method based on multispectral remote sensing images, the specific steps are as follows:

[0008] S1: Obtain multi-temporal multispectral remote sensing images of the target water area in different seasons and tidal states, perform radiation calibration, atmospheric correction and geometric registration on the multi-temporal multispectral remote sensing images, perform dynamic water level correction based on the astronomical tide model, and unify the multi-temporal multispectral remote sensing images to the average sea level reference surface;

[0009] S2: Extract the water body optical characteristic parameters of the multi-temporal multispectral remote sensing images, divide the optical dynamic zones using a time series clustering algorithm, establish an optical parameter time series database for each optical dynamic zone, and establish an optical parameter prediction model according to the environmental driving factors;

[0010] S3: Design a sampling scheme according to the optical dynamic zones, collect measured data of water depth, suspended matter concentration and chlorophyll concentration in each optical dynamic zone, and supplement sampling points in high uncertainty areas using an active learning strategy;

[0011] S4: Establish a semi-analytical water depth inversion model, express the water body attenuation coefficient as a function of optical component concentration and space-time factor, construct an optical parameter time series prediction module of long short-term memory network, and adjust the parameters of the semi-analytical water depth inversion model according to the output of the optical parameter time series prediction module;

[0012] S5: Establish a bottom type classification system, separate the water body signal and the bottom signal by using a spectral unmixing algorithm, construct a deep learning decoupling network, and invert the bottom reflectivity and water depth through the deep learning decoupling network;

[0013] S6: Temporal and spatial fusion is performed on the multi-period water depth inversion results by using a Bayesian fusion method, a topographic continuity constraint condition is introduced, a water depth change quantity of adjacent periods is calculated and a significance test is performed, and a topographic evolution area is identified;

[0014] S7: Hierarchical precision evaluation and whole-chain uncertainty analysis are performed by using an independent verification sample set, an uncertainty quantification model of ensemble learning is constructed, and a pixel-level precision distribution map is generated.

[0015] Preferably, the step S1 is specifically as follows:

[0016] S1.1: Obtain multi-temporal and multi-spectral remote sensing images of the target water area in the wet season, the dry season and the transition period, record the acquisition time and satellite overpass time of each image, perform radiation calibration, atmospheric correction and geometric registration on the multi-temporal and multi-spectral remote sensing images in turn, convert the original DN value to apparent reflectance by radiation calibration, perform atmospheric correction by using a water body self-adaptive correction algorithm based on the dark pixel principle, and perform geometric registration to register all images to a unified geographic coordinate system, select one image as a radiation reference image, normalize other images to the radiation reference image by histogram matching method, and obtain a multi-temporal standardized image dataset;

[0017] S1.2: Obtain the astronomical tide harmonic constant of the target water area, the astronomical tide harmonic constant includes the amplitude and the lag angle of each component tide, calculate the instantaneous tide level value according to the astronomical tide harmonic constant and the acquisition time of each image by using the astronomical tide prediction formula, set the mean sea level as the water level reference surface, calculate the water level deviation, and establish a pixel-level water level correction amount for each pixel in the multi-temporal standardized image dataset. Subtract the water level deviation from the apparent water depth observation value to obtain the true water depth value corrected to the mean sea level reference surface.

[0018] Preferably, the step S2 is specifically as follows:

[0019] S2.1: Extracting blue band reflectance, green band reflectance and red band reflectance from the multi-temporal normalized image dataset, calculating blue-green band ratio, normalized suspended substance index and chlorophyll fluorescence peak index as water optical characteristic parameters, arranging the water optical characteristic parameters in time sequence for each pixel to construct a time series vector, calculating the similarity distance between the time series vectors of the pixels using dynamic time warping algorithm to form a similarity distance matrix, dividing the pixels into optical dynamic partitions based on the similarity distance matrix using K-means clustering algorithm, calculating the average value of the water optical characteristic parameters of each optical dynamic partition at each time phase to form an optical characteristic time variation curve, and storing it as an optical parameter time series database;

[0020] S2.2: Obtaining the environmental driving factor data corresponding to the multi-temporal normalized image dataset, the environmental driving factor data including daily rainfall, daily runoff, daily average wind speed and daily average water temperature, corresponding the environmental driving factor data to each optical dynamic partition through spatial interpolation method, converting the water optical characteristic parameters into suspended substance concentration value and chlorophyll concentration value as the main control optical parameters through empirical formula, taking the environmental driving factor data as the independent variable and the main control optical parameters as the dependent variable, establishing an optical parameter prediction model using multiple linear regression method, and establishing an independent optical parameter prediction model for each optical dynamic partition.

[0021] Preferably, the step S3 is specifically as follows:

[0022] S3.1: Reading the optical dynamic partition result, calculating the area proportion and fluctuation amplitude of each optical dynamic partition, the fluctuation amplitude being represented by the standard deviation of the water optical characteristic parameters, calculating the initial sampling point number distribution weight according to the area proportion and fluctuation amplitude, distributing the preset total sampling point number to each optical dynamic partition according to the weight, determining the initial sampling point position using stratified random sampling method, measuring the water depth value using a depth finder, collecting water samples and measuring the suspended substance concentration value and chlorophyll concentration value, establishing an initial sampling point data table containing geographic coordinates, sampling time and measured parameters, and performing spatio-temporal matching between the sampling point data and the multi-temporal normalized image dataset, extracting the band reflectance and water optical characteristic parameters of the corresponding pixels to form an initial training sample set;

[0023] S3.2: Divide the initial training sample set into a training subset and a validation subset, train the water depth inversion model and the optical parameter inversion model using the training subset, apply the trained model to the pixels not involved in training to calculate the predicted values, calculate the prediction uncertainty of each pixel using the Gaussian process regression method, sort the pixels by prediction uncertainty, select the top N pixels as candidate supplementary sampling points, perform spatial clustering on the candidate sampling points, select the pixel with the highest prediction uncertainty in each cluster as the actual supplementary sampling point, repeat the measurement process of step S3.1 to obtain supplementary sampling data, and add the supplementary data to the initial training sample set to form an expanded training sample set.

[0024] Preferably, the step S4 is specifically as follows:

[0025] S4.1: Establish a semi-analytical water depth inversion model based on radiation transfer theory, express the water surface off-water radiation as the sum of the water scattering contribution and the bottom reflection contribution, express the water attenuation coefficient as a linear combination of the suspended matter absorption coefficient, the chlorophyll absorption coefficient and the colored dissolved organic matter absorption coefficient, each absorption coefficient has a power function relationship with the corresponding concentration, introduce a seasonal correction factor and a tidal correction factor to correct the absorption coefficient baseline value, extract the measured data from the expanded training sample set, use the nonlinear least squares method to solve the minimum value of the objective function to obtain the optimal parameter value, set parameters for each optical dynamic partition, and establish a correspondence table between the optical dynamic partition number and the model parameter group;

[0026] S4.2: Construct an optical parameter time series prediction module of the long short-term memory network, which includes an input layer, 2-4 layers of LSTM hidden layers and a fully connected output layer, the input layer receives historical 3-12 time phase water optical feature parameters and environmental driving factors, generates LSTM training data set by time sliding window method, trains the model using Adam optimization algorithm, inputs the historical data before the target time phase into the trained model to obtain the predicted value of suspended matter concentration and chlorophyll concentration, calculates the dynamic water attenuation coefficient according to the predicted value and the space-time correction factor, and substitutes it into the semi-analytical water depth inversion model to calculate the water depth of the pixel.

[0027] Preferably, the step S5 is specifically as follows:

[0028] S5.1: Establish a bottom type classification system including sandy bottom, muddy bottom, rocky bottom and seagrass bed bottom, screen shallow water area sampling points with water depth less than 5 meters to collect bottom samples, measure spectral reflectance in 400-900 nm band by using laboratory spectrometer, establish bottom type spectral library, establish pure water body spectral model for image elements in deep water area with water depth greater than 20 meters, decompose observation spectral vector into linear weighted combination of pure water body endmember spectrum and bottom endmember spectrum by using linear spectral unmixing algorithm, solve mixing proportion coefficient by using least squares method, decompose into water body spectral contribution vector and bottom spectral contribution vector according to mixing proportion coefficient, and identify bottom type by using spectral angle mapping method;

[0029] S5.2: Construct a deep learning decoupling network including water body attenuation feature extraction branch and bottom reflectance feature extraction branch, two branches respectively include 3-5 layers of convolutional layers and pooling layers to output feature vectors, and the feature vectors are weighted and modulated through an attention mechanism module, a feature fusion layer is used to splice the feature vectors to form a fused feature vector, a fully connected decoding layer is used to output water depth prediction value, bottom reflectance prediction value and bottom type probability distribution, a composite loss function including data-driven loss term and physical constraint loss term is constructed, an Adam optimization algorithm is used to train the network, and the trained network is applied to shallow water area image elements for water depth inversion and bottom type identification.

