CN115376002A - A Multispectral Satellite Remote Sensing Bathymetry Method Based on Stacking Integrated Model - Google Patents

Abstract

The invention discloses a multispectral satellite remote sensing depth measurement method based on a stacking integrated model, and belongs to the field of ocean remote sensing detection. The algorithm comprises the following steps: acquiring a multispectral satellite image of a water area to be measured and a corresponding actual measurement sample data set; establishing a stacking integration model which comprises a base learning layer of a first layer and a generalization layer of a second layer, wherein the base learning layer comprises a plurality of base learners; inputting the preprocessed multispectral satellite image and the actual measurement sample data set into a stacking integration model for training, and acquiring the overall water depth information of the water area to be measured; and carrying out quality evaluation and water depth mapping on the obtained product. The method for combining the training data sets and the machine learning model is effective in estimating the water depth aiming at the remote sensing image and is superior to a single model. Compared with the traditional acoustic depth sounding method, the method has the advantages of lower cost and stronger universality.

Description

A Multispectral Satellite Remote Sensing Bathymetry Method Based on Stacking Integrated Model

technical field

The invention relates to the field of marine remote sensing detection, in particular to a multispectral satellite remote sensing depth sounding method based on a stacking integrated model.

Background technique

Water depth is an important topographic element, bathymetry is the key to obtaining seabed information, and underwater topography is key data for marine construction management and coastal environment monitoring. Through bathymetry, it is possible to detect underwater topography and underwater channel obstacles, and serve offshore economic construction and military struggle.

Ocean bathymetry, as one of the important hydrological characteristics, has always been the most basic work in marine survey and mapping. Shallow water bathymetry has always been a difficult point in the field of marine survey and mapping. Mapping using ship-based acoustic echosounding, an approach capable of producing precise point measurements or depth profiles along transects, is limited by inefficiency, expense, and lack of availability. Because shallow waters restrict navigation safety, especially in areas with large tolerances, traditional shipborne sonar sounding cannot be applied to nearshore shallow waters. Another alternative method is to combine measured water depth data and satellite remote sensing data to measure Improve the prediction of remote sensing water depth, and it will be more accurate with the improvement of spatial resolution and spectral resolution of optical sensors. The emergence of satellite images with spatial resolution higher than 30m, such as Landsat satellite, SPOT and WorldView satellite, makes airborne Satellite Derived Bathymetry (SDB) has gradually become a practical way of sounding.

As a sounding method for bathymetry inversion, empirical methods and physics-based semi-empirical methods have been developed for satellite bathymetry research. Although variations in bottom reflectivity and variations in the optical properties of the water column can affect accurate depth extraction, using multiple bands can improve the robustness of depth estimation. However, bathymetry based solely on multispectral sensors is still problematic in practical applications. Satellite sounding requires supervised training data of certain satellite images, and the sounding error is relatively large. Therefore, it is necessary to combine the measured data and invert the water depth data through the machine learning model to improve the sounding accuracy.

In recent years, scholars at home and abroad have done various researches on the research of water depth inversion for multispectral remote sensing using machine learning models. At present, multiple machine learning methods have been applied to the inversion of remote sensing water depth. Zhang Ying, Wang Yanjiao, etc. introduced the momentum BP neural network model to invert the water depth of the Nangang section of the Yangtze Estuary, and found that the inversion accuracy of the model in waters shallower than 5 meters Higher; Vanesa Mateo-Pérez et al. used the support vector machine model to calculate the water depth of Sentinel-2 remote sensing images in Luarca and Candás ports in Spain, and obtained inversion results close to the actual measurement; Tatsuyuki Sagaw et al. used random forest machine learning and A generalized depth estimation model was created from multi-temporal satellite images, and a total of 135 Landsat-8 images and a large number of training bathymetric data from five regions were analyzed using the Google Earth engine. The accuracy of the SDB was assessed by comparison with reference bathymetric data. The root mean square error of the final estimated water depth in the five test areas is 1.41m, and the depth is 0 to 20m. It can be seen that the calculation of the random forest model is suitable for various shallow water areas under high transparency conditions; Hassan Mohamed uses the least squares enhancement (LSB) An ensemble learning fitting algorithm for computing shallow lake bathymetric maps from high-resolution satellite imagery and bathymetric samples from ecosounders and regressing them with principal component analysis (PCA) and generalized linear models (GLM) Compared with the methods, it can be seen that the integrated learning algorithm has a higher potential than a single model.