[0030] Preferably, the step S6 is specifically as follows:

[0031] S6.1: Collect multi-period water depth inversion results, extract water depth inversion values of each time phase for each image element to form time series data, establish a Bayesian fusion framework, use multi-period water depth inversion values as observation data, construct observation model and state transition model, introduce spatial smoothness constraint and slope rationality constraint as Bayesian prior distribution, encode spatial smoothness constraint as Markov random field prior distribution, and encode slope rationality constraint as slope prior distribution, solve posterior probability distribution by using Markov chain Monte Carlo method or variational Bayesian method, and extract posterior mean value from the posterior distribution as the fused water depth value, and extract posterior standard deviation as the fusion uncertainty index;

[0032] S6.2: Select two adjacent time interval fused water depth products, calculate water depth change for each image element, calculate the uncertainty of water depth change according to error propagation law, construct significance test statistic as the ratio of water depth change to its uncertainty, set the significance level to 0.05 or 0.01 to determine the critical value, and determine the significant change image element when the absolute value of the test statistic is greater than the critical value, perform spatial clustering analysis on the significance test results, set the minimum cluster area threshold to remove fragmented areas, calculate the area, average water depth change and change type of each topographic evolution area, and establish a topographic evolution area database.

[0033] Preferably, the step S7 is specifically as follows:

[0034] S7.1: Collect an independent verification sample set, the sample quantity of which is 20% to 30% of the total sample quantity, stratify the independent verification sample set according to the water depth range, the bottom type and the optical dynamic partition, extract the water depth inversion value of the pixel corresponding to the verification sample point position from the fusion water depth product, calculate the root mean square error, the mean absolute error, the mean deviation and the determination coefficient of each stratified subset, decompose the water depth inversion process into five links of atmospheric correction, water level correction, optical parameter estimation, model structure selection and parameter setting, calculate the uncertainty of each link respectively, synthesize the overall uncertainty according to the error propagation theory, and calculate the contribution rate of each link to the overall uncertainty;

[0035] S7.2: Construct an ensemble learning uncertainty quantification model including 5 to 20 base models, the base model types including semi-analytical model, deep learning network, random forest regression, support vector regression and gradient boosting regression, apply the trained base model to each pixel to calculate the water depth prediction value, calculate the mean value of the base model prediction value as the ensemble prediction value, calculate the standard deviation as the prediction uncertainty, use the independent verification sample set to establish an uncertainty calibration function, adopt piecewise linear interpolation or cubic spline interpolation method, apply the calibration function to each pixel to generate the calibrated prediction error estimation value, generate the pixel-level precision distribution map, calculate the confidence interval according to the confidence level, and generate the confidence interval width distribution map.

[0036] The application also includes a water depth inversion system based on multi-spectral remote sensing images, which comprises an image preprocessing module, an optical partition module, a sample collection module, a model inversion module, a bottom decoupling module, a Bayesian fusion module and a precision evaluation module.

[0037] The image preprocessing module is used to obtain multi-temporal multi-spectral remote sensing images in different seasons and tidal states, perform radiation calibration, atmospheric correction, geometric registration and water level dynamic correction on the images, and unify the images to the average sea level reference.

[0038] The optical partition module is used to extract water body optical characteristic parameters, divide the optical dynamic partition by using a time series clustering algorithm, establish a partition optical parameter time series database and construct an optical parameter prediction model.

[0039] The sample collection module is used to design a sampling scheme according to the optical partition, collect water depth and optical component measured data, and supplement sampling points in high uncertainty areas by using an active learning strategy.

[0040] The model inversion module is used to establish a semi-analytical water depth inversion model, combine the time series prediction to output dynamic adjustment of model parameters, and realize water depth calculation.

[0041] The substrate decoupling module is used for establishing a substrate classification system, a spectral unmixing algorithm is used for separating water body and substrate signals, and a deep learning network is used for jointly inverting substrate reflectivity and water depth;

[0042] The Bayesian fusion module is used for spatio-temporal fusion of multi-period water depth inversion results, spatial smoothing and slope constraints are introduced, water depth changes are calculated, and significant tests are performed to identify topographic evolution areas;

[0043] The precision evaluation module is used for hierarchical precision evaluation and uncertainty analysis based on an independent validation sample set, and an integrated learning uncertainty quantification model is constructed to generate pixel-level precision distribution and confidence interval maps.

[0044] Compared with the prior art, the present application has the following advantages:

[0045] 1. Hierarchical precision evaluation is performed using an independent validation sample set, and model performance differences are analyzed in dimensions such as water depth range and substrate type. The whole-chain uncertainty analysis quantifies the error propagation of atmospheric correction, water level correction, etc., and an integrated learning uncertainty quantification model is constructed to generate pixel-level precision distribution maps. Thus, this technical solution forms a closed-loop dynamic adaptation framework by systematically integrating multi-temporal data dynamic modeling and uncertainty quantification mechanisms, and embeds a spatio-temporal change compensation mechanism throughout the data acquisition and result verification process, effectively addressing the interference of spatio-temporal heterogeneity of water body optical properties on water depth inversion, and solving the problem of model parameter mismatch and precision degradation caused by the dynamic changes of nearshore water body optical properties in seasonal and tidal scales.

[0046] 2. The initial training sample set is divided into a training subset and a validation subset to provide independent data support for model evaluation. The prediction model established using the training subset can identify potential knowledge blind spots. The Gaussian process regression method is used to calculate the prediction uncertainty, which can effectively quantify the prediction confidence and focus on the most uncertain areas of the model. By spatial clustering of candidate sampling points and selecting the most informative points in each cluster, the efficiency of supplementary sampling is maximized. By iteratively integrating new data, the sample set gradually adapts to the complex evolution of water body optical properties, significantly improving the representativeness of the training data and the model generalization ability. The whole scheme solves the problem of insufficient sample representativeness in optical dynamic partitioning through dynamic optimization of sampling resource allocation and active learning mechanism, ensuring that the measured data can accurately capture the spatio-temporal changes of water body optical properties and provide a high-quality training basis for the water depth inversion model. BRIEF DESCRIPTION OF DRAWINGS

[0047] Figure 1 The method steps of the present application are shown in the figure.

[0048] Figure 2 The system flowchart of the present application is shown in the figure. DETAILED DESCRIPTION

[0049] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the present application.

[0050] Embodiment one: please refer to Figure 1 A water depth inversion method based on multispectral remote sensing images, the specific steps are as follows:

[0051] S1: Obtain multispectral remote sensing images of different seasons and tidal states of the target water area, perform radiation calibration, atmospheric correction and geometric registration on the multispectral remote sensing images, perform water level dynamic correction based on the astronomical tide model, and unify the multispectral remote sensing images to the average sea level datum;

[0052] S2: Extract the water optical characteristic parameters of the multispectral remote sensing images, divide the optical dynamic zones by using a time series clustering algorithm, establish an optical parameter time series database of each optical dynamic zone, and establish an optical parameter prediction model according to environmental driving factors;

[0053] S3: According to the optical dynamic zoning design, the sampling scheme is designed, and the measured data of water depth, suspended matter concentration and chlorophyll concentration are collected in each optical dynamic zone, and the active learning strategy is used to supplement the sampling points in the high uncertainty area;

[0054] S4: Establish a semi-analytical water depth inversion model, express the water body attenuation coefficient as a function of the optical component concentration and the space-time factor, construct an optical parameter time series prediction module of the long short-term memory network, and adjust the parameters of the semi-analytical water depth inversion model according to the output of the optical parameter time series prediction module;

[0055] S5: Establish a bottom type classification system, separate the water body signal and the bottom signal by using a spectral unmixing algorithm, construct a deep learning decoupling network, and invert the bottom reflectivity and water depth through the deep learning decoupling network;

[0056] S6: Temporal and spatial fusion is performed on the multi-period water depth inversion results by using a Bayesian fusion method, a topographic continuity constraint condition is introduced, the water depth change amount of adjacent periods is calculated and significance test is performed, and a topographic evolution area is identified;

[0057] S7: Use an independent verification sample set to perform hierarchical precision evaluation and whole-chain uncertainty analysis, construct an uncertainty quantification model of ensemble learning, and generate a pixel-level precision distribution map.

[0058] In this embodiment: The multi-temporal and multi-spectral remote sensing image can be understood as image data with multiple band information obtained from satellite or aerial platform, which can be obtained by Landsat series satellites, Sentinel series satellites or other remote sensing devices with multi-spectral imaging capability. The process of radiation calibration can convert the original digital value into physical quantity, such as apparent reflectance, by using the sensor parameters calibrated in the laboratory. The specific implementation of atmospheric correction includes but is not limited to correction algorithm based on dark pixel principle, MODTRAN radiation transfer model simulation method or 6S model correction method. The realization of astronomical tide model can calculate the tidal component by harmonic analysis method, or predict the tidal level change by numerical simulation method.

[0059] The division of optical dynamic zoning can be realized by various clustering algorithms, such as hierarchical clustering algorithm, density clustering algorithm or fuzzy C mean clustering algorithm, the purpose of which is to divide the water area according to the time series homology of water optical characteristics. The establishment of optical parameter prediction model can use machine learning methods such as support vector machine regression, random forest regression or artificial neural network, combined with environmental driving factor data to predict the trend of optical parameter change.

[0060] The implementation of active learning strategy can be completed in many ways, such as selecting the area with larger prediction error as the supplementary sampling point based on uncertainty sampling method, or using diversity based sampling method to ensure the spatial distribution uniformity of sampling points. The water body attenuation coefficient in semi-analytical water depth inversion model can be described by various mathematical expressions, such as polynomial fitting, exponential function fitting or piecewise linear interpolation method. The training of long short-term memory network can optimize the model performance by adjusting the number of hidden layers, learning rate or regularization parameter.

[0061] The specific implementation of spectral unmixing algorithm can use non-negative matrix factorization method, independent component analysis method or unmixing algorithm based on physical model to separate water body signal and substrate signal. The construction of deep learning decoupling network can optimize the network structure by adjusting the convolution kernel size, pooling layer step or activation function type. The realization of Bayesian fusion method can solve the posterior probability distribution by Markov chain Monte Carlo sampling method or variational inference method. The coding of topographic continuity constraint condition can be realized by neighborhood pixel difference minimization method or slope gradient smoothing method.

[0062] By systematically integrating multi-temporal data dynamic modeling and uncertainty quantification mechanisms, the spatial and temporal heterogeneity of water optical properties is effectively addressed, which interferes with water depth inversion. The core is to build a closed-loop dynamic adaptation framework, which embeds a temporal and spatial variation compensation mechanism throughout the data acquisition to result verification process, avoiding systematic bias caused by ignoring optical parameter drift in traditional static models. By introducing optical dynamic zoning, optical parameter prediction models, and active learning strategies, the model's adaptability to dynamic water environments is significantly improved. At the same time, by using Bayesian fusion methods and ensemble learning uncertainty quantification models, the spatiotemporal consistency and reliability of multi-period water depth inversion results are achieved, providing technical support for precise applications in dynamic water environments.

[0063] By acquiring multi-temporal multi-spectral remote sensing images of the target water area in different seasons and tidal states, the images are first radiometrically calibrated, atmospherically corrected, and geometrically registered to eliminate interference caused by imaging conditions. Based on the astronomical tide model, the water level is dynamically corrected, and all images are unified to the mean sea level reference surface, ensuring the comparability of multi-temporal data in space and time, and providing consistent basic data sources for subsequent analysis. On this basis, the water optical characteristic parameters of multi-temporal images are extracted, and the time series clustering algorithm is used to divide the optical dynamic zoning, establishing the optical parameter time series database for each zoning. Combined with environmental driving factors such as rainfall, wind speed, and other variables, an optical parameter prediction model is constructed, allowing the model to dynamically adjust the predicted value of the optical parameter based on external environmental changes.

[0064] Within the optical dynamic zoning, a sampling scheme is designed, with the initial number of sampling points allocated based on the area proportion and fluctuation amplitude of the zoning, and active learning strategies used to supplement sampling points in high-uncertainty areas, thereby improving the coverage of dynamic processes by measured data. Subsequently, based on the radiative transfer theory, a semi-analytical water depth inversion model is established, expressing the water body attenuation coefficient as a function of optical component concentrations such as suspended solids and chlorophyll, and temporal and spatial factors, and introducing seasonal and tidal correction factors to enable the model to adapt to the dynamic changes in water state. At the same time, a long short-term memory network optical parameter time series prediction module is constructed, which uses historical optical parameter sequences to predict suspended solids and chlorophyll concentrations, and adjusts the model parameters in real time based on the prediction results, achieving immediate response to the optical state of the water body.

[0065] To further separate the water body signal from the bottom signal, a bottom type classification system is established, and a spectral unmixing algorithm is used to decompose the observed spectrum into a linear combination of pure water body and bottom endmember spectra. By constructing a deep learning decoupling network, the attention mechanism module is used to weight and modulate the water body attenuation characteristics and bottom reflectance characteristics, and finally output the water depth prediction value and bottom reflectance prediction value. In addition, a Bayesian fusion method is used to spatio-temporally fuse the multi-period water depth inversion results, and a topographic continuity constraint is introduced to ensure that the fusion results conform to the topographic evolution law. By calculating the water depth change between adjacent periods and performing significance test, the real topographic evolution area is identified, and the environmental noise and actual change are effectively distinguished.

[0066] Finally, the independent verification sample set is used for stratified precision evaluation, and the model performance difference is analyzed in the dimensions of water depth range and bottom type. The whole-chain uncertainty analysis quantifies the error propagation of atmospheric correction, water level correction, etc., and constructs an uncertainty quantification model of ensemble learning to generate a pixel-level precision distribution map. Thus, through the systematic integration of multi-temporal data dynamic modeling and uncertainty quantification mechanism, the technical scheme forms a closed-loop dynamic adaptation framework, which embeds the spatio-temporal change compensation mechanism throughout the data acquisition to result verification process, effectively deals with the interference of spatio-temporal heterogeneity of nearshore water optical properties on water depth inversion, and solves the problem of parameter mismatch and precision decline of water depth inversion model caused by the dynamic changes of nearshore water optical properties in seasonal and tidal scales.

[0067] Embodiment two: please refer to Figure 1 , the specific way of step S1 is as follows:

[0068] S1.1: Obtain multi-temporal multi-spectral remote sensing images of the target water area in the wet season, dry season and transition period, record the acquisition time and satellite overpass time of each image, and sequentially perform radiation calibration, atmospheric correction and geometric registration on the multi-temporal multi-spectral remote sensing images. Radiation calibration converts the original DN value to apparent reflectance, atmospheric correction uses a water self-adaptive correction algorithm based on the dark pixel principle, and geometric registration registers all images to a unified geographic coordinate system. Select one image as the radiation reference image, normalize other images to the radiation reference image through histogram matching method, and obtain the multi-temporal standardized image dataset.

[0069] S1.2: Obtain the astronomical tide harmonic constant of the target water area, which includes the amplitude and phase angle of each tidal component. According to the astronomical tide harmonic constant and the acquisition time of each image, the instantaneous tide level value is calculated using the astronomical tide prediction formula. The average sea level is set as the water level reference surface, and the water level deviation is calculated. For each pixel in the multi-temporal standardized image dataset, the pixel-level water level correction amount is established, and the apparent water depth observation value is subtracted by the water level deviation to obtain the true water depth value corrected to the average sea level reference surface.

[0070] In this embodiment: multi-temporal multi-spectral remote sensing image refers to multi-spectral image data of the target water area obtained in different time periods, which can be obtained by remote sensing platforms such as Landsat series satellites or Sentinel-2 satellites. Radiometric calibration refers to the process of converting the digital quantization value (DN value) recorded by the sensor into apparent reflectance with physical meaning, which aims to eliminate the influence of sensor response difference on image data. Atmospheric correction refers to the process of removing atmospheric scattering and absorption effects by a specific algorithm to obtain the true reflectance of the ground surface, wherein the water body adaptive correction algorithm based on the dark pixel principle can optimize the correction effect for water body characteristics. Geometric registration refers to the process of unifying image data obtained at different times to the same geographic coordinate system, which aims to eliminate geometric distortion caused by imaging angle, terrain undulation and other factors. Histogram matching method is a radiometric consistency correction technology, which adjusts the brightness distribution of the image to be consistent with the reference image, thereby ensuring the comparability of multi-temporal image data.

[0071] The astronomical tide harmonic constant refers to a set of parameters describing the law of tidal variation, which can be obtained by analyzing long-term tidal level observation data. The astronomical tide prediction formula is a mathematical model based on the principle of harmonic analysis, which can accurately calculate the tidal level at any time according to the harmonic constant and time variable. The pixel-level water level correction amount refers to the tidal level correction value calculated for each pixel position, which aims to reflect the influence of spatial position on tidal level and solve the problem of tidal flat area caused by terrain undulation.

[0072] This scheme realizes dynamic water level correction through a systematic operation process. First, multi-temporal multi-spectral remote sensing images of the target water area in the wet season, dry season and transition period are obtained, and the acquisition time and satellite overflight time of each image are recorded. This operation ensures that the image data covers the key hydrological period and provides a time dimension basis for capturing the seasonal changes in water optical properties. Subsequently, through radiometric calibration, atmospheric correction and geometric registration of multi-temporal multi-spectral remote sensing images in turn, the sensor response difference, atmospheric scattering effect and geometric distortion are eliminated, ensuring the consistency of image data in radiation and space. The water body adaptive correction algorithm based on the dark pixel principle optimizes the atmospheric influence correction for water body characteristics, avoiding the lack of applicability of general algorithms in nearshore waters. Select one image as the radiometric reference image and normalize other images to the radiometric reference image by histogram matching method, which ensures the comparability of reflectance data between different images and lays a reliable foundation for subsequent time series analysis.

[0073] On this basis, by acquiring the astronomical tide harmonic constant of the target water area, and combining the acquisition time of each image, the instantaneous tide value is calculated by using the astronomical tide prediction formula, so that the tide prediction and the image acquisition time are strictly synchronized. The average sea level is set as the water level reference surface, a stable water depth reference standard is established, and the reference surface drift caused by tidal fluctuation is eliminated. The pixel-level water level correction amount is established for each pixel, the influence of spatial position on the tide level is considered, and the problem of spatial heterogeneity of the tide level caused by the terrain undulation in the tidal flat area is solved. Finally, the apparent water depth observation value is subtracted by the water level deviation amount to obtain the true water depth value, so that the water depth data is accurately converted to the unified reference surface, and the reliability of the terrain monitoring in the dynamic water environment is significantly improved.

[0074] Through the above technical scheme, the problem of non-uniform water depth reference surface caused by tidal fluctuation is effectively solved, especially in the area with strong tidal power. Through the pixel-level tide level accurate calculation mechanism, the tide level deviation caused by ignoring the time details and spatial heterogeneity in the traditional method is avoided, and accurate input is provided for subsequent water depth inversion.

[0075] Embodiment three: please refer to Figure 1 , the specific way of step S2 is as follows:

[0076] S2.1: Extract the blue light band reflectivity, green light band reflectivity and red light band reflectivity from the multi-temporal standardized image data set, calculate the blue-green band ratio, normalized suspended matter index and chlorophyll fluorescence peak index as water body optical characteristic parameters, arrange the water body optical characteristic parameters in time sequence for each pixel to construct a time sequence vector, calculate the similarity distance between the time sequence vectors of the pixels by using dynamic time warping algorithm to form a similarity distance matrix, divide the pixels into optical dynamic partitions based on the similarity distance matrix by using K-means clustering algorithm, and form the optical characteristic time variation curve by calculating the average value of the water body optical characteristic parameters of each optical dynamic partition at each phase, and store it as an optical parameter time sequence database;

[0077] S2.2: Obtain the environmental driving factor data corresponding to the multi-temporal standardized image data set, the environmental driving factor data includes daily rainfall, daily runoff, daily average wind speed and daily average water temperature, the environmental driving factor data is corresponded to each optical dynamic partition by using spatial interpolation method, the water body optical characteristic parameters are converted into suspended matter concentration value and chlorophyll concentration value as main control optical parameters by using empirical formula, and the environmental driving factor data is used as independent variable and the main control optical parameter is used as dependent variable, and a multivariate linear regression method is used to establish an optical parameter prediction model, and an independent optical parameter prediction model is established for each optical dynamic partition.

[0078] In this embodiment: the optical dynamic partition refers to the spatial unit divided based on the time series clustering results of water body optical characteristic parameters, which can be realized by K-means clustering algorithm or hierarchical clustering algorithm, aiming to reflect the spatial heterogeneity of water body optical characteristics. The fluctuation amplitude refers to the dynamic change intensity index quantified by the standard deviation of water body optical characteristic parameters, which can be realized by the standard deviation calculation method or the coefficient of variation analysis method in statistics, aiming to identify the key areas with intense dynamic changes. The initial sampling point number allocation weight refers to the sampling resource allocation basis calculated by combining the area proportion and the fluctuation amplitude, which can be realized by the weighted average method or the normalization processing method, aiming to ensure higher sampling density in high dynamic areas. The stratified random sampling method refers to the sampling strategy of random point distribution in each partition, which can be realized by the systematic sampling method or the stratified proportional sampling method, aiming to avoid human selection bias while ensuring the uniformity of spatial distribution.

[0079] This scheme first reads the optical dynamic partition results divided in the early stage, providing a spatial framework for targeted sampling. On this basis, the area proportion and fluctuation amplitude of each partition are calculated to quantify the dynamic change intensity of each partition, which enables the identification of key areas with small area but intense changes. Then, the fluctuation amplitude is included in the weight calculation system to realize the optimal allocation of limited sampling resources, making the sampling resources preferentially cover the areas with intense changes and improving the overall sample's representation ability of water body optical characteristics. The stratified random sampling method is used to determine the initial sampling point position, which not only avoids human selection bias, but also ensures the uniformity of spatial distribution and reduces the influence of local anomalies on the sample set. Key measured parameters are obtained by depth sounder and water sample analysis, and a data table containing spatial and temporal dimension information is established, which facilitates subsequent accurate matching with remote sensing images. The sampling point data and multi-temporal normalized image data set are spatio-temporally matched to extract the band reflectance and water body optical characteristic parameters of the corresponding pixels, forming a high-quality initial training sample set.

[0080] In the subsequent link, the initial training sample set is divided into training subset and validation subset, which provides independent data support for model evaluation. The prediction model established by the training subset can identify potential knowledge blind area. The prediction uncertainty is calculated by Gaussian process regression method, which can effectively quantify the prediction confidence and focus on the most uncertain area of the model. Through spatial clustering of candidate sampling points and selecting the most informative points in each cluster, the efficiency of supplementary sampling is maximized. By iteratively integrating new data, the sample set gradually adapts to the complex evolution of water body optical properties, significantly improving the representativeness of training data and the generalization ability of the model. The whole scheme solves the problem of insufficient sample representativeness in optical dynamic zoning through dynamic optimization of sampling resource allocation and active learning mechanism, ensuring that the measured data can accurately capture the spatio-temporal variation of water body optical properties, and providing a high-quality training basis for water depth inversion model.

[0081] Embodiment four: please refer to Figure 1 , step S3 is specifically as follows:

[0082] S3.1: read the optical dynamic zoning result, calculate the area proportion and fluctuation amplitude of each optical dynamic zoning, the fluctuation amplitude is represented by the standard deviation of water body optical characteristic parameters, calculate the initial sampling point number allocation weight according to the area proportion and fluctuation amplitude, allocate the preset total sampling point number to each optical dynamic zoning according to the weight, determine the initial sampling point position by using stratified random sampling method, measure the water depth value using the depth sounder, collect water sample and measure the suspended matter concentration value and chlorophyll concentration value, establish the initial sampling point data table containing geographic coordinates, sampling time and measured parameters, perform spatio-temporal matching of the sampling point data and the multi-temporal standardized image data set, extract the band reflectance and water body optical characteristic parameters of the corresponding pixels to form the initial training sample set;

[0083] S3.2: divide the initial training sample set into training subset and validation subset, train the water depth inversion model and optical parameter inversion model using the training subset, apply the trained model to the pixels not involved in the training to calculate the prediction value, calculate the prediction uncertainty of each pixel by using Gaussian process regression method, sort the pixels according to the prediction uncertainty, select the first N pixels as candidate supplementary sampling points, spatially cluster the candidate sampling points, select the pixel with the highest prediction uncertainty in each cluster as the actual supplementary sampling point, repeat the measurement process of step S3.1 to obtain supplementary sampling data, add the supplementary data to the initial training sample set to form an expanded training sample set.

[0084] In this embodiment: the optical dynamic zoning refers to the region divided based on the change law of water body optical characteristics in time and space, which can be realized by time series clustering algorithm or spatial clustering algorithm. The purpose of introducing this feature is to provide a scientific basis for the subsequent optimization of the distribution of sampling points, ensuring that the sampling points can cover the main change mode of water body optical characteristics. The fluctuation amplitude refers to the degree of change of water body optical characteristic parameters in the region, which can be measured by standard deviation, variance or other statistical indicators, for quantifying the instability of optical characteristics in the region. The initial sampling point number allocation weight refers to the sampling priority index calculated comprehensively according to the zoning area and fluctuation amplitude, which can be generated by linear weighting, nonlinear function or other mathematical methods to realize the reasonable allocation of sampling resources. The hierarchical random sampling method is a hybrid sampling strategy combining stratified sampling and random sampling, which aims to ensure the uniform distribution of sampling points while taking into account the characteristic differences of different partitions.

[0085] The above scheme realizes the goal of dynamic optimization sampling strategy through a series of ordered steps. First, based on the optical dynamic zoning results, the importance of each partition is quantified by calculating the area ratio and fluctuation amplitude, providing an objective basis for the allocation of sampling points. This process not only considers the spatial range of the partition, but also pays special attention to the change intensity of the optical characteristics, so that the high fluctuation region can obtain higher sampling density. Second, the hierarchical random sampling method is used to determine the initial sampling point position, which not only ensures the spatial uniformity of the sampling points, but also avoids the potential bias of pure random sampling. Then, the key measured parameters are obtained by using the depth sounder and water sample analysis, and are spatio-temporally matched with the multi-temporal standardized image dataset to form a high-quality initial training sample set. On this basis, the sample set is divided into training subset and validation subset, which are used for model training and performance evaluation respectively, ensuring the generalization ability of the model. The prediction uncertainty of each pixel is calculated using the Gaussian process regression method, and the candidate supplementary sampling points are selected based on this, which embodies the core idea of the active learning strategy. By spatial clustering of the candidate sampling points and selecting the pixel with the highest prediction uncertainty in each clustering cluster as the actual supplementary sampling point, the problem of excessive concentration of sampling points is effectively avoided, and the diversity and representativeness of spatial coverage are improved. Finally, by iteratively supplementing the sampling data, the training sample set is gradually optimized, significantly enhancing the adaptability and precision of the model in dynamic water body environment.

[0086] The above scheme solves the problem that the initial sampling points cannot fully cover the high fluctuation region of optical characteristics by dynamically adjusting the sampling strategy, and performs targeted supplementary sampling for the region with high model prediction uncertainty, thereby significantly improving the precision and reliability of the water depth inversion model in the spatio-temporally heterogeneous water body environment.

[0087] Embodiment five: please refer to Figure 1 , the specific way of step S4 is as follows:

[0088] S4.1: Establish a semi-analytical water depth inversion model based on radiative transfer theory, express the water surface off-water radiation as the sum of water scattering contribution and bottom reflection contribution, express the water attenuation coefficient as a linear combination of suspended matter absorption coefficient, chlorophyll absorption coefficient and colored dissolved organic matter absorption coefficient, each absorption coefficient has a power function relationship with the corresponding concentration, introduce seasonal correction factor and tidal correction factor to modify the absorption coefficient baseline value, extract measured data from the extended training sample set, solve the minimum value of the objective function to obtain the optimal parameter value using nonlinear least squares method, set parameters for each optical dynamic partition, and establish a correspondence table between optical dynamic partition number and model parameter group;

[0089] S4.2: Construct an optical parameter time series prediction module of long short-term memory network, which includes an input layer, 2 to 4 layers of LSTM hidden layer and a fully connected output layer, the input layer receives historical 3 to 12 time phase water optical characteristic parameters and environmental driving factors, generates LSTM training data set by time sliding window method, trains the model using Adam optimization algorithm, inputs the historical data before the target time phase into the trained model to obtain the predicted value of suspended matter concentration and chlorophyll concentration, calculates the dynamic water attenuation coefficient according to the predicted value and the space-time correction factor, and substitutes it into the semi-analytical water depth inversion model to calculate the water depth of the pixel.

[0090] In this embodiment: the water surface off-water radiation refers to the water surface reflection signal obtained through remote sensing image, which can be realized by using multispectral sensor to capture reflectivity of different wave bands. The water scattering contribution item refers to the light scattering effect caused by the particles in the water body, and the bottom reflection contribution item refers to the reflection signal from the water bottom, which together constitute the main components of the water surface off-water radiation, and the purpose is to accurately separate the water optical process and the physical action of bottom reflection, and lay a theoretical foundation for water depth calculation.

[0091] The water attenuation coefficient refers to the degree of absorption and scattering of light in the water during the propagation process, which can be quantified by the linear combination of suspended matter absorption coefficient, chlorophyll absorption coefficient and colored dissolved organic matter absorption coefficient. This expression can reflect the independent contribution of different optical components to light attenuation, and the purpose is to make the model respond to the complex changes of water composition. In addition, the absorption coefficient has a power function relationship with the corresponding concentration, which reflects the nonlinear relationship between concentration and optical characteristics, and is more consistent with the actual water behavior.

[0092] The seasonal correction factor and the tidal correction factor are key parameters for dynamically compensating the reference value of the absorption coefficient, which can be achieved by statistical analysis or empirical formula based on historical observation data. The purpose is to dynamically adjust the seasonal land source input and the suspended matter fluctuation caused by tides to avoid systematic deviation caused by static parameters.

[0093] The expanded training sample set refers to the measured data set containing multiple time phases and multiple regions obtained by preliminary sampling and supplementary sampling, which can use active learning strategy to supplement sampling points in high uncertainty area to optimize sample distribution. The purpose is to optimize model accuracy by using measured information to ensure that parameters match the actual water body state.

[0094] The correspondence table of optical dynamic zoning number and model parameter group refers to assigning independent model parameter groups to different optical dynamic zoning according to their characteristics, which can be stored and called in the form of database or mapping table. The purpose is to consider the spatial heterogeneity of water optical characteristics, prevent errors at the partition boundary of the global unified parameters, and improve the inversion efficiency.

[0095] The long short-term memory network is a deep learning model suitable for time series prediction, which captures long-term temporal dependencies through multiple LSTM hidden layers. The input layer receives historical 3 to 12 time phase water optical feature parameters and environmental driving factors, and these data can generate continuous time series samples using time sliding window method to enhance the model fitting ability for dynamic process. The Adam optimization algorithm is used to accelerate model convergence and stabilize the training process, the purpose is to improve the prediction accuracy and training efficiency.

[0096] The above technical solutions realize the dynamic expression of water attenuation coefficient through the deep integration of physical model and time series prediction. First, the semi-analytical water depth inversion model based on radiation transfer theory decomposes the water surface off-water radiation into water scattering contribution and bottom reflection contribution, thereby accurately separating the physical effects of water optical process and bottom reflection. Second, by expressing the water attenuation coefficient as a linear combination of suspended matter absorption coefficient, chlorophyll absorption coefficient and colored dissolved organic matter absorption coefficient, and introducing a power function relationship, the independent contribution of different optical components to light attenuation is quantified. On this basis, the introduction of seasonal correction factor and tidal correction factor effectively compensates the seasonal land source input and the suspended matter fluctuation caused by tides, avoiding systematic deviation caused by static parameters.

[0097] The measured data of the extended training sample set optimizes the objective function by a nonlinear least square method to obtain optimal parameter values, ensuring that the model parameters match the actual water body state. Parameter setting is performed for each optical dynamic partition, and a correspondence table of optical dynamic partition number and model parameter group is established, which not only considers the spatial heterogeneity of the optical characteristics of the water body, but also improves the inversion efficiency. In terms of optical parameter time series prediction, the long short-term memory network integrates historical data and environmental driving factors to mine the time series dependence law of the optical parameters, generates a dynamic water body attenuation coefficient, and thus adapts to the change of the water body state in real time.

[0098] By the above technical solution, the model mismatch problem caused by the spatial and temporal variation of the optical characteristics of the water body is solved, and the adaptability of the model in the dynamic water body environment is fundamentally improved. For example, during the season replacement or tidal fluctuation, the introduction of the dynamic water body attenuation coefficient effectively avoids the phenomenon that the high suspended matter concentration in the high flow period is misjudged as an increase in water depth or the chlorophyll absorption is enhanced during the algal bloom, causing false lightening, and significantly improves the accuracy and reliability of the water depth inversion.

[0099] Embodiment six: please refer to Figure 1 , the specific way of step S5 is as follows:

[0100] S5.1: a bottom type classification system including sandy bottom, muddy bottom, rocky bottom and seagrass bed bottom is established, bottom samples are collected at sampling points in shallow water areas with a water depth of less than 5 meters, a spectral reflectance in a wave band of 400 nanometers to 900 nanometers is measured by a laboratory spectrometer, a bottom type spectral library is established, a pure water body spectral model is established for pixels in deep water areas with a water depth of more than 20 meters, a linear spectral unmixing algorithm is used to decompose the observed spectral vector into a linear weighted combination of pure water body endmember spectrum and bottom endmember spectrum, a least square method is used to solve the mixing proportion coefficient, the mixing proportion coefficient is decomposed into water body spectral contribution vector and bottom spectral contribution vector, and a spectral angle mapping method is used for bottom type identification;

[0101] S5.2: a deep learning decoupling network including a water body attenuation feature extraction branch and a bottom reflectance feature extraction branch is constructed, the two branches respectively include 3 to 5 layers of convolutional layers and pooling layers to output feature vectors, the feature vectors are weighted and modulated through an attention mechanism module, a feature fusion layer is used to splice the feature vectors to form a fused feature vector, a fully connected decoding layer is used to output water depth prediction value, bottom reflectance prediction value and bottom type probability distribution, a composite loss function including a data-driven loss term and a physical constraint loss term is constructed, an Adam optimization algorithm is used to train the network, and the trained network is applied to shallow water area pixels for water depth inversion and bottom type identification.

[0102] In the present embodiment: the substrate type classification system refers to a classification framework designed for the spectral reflectance differences of different substrates in the visible to near-infrared band, which can be realized by field sampling combined with remote sensing image analysis, and the purpose is to customize the process for each substrate type to reduce systematic bias. The shallow water sampling point refers to the area with water depth less than 5 meters, because the contribution of substrate reflection signal to water surface spectrum cannot be ignored in this depth range, which can be determined by field sampling combined with remote sensing image, and the purpose is to ensure the accuracy of model input data.

[0103] Spectral reflectance refers to the reflection ability of substrate in a specific waveband range, which can be measured by laboratory spectrometer or high-precision field spectrometer, and the purpose is to fully capture the spectral information of substrate characteristics. The substrate type spectral library refers to the database storing standardized substrate spectral data, which can be realized by spectral measurement and sorting of multiple substrate samples, and the purpose is to provide reliable endmember reference for subsequent spectral unmixing.

[0104] The pure water body spectral model refers to the water body optical property benchmark model established based on deep water pixels, which can be realized by statistical modeling of pixels with water depth greater than 20 meters, and the purpose is to provide pure water body optical property reference. Linear spectral unmixing algorithm refers to the technique of decomposing shallow water spectrum into water transmission and substrate reflection signal, which can be solved by least square method or other optimization algorithm, and the purpose is to separate the two signal sources to improve the inversion accuracy.

[0105] Spectral angle mapping method refers to a classification method based on the similarity measurement of spectral vector angle, which can be used to identify substrate types by calculating the angle between spectral vectors, and the purpose is to reduce misjudgment. Deep learning decoupling network refers to a double branch network structure including water body attenuation feature extraction branch and substrate reflection feature extraction branch, which can be realized by multi-layer convolutional neural network, and the purpose is to model water body optical properties and substrate reflection properties independently.

[0106] Composite loss function refers to the optimization objective function combining data-driven loss term and physical constraint loss term, which can be realized by embedding prior knowledge such as radiative transfer equation, and the purpose is to prevent the model from deviating from the physical law.

[0107] The above technical solutions solve the key problem of substrate reflection interference in shallow water depth inversion by systematically separating water and substrate signals. First, the establishment of substrate type classification system can classify the spectral reflectance characteristics of different substrate types such as sandy bottom, muddy bottom, rocky bottom and seaweed bed bottom, avoiding systematic bias caused by uniform model due to substrate signal mixing.

[0108] The shallow water area sampling points are screened and the bottom material samples are collected, and the spectral reflectivity of 400-900 nm wave band is measured by a laboratory spectrometer, so as to comprehensively capture the spectral information of the bottom material characteristics, and provide a reliable reference for subsequent spectral unmixing.

[0109] The water body spectrum model is established by identifying the deep water area pixel, and the pure water body optical characteristic benchmark is provided for the spectral unmixing of the shallow water area. In addition, the shallow water area spectrum is decomposed into water body and bottom material signals by using the linear spectral unmixing algorithm, and the mixing proportion coefficient is solved by the least square method, so as to effectively separate the two signal sources. Finally, the deep learning decoupling network is constructed, the water body attenuation and bottom material reflection characteristics are independently modeled by the double-branch design, and the composite loss function is combined for optimization, so as to significantly improve the accuracy and reliability of the water depth inversion and bottom material classification.

[0110] By the above technical solutions, the problem of bottom material reflection interference in the water depth inversion of the shallow water area is solved, and the synchronous inversion of water depth and bottom material type is realized, so as to improve the practicability and scientificity of the overall solution.

[0111] Embodiment seven: please refer to Figure 1 , and the specific manner of step S6 is as follows:

[0112] S6.1: Collecting multi-period water depth inversion results, extracting the water depth inversion values of each time phase for each pixel to form time sequence data, establishing a Bayesian fusion framework, taking the multi-period water depth inversion values as observation data, constructing an observation model and a state transition model, introducing spatial smoothness constraints and slope rationality constraints as Bayesian prior distribution, coding the spatial smoothness constraints as Markov random field prior distribution, and coding the slope rationality constraints as slope prior distribution, solving the posterior probability distribution by using Markov chain Monte Carlo method or variational Bayesian method, and extracting the posterior mean value as the fused water depth value and the posterior standard deviation as the fusion uncertainty index from the posterior distribution;

[0113] S6.2: Selecting two adjacent time interval fused water depth products, calculating the water depth change for each pixel, calculating the uncertainty of the water depth change according to the error propagation law, constructing the significance test statistic as the ratio of the water depth change to its uncertainty, setting the significance level to 0.05 or 0.01 to determine the critical value, determining the significant change pixel when the absolute value of the test statistic is greater than the critical value, performing spatial clustering analysis on the significance test results, setting the minimum clustering area threshold to remove fragmented areas, calculating the area, average water depth change and change type of each topographic evolution area, and establishing a topographic evolution area database.

[0114] In this embodiment: the Bayesian fusion framework refers to a probability inference method based on Bayes' theorem, which can be implemented in various mathematical modeling methods, such as by constructing a complex probabilistic graphical model or using approximate inference algorithms. The purpose of introducing this framework is to integrate multi-period water depth inversion results, thereby effectively suppressing noise interference in single-period inversion and improving estimation stability.

[0115] The spatial smoothness constraint can be understood as a prior assumption reflecting the continuity of natural terrain, which can be expressed by a probability model such as Markov random field, aiming to eliminate the influence of random noise on the inversion result. The slope rationality constraint is to ensure that the terrain change conforms to the natural law and avoid unreasonable abrupt changes, which can be achieved by defining a reasonable slope distribution function.

[0116] The significance test statistic refers to a quantitative index for evaluating whether the water depth change has statistical significance, which can be constructed by various statistical methods, such as t-test or z-test, etc. The purpose of introducing this statistic is to distinguish between real terrain evolution and random fluctuations, thereby improving the reliability of terrain monitoring.

[0117] The above scheme provides a time series basis for subsequent analysis by collecting multi-period water depth inversion results, making the historical change pattern of each spatial position individualized, thereby reflecting the influence of dynamic drift of water body optical properties on the inversion result. The Bayesian fusion framework is established to take multi-period water depth inversion values as observation data, fully utilize all available information to avoid single-period limitations, construct observation models and state transition models to describe the relationship between inversion values and true water depth and the evolution law of water depth over time, and enhance the physical rationality of the model. The spatial smoothness constraint and the slope rationality constraint are introduced as Bayesian prior distribution, aiming at the natural characteristics of spatially continuous terrain, effectively modeling the spatial correlation between adjacent pixels to suppress random noise and ensure that the terrain change conforms to the natural slope law and avoids unreasonable abrupt changes.

[0118] The Markov Chain Monte Carlo method or the variational Bayesian method is used to solve the posterior probability distribution, which efficiently handles high-dimensional probability distributions to accurately estimate the fusion result, provides optimal estimation to reduce bias, and quantifies the estimation reliability of each pixel.

[0119] The two-period fusion water depth products selected with adjacent time intervals focus on the changes in the continuous period, avoiding long-term cumulative error interference, calculating the water depth change for each pixel to directly measure the terrain evolution amplitude, and ensuring that the change evaluation considers the transmission of original inversion errors.

[0120] The significance test statistic is constructed as the ratio of the water depth change and its uncertainty, which can effectively distinguish the real topographic evolution from random fluctuations. Spatial clustering analysis is performed on the significance test results to merge adjacent significant change pixels into regions, which conforms to the characteristics of topographic evolution usually occurring in patches, and eliminates small-scale noise points to focus on meaningful evolution regions. The area, average water depth change and change type of each topographic evolution region are calculated to quantify the evolution characteristics to support subsequent analysis, and a topographic evolution region database system is established to store the information, which is convenient for long-term dynamic monitoring and management, so as to realize reliable identification and quantitative evaluation of topographic evolution.

[0121] Through the above technical scheme, the problem existing in the process that the Bayesian fusion framework of multi-period water depth inversion results is used to integrate the space-time information and improve the stability of water depth estimation is solved. The scheme effectively integrates multi-period water depth inversion results, solves the problem of unreliable topographic monitoring caused by unstable single-period inversion results in a dynamic water environment, and realizes reliable identification and quantitative evaluation of topographic evolution.

[0122] Embodiment eight: please refer to Figure 1 , the specific way of step S7 is as follows:

[0123] S7.1: Collect an independent validation sample set, the number of samples in the independent validation sample set is 20% to 30% of the total sample size, stratify the independent validation sample set according to the water depth range, bottom type and optical dynamic partition, extract the water depth inversion value of the pixel corresponding to the validation sample point position from the fusion water depth product, calculate the root mean square error, mean absolute error, mean deviation and determination coefficient of each stratified subset, divide the water depth inversion process into five links: atmospheric correction, water level correction, optical parameter estimation, model structure selection and parameter setting, calculate the uncertainty of each link, synthesize the overall uncertainty according to the error propagation theory, and calculate the contribution rate of each link to the overall uncertainty;

[0124] S7.2: Construct an ensemble learning uncertainty quantification model including 5 to 20 base models, the base model types include semi-analytical model, deep learning network, random forest regression, support vector regression and gradient boosting regression, apply the trained base model to each pixel to calculate the water depth prediction value, calculate the mean value of the base model prediction value as the ensemble prediction value, calculate the standard deviation as the prediction uncertainty, use the independent validation sample set to establish an uncertainty calibration function, use piecewise linear interpolation or cubic spline interpolation method, apply the calibration function to each pixel to generate calibrated prediction error estimate value, generate pixel-level precision distribution map, calculate confidence interval according to confidence level, and generate confidence interval width distribution map.

[0125] In the present embodiment: the independent verification sample set refers to a set of independent data used to evaluate the accuracy of water depth inversion results, which can be achieved by using field measurement data or high-precision reference data, and the purpose is to verify the generalization ability of the model through independent data.

[0126] The hierarchical processing refers to classifying the verification sample set according to environmental characteristics such as water depth range, bottom type and optical dynamic partition, which can be achieved by clustering algorithm or rule division, and the purpose is to evaluate the accuracy of different environmental conditions.

[0127] Error propagation theory refers to a mathematical method for describing the error propagation law in multi-step processing, which can be achieved based on Taylor expansion or Monte Carlo simulation, and the purpose is to reveal the influence path of errors at each link on the final result.

[0128] The integrated learning uncertainty quantification model refers to capturing the uncertainty of complex systems through the combination of multiple base models, which can be achieved by weighted average or Bayesian inference, and the purpose is to enhance the robustness of uncertainty estimation.

[0129] The above technical solution effectively solves the error distribution misalignment problem caused by environmental dynamics through systematic hierarchical verification and whole-chain error analysis mechanism. First, the hierarchical processing of the independent verification sample set can evaluate the accuracy of specific environmental conditions, avoiding the phenomenon of global average masking local accuracy degradation, especially suitable for the drift of water optical properties caused by tide and seasonal change.

[0130] Secondly, the water depth inversion process is divided into five key links and the uncertainty is calculated respectively, which reveals the error propagation path of each link based on radiation transfer theory and error propagation law, making the uncertainty quantification cover the complete chain from raw data to final product. On this basis, the total uncertainty is synthesized by error propagation theory, ensuring the physical consistency of uncertainty estimation and avoiding the error amplification or offset caused by simple superposition. In addition, the integrated learning uncertainty quantification model uses the average and standard deviation of the prediction values of multiple base models to capture the complex influence of dynamic changes of water optical parameters, enhancing the robustness of uncertainty estimation. Finally, the calibration function is applied by piecewise linear interpolation or cubic spline interpolation method, which finely adjusts the nonlinear error characteristics under different confidence levels, ensuring that the calibration process adapts to the local characteristics of water depth and bottom type changes.

[0131] Through the above technical solution, the pixel-level accuracy distribution map and confidence interval width distribution map are generated, providing users with location-dependent reliability evaluation, supporting accurate identification of high uncertainty areas in dynamic water environment, thereby significantly improving the practical value of water depth inversion results in topography monitoring.

[0132] The application also comprises a water depth inversion system based on multispectral remote sensing images, please refer to Figure 2 The system comprises an image preprocessing module, an optical zoning module, a sample collection module, a model inversion module, a bottom material decoupling module, a Bayesian fusion module, and an accuracy evaluation module.

[0133] The image preprocessing module is used to obtain multi-temporal multispectral remote sensing images in different seasons and tidal states, and perform radiation calibration, atmospheric correction, geometric registration, and water level dynamic correction on the images to unify the images to the average sea level reference.

[0134] The optical zoning module is used to extract water body optical characteristic parameters, divide optical dynamic zones using a time series clustering algorithm, establish a time series database of zoned optical parameters, and construct an optical parameter prediction model.

[0135] The sample collection module is used to design a sampling scheme according to the optical zoning, collect water depth and optical component measured data, and supplement sampling points in high uncertainty areas using an active learning strategy.

[0136] The model inversion module is used to establish a semi-analytical water depth inversion model, combine time series prediction to output dynamic adjustment of model parameters, and realize water depth calculation.

[0137] The bottom material decoupling module is used to establish a bottom material classification system, separate water body and bottom material signals using a spectral unmixing algorithm, and realize joint inversion of bottom material reflectivity and water depth using a deep learning network.

[0138] The Bayesian fusion module is used to perform spatio-temporal fusion on multi-period water depth inversion results, introduce spatial smoothing and slope constraints, calculate water depth changes, and perform significance test to identify topographic evolution areas.

[0139] The accuracy evaluation module is used to perform hierarchical accuracy evaluation and uncertainty analysis based on an independent verification sample set, construct an integrated learning uncertainty quantification model to generate pixel-level accuracy distribution and confidence interval maps.

[0140] The contents not described in detail in the specification belong to the prior art known to those skilled in the art.

[0141] Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements to part of the technical features, and any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for water depth inversion based on multi-spectral remote sensing image, characterized in that: The specific steps are as follows: S1: Obtain multi-temporal and multi-spectral remote sensing images of different seasons and tidal states of the target water area, perform radiation calibration, atmospheric correction and geometric registration on the multi-temporal and multi-spectral remote sensing images, perform water level dynamic correction based on an astronomical tide model, and unify the multi-temporal and multi-spectral remote sensing images to an average sea level reference surface; S2: Extract water body optical characteristic parameters from the multi-temporal and multi-spectral remote sensing images, divide the optical dynamic zones by using a time series clustering algorithm, establish an optical parameter time series database for each optical dynamic zone, and establish an optical parameter prediction model according to environmental driving factors; The specific manner of the step S2 is as follows: S2.1: Extract blue light band reflectivity, green light band reflectivity and red light band reflectivity from the multi-temporal standardized image dataset, calculate blue-green band ratio, normalized suspended matter index and chlorophyll fluorescence peak index as water body optical characteristic parameters, arrange the water body optical characteristic parameters in time sequence for each pixel to construct a time series vector, calculate the similarity distance between the time series vectors of the pixels by using a dynamic time warping algorithm to form a similarity distance matrix, divide the pixels into optical dynamic zones by using a K-means clustering algorithm based on the similarity distance matrix, and statistically form optical characteristic time variation curves by taking the average of the water body optical characteristic parameters of each optical dynamic zone at each phase to store as an optical parameter time series database; S2.2: Obtain environmental driving factor data corresponding to the multi-temporal standardized image dataset, the environmental driving factor data including daily rainfall, daily runoff, daily average wind speed and daily average water temperature, correspond the environmental driving factor data to each optical dynamic zone by using a spatial interpolation method, convert the water body optical characteristic parameters into suspended matter concentration values and chlorophyll concentration values as main control optical parameters by using an empirical formula, take the environmental driving factor data as independent variables and the main control optical parameters as dependent variables, establish an optical parameter prediction model by using a multiple linear regression method, and establish an independent optical parameter prediction model for each optical dynamic zone; S3: Design a sampling scheme according to the optical dynamic zones, collect measured data of water depth, suspended matter concentration and chlorophyll concentration in each optical dynamic zone, and supplement sampling points in high-uncertainty areas by using an active learning strategy; S4: Establish a semi-analytical water depth inversion model, express the water body attenuation coefficient as a function relationship of optical component concentration and space-time factors, construct an optical parameter time sequence prediction module of a long short-term memory network, and adjust the parameters of the semi-analytical water depth inversion model according to the output of the optical parameter time sequence prediction module; S5: Establish a bottom type classification system, separate water body signals and bottom signals by using a spectral unmixing algorithm, construct a deep learning decoupling network, and invert the bottom reflectivity and water depth by using the deep learning decoupling network; S6: Temporally and spatially fuse the multi-period water depth inversion results by using a Bayesian fusion method, introduce a topographic continuity constraint condition, calculate the water depth change amount of adjacent periods and perform a significance test, and identify a topographic evolution area; S7: Perform stratified accuracy evaluation and full-chain uncertainty analysis using independent validation sample set, build uncertainty quantification model for ensemble learning, and generate pixel-level accuracy distribution map. 2.The water depth inversion method based on multispectral remote sensing image according to claim 1, characterized in that: The step S1 is specifically as follows: S1.1: Obtain multi-temporal and multi-spectral remote sensing images of the target water area in the wet season, dry season and transition period, record the acquisition time and satellite passing time of each image, and sequentially perform radiation calibration, atmospheric correction and geometric registration on the multi-temporal and multi-spectral remote sensing images. The radiation calibration converts the original DN value into apparent reflectivity. The atmospheric correction adopts a water self-adaptive correction algorithm based on the dark pixel principle. The geometric registration registers all images to a unified geographic coordinate system. Select one image as the radiation reference image. Normalize other images to the radiation reference image through histogram matching method to obtain a multi-temporal standardized image dataset; S1.2: Obtain the astronomical tide harmonic constant of the target water area. The astronomical tide harmonic constant includes the amplitude and phase angle of each component. According to the astronomical tide harmonic constant and the acquisition time of each image, the instantaneous tide value is calculated by using the astronomical tide prediction formula. The average sea level is set as the water level reference surface. The water level deviation is calculated. The pixel-level water level correction amount is established for each pixel in the multi-temporal standardized image dataset. The apparent water depth observation value is subtracted by the water level deviation to obtain the true water depth value corrected to the average sea level reference surface.

3. The method according to claim 2, wherein: The step S3 is specifically as follows: S3.1: Read the optical dynamic zoning result, calculate the area proportion and fluctuation amplitude of each optical dynamic zoning, the fluctuation amplitude is represented by the standard deviation of the water optical characteristic parameter, calculate the initial sampling point number distribution weight according to the area proportion and fluctuation amplitude, distribute the preset total sampling point number to each optical dynamic zoning according to the weight, determine the initial sampling point position by using the stratified random sampling method, measure the water depth value using the depth sounder, collect the water sample and measure the suspended matter concentration value and chlorophyll concentration value, establish the initial sampling point data table containing geographic coordinates, sampling time and measured parameters, match the sampling point data with the multi-temporal standardized image dataset in time and space, extract the corresponding pixel band reflectance and water optical characteristic parameter to form the initial training sample set; S3.2: Divide the initial training sample set into training subset and validation subset, train the water depth inversion model and optical parameter inversion model using the training subset, apply the trained model to the pixels not participating in the training to calculate the predicted value, calculate the prediction uncertainty of each pixel by using the Gaussian process regression method, sort the pixels according to the prediction uncertainty, select the first N pixels as candidate supplementary sampling points, perform spatial clustering on the candidate sampling points, select the pixel with the highest prediction uncertainty in each clustering cluster as the actual supplementary sampling point, repeat the measurement process of step S3.1 to obtain supplementary sampling data, and add the supplementary data to the initial training sample set to form an expanded training sample set.

4. The water depth retrieval method based on multispectral remote sensing images according to claim 3, characterized in that: The step S4 is specifically as follows: S4.1: Based on the radiation transfer theory, a semi-analytical water depth inversion model is established, the water surface radiance is expressed as the sum of the water scattering contribution and the bottom reflection contribution, the water attenuation coefficient is expressed as the linear combination of the suspended matter absorption coefficient, the chlorophyll absorption coefficient and the colored dissolved organic matter absorption coefficient, the power function relationship between each absorption coefficient and the corresponding concentration is introduced, the seasonal correction factor and the tidal correction factor are introduced to modify the baseline value of the absorption coefficient, the measured data is extracted from the extended training sample set, the nonlinear least squares method is used to solve the minimum value of the objective function to obtain the optimal parameter value, the parameter setting is carried out for each optical dynamic partition, and a corresponding relationship table of the optical dynamic partition number and the model parameter group is established; S4.2: An optical parameter time sequence prediction module of a long short-term memory network is constructed, the long short-term memory network includes an input layer, 2 to 4 layers of LSTM hidden layers and a fully connected output layer, the input layer receives the historical 3 to 12 time phase water optical characteristic parameters and environmental driving factors, the LSTM training data set is generated by the time sliding window method, the model is trained by using the Adam optimization algorithm, the historical data before the target time phase is input into the trained model to obtain the suspended matter concentration prediction value and the chlorophyll concentration prediction value, the dynamic water attenuation coefficient is calculated according to the prediction value and the space-time correction factor, and the semi-analytical water depth inversion model is substituted to perform water depth inversion calculation on the pixel.

5. The method according to claim 4, wherein: The step S5 is specifically as follows: S5.1: A bottom type classification system including sandy bottom, muddy bottom, rocky bottom and seagrass bed bottom is established, bottom samples are collected at sampling points in shallow water areas with a water depth of less than 5 meters, a laboratory spectrometer is used to measure the spectral reflectance in a wave band of 400 nanometers to 900 nanometers, a bottom type spectral library is established, a pure water body spectral model is established for pixels in deep water areas with a water depth of more than 20 meters, a linear spectral unmixing algorithm is used to decompose the observed spectral vector into a linear weighted combination of pure water body endmember spectrum and bottom endmember spectrum, the mixing proportion coefficient is solved by using the least squares method, the mixing proportion coefficient is decomposed into a water body spectral contribution vector and a bottom spectral contribution vector, and the spectral angle mapping method is used for bottom type identification; S5.2: A deep learning decoupling network including a water body attenuation feature extraction branch and a bottom reflection feature extraction branch is constructed, the two branches respectively include 3 to 5 layers of convolution layers and pool layers to output feature vectors, the feature vectors are weighted and modulated through an attention mechanism module, the feature fusion layer is used for splicing the feature vectors to form a fusion feature vector, the fully connected decoding layer is used for outputting the water depth prediction value, the bottom reflectivity prediction value and the bottom type probability distribution, a composite loss function including a data-driven loss term and a physical constraint loss term is constructed, the network is trained by using the Adam optimization algorithm, and the trained network is applied to the shallow water area pixels for water depth inversion and bottom type identification.

6. The water depth retrieval method based on multispectral remote sensing images according to claim 5, characterized in that: The step S6 is specifically as follows: S6.1: Collect multi-period water depth inversion results, extract the water depth inversion value of each time phase for each pixel to form time series data, establish a Bayesian fusion framework, take the multi-period water depth inversion value as observation data, construct an observation model and a state transition model, introduce spatial smoothness constraints and slope rationality constraints as Bayesian prior distribution, encode the spatial smoothness constraints as Markov random field prior distribution, encode the slope rationality constraints as slope prior distribution, solve the posterior probability distribution by Markov chain Monte Carlo method or variational Bayesian method, extract the posterior mean value from the posterior distribution as the fused water depth value, and extract the posterior standard deviation as the fusion uncertainty index; S6.2: Select two adjacent time interval fused water depth products, calculate the water depth change for each pixel, calculate the uncertainty of the water depth change according to the error propagation law, construct the significance test statistic as the ratio of the water depth change to its uncertainty, set the significance level to 0.05 or 0.01 to determine the critical value, and determine the significant change pixel when the absolute value of the test statistic is greater than the critical value, perform spatial clustering analysis on the significance test results, set the minimum clustering area threshold to remove fragmented areas, calculate the area, average water depth change and change type of each topographic evolution area, and establish a topographic evolution area database.

7. The method according to claim 6, wherein: The step S7 is specifically as follows: S7.1: Collect an independent validation sample set, the number of samples in the independent validation sample set is 20% to 30% of the total sample size, stratify the independent validation sample set according to the water depth range, bottom type and optical dynamic partition, extract the water depth inversion value of the pixel corresponding to the validation sample point position from the fused water depth product, calculate the root mean square error, mean absolute error, mean deviation and determination coefficient of each stratified subset, decompose the water depth inversion process into five links: atmospheric correction, water level correction, optical parameter estimation, model structure selection and parameter setting, calculate the uncertainty of each link, synthesize the overall uncertainty according to the error propagation theory, and calculate the contribution rate of each link to the overall uncertainty; S7.2: Construct an ensemble learning uncertainty quantification model including 5 to 20 base models, the base model types include semi-analytical model, deep learning network, random forest regression, support vector regression and gradient boosting regression, apply the trained base model to each pixel to calculate the water depth prediction value, calculate the mean value of the base model prediction value as the ensemble prediction value, calculate the standard deviation as the prediction uncertainty, use the independent validation sample set to establish an uncertainty calibration function, use piecewise linear interpolation or cubic spline interpolation method, apply the calibration function to each pixel to generate calibrated prediction error estimates, generate a pixel-level precision distribution map, calculate the confidence interval according to the confidence level, and generate a confidence interval width distribution map. 8.A water depth inversion system based on multispectral remote sensing images, characterized in that: The water depth inversion system based on multi-spectral remote sensing image is used to execute the water depth inversion method based on multi-spectral remote sensing image in any one of claims 1-7, and the system comprises an image preprocessing module, an optical partition module, a sample collection module, a model inversion module, a bottom decoupling module, a Bayesian fusion module, and a precision evaluation module. The image preprocessing module is used for acquiring multi-temporal and multi-spectral remote sensing images in different seasons and tidal states, performing radiation calibration, atmospheric correction, geometric registration and water level dynamic correction on the images, and unifying the images to an average sea level reference; The optical zoning module is used for extracting water body optical characteristic parameters, dividing optical dynamic zones by using a time series clustering algorithm, establishing a time series database of optical parameters of the zones, and constructing an optical parameter prediction model; Blue light band reflectivity, green light band reflectivity and red light band reflectivity are extracted from a multi-temporal standardized image dataset, blue-green band ratio, normalized suspended matter index and chlorophyll fluorescence peak index are calculated as water body optical characteristic parameters, time series vectors are constructed by arranging water body optical characteristic parameters in time sequence for each pixel, a similarity distance matrix is constructed by calculating the similarity distance between time series vectors of pixels by using a dynamic time warping algorithm, pixels are divided into optical dynamic zones by using a K-means clustering algorithm based on the similarity distance matrix, and average values of water body optical characteristic parameters of each optical dynamic zone at each phase are counted to form an optical characteristic time variation curve, which is stored as an optical parameter time series database; Environment driving factor data corresponding to the multi-temporal standardized image dataset is acquired, the environment driving factor data includes daily rainfall, daily runoff, daily average wind speed and daily average water temperature, the environment driving factor data is corresponded to each optical dynamic zone by using a spatial interpolation method, suspended matter concentration values and chlorophyll concentration values are converted from water body optical characteristic parameters as main control optical parameters by using an empirical formula, and the environment driving factor data is used as an independent variable and the main control optical parameters are used as a dependent variable, a multivariate linear regression method is used to establish an optical parameter prediction model, and independent optical parameter prediction models are established for each optical dynamic zone; The sample collection module is used for designing a sampling scheme according to the optical zoning, collecting water depth and optical component measured data, and supplementing sampling points in high uncertainty areas by using an active learning strategy; The model inversion module is used for establishing a semi-analytical water depth inversion model, and combining a time series prediction to output dynamic adjustment of model parameters to realize water depth calculation; The substrate decoupling module is used for establishing a substrate classification system, separating water body and substrate signals by using a spectral unmixing algorithm, and realizing joint inversion of substrate reflectivity and water depth by using a deep learning network; The Bayesian fusion module is used for spatio-temporal fusion of multi-period water depth inversion results, introducing spatial smoothing and slope constraints, calculating water depth changes, and performing significance test to identify topographic evolution areas; The precision evaluation module is used for developing hierarchical precision evaluation and uncertainty analysis based on an independent verification sample set, constructing an integrated learning uncertainty quantification model to generate a pixel-level precision distribution and a confidence interval map.

Images