Most of the existing machine learning models use the same classifier, or multiple classifiers from the same source are integrated together. This integration method can be called homologous integration. For homologous integration, there is a very important point to note: that is, the errors of models from the same source are related, that is, the causes of each model error are the same, and the factors that cause errors in each model are Similar, then the combination still cannot effectively improve the effect of the model. And because each model has its excellent areas of expertise in different water depth ranges, it is difficult to find a machine learning model that is suitable for all water depth inversion conditions.

Contents of the invention

In view of the existing technical defects, the present invention proposes an integration method using Stacking to perform different weighted outputs for different models, so that the characteristics of each model can be brought into play and better inversion results can be obtained.

In order to achieve the above object, the present invention provides a kind of multispectral satellite remote sensing sounding method based on stacking integrated model, comprising the steps:

(1) Obtain the multispectral satellite image of the water area to be measured and the corresponding measured sample data set, carry out preprocessing for the multispectral satellite image; divide the measured sample data set into a training set and a verification set;

(2) set up the stacking integration model, the stacking integration model has two layers, including the base learning layer of the first layer and the generalization layer of the second layer, and the base learning layer includes a plurality of basic learners;

(3) input the multispectral satellite image and the measured sample data set into the stacking integrated model after preprocessing to obtain the overall water depth information of the water area to be measured;

(4) Carry out quality assessment and bathymetry mapping for the overall water depth information of the water area to be measured.

Further, the preprocessing of the multi-spectral satellite image includes radiation correction, geometric registration and flare elimination.

Further, the multiple basic learners of the basic learning layer include: random forest, support vector machine, GBDT and neural network four water depth inversion machine learning models.

Further, the learner of the generalization layer is MARS multiple adaptive regression spline algorithm.

Further, the training method of the stacking integrated model includes:

(2.1) Input the preprocessed multispectral satellite images and the training set data to a plurality of basic learners of the basic learning layer to perform water depth inversion machine learning, and the result is used as the output of the basic learning layer;

(2.2) using the verification set data to verify and evaluate the output of the base learning layer;

(2.3) Using the output of the basic learning layer as the training data set of the generalization layer for training, and using the verification set data for verification and evaluation, and obtaining the overall water depth information of the water area to be tested.

The present invention also provides a multispectral satellite remote sensing depth-sounding device based on the stacking integrated model, comprising:

Preprocessing module: used to obtain the multispectral satellite image of the water area to be measured and the corresponding measured sample data set, and perform preprocessing on the multispectral satellite image; divide the measured sample data set into a training set and a verification set;

Integrated model modeling module: for setting up stacking integrated model, described stacking integrated model comprises the basic learning layer of the first layer and the generalization layer of the second layer, and described basic learning layer comprises a plurality of basic learners;

Water depth calculation module: used for inputting the preprocessed multispectral satellite image and the measured sample data set into the stacking integrated model to obtain the water depth information of the water area to be measured;

Evaluation and mapping module: used for quality assessment and bathymetric mapping of the overall water depth information of the water area to be measured.

The present invention also provides a device, including a memory, a processor, and a computer program stored in the memory and capable of being executed on the memory. When the processor executes the computer program, the above-mentioned multispectral satellite remote sensing depth sounding based on the stacking integrated model is realized. method.

The present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed in a computer, the computer is made to execute the above-mentioned method for multi-spectral satellite remote sensing depth sounding based on the stacking integrated model.

Beneficial effects of the present invention:

The method of combining multiple training data sets and machine learning models proposed by the present invention is effective in estimating water depth for remote sensing images, and the result of the model method is better than that of a single model. It can successfully estimate water depth from observed satellite imagery, and has the advantages of lower cost and greater versatility than traditional acoustic bathymetry methods. In addition, considering the non-parametric nature of machine learning methods, the integrated model can be applied not only to the estimation of coastal water depth, but also to other observable coastal quantitative values.

Description of drawings

Fig. 1 is a schematic flow chart of a multi-spectral satellite remote sensing depth sounding method based on a stacking integrated model according to an embodiment of the present invention.

Fig. 2 is the comparison chart ((a) random forest (RF) of the inversion result of measured depth and a kind of machine learning method of the embodiment of the present invention; (b) GBDT; (c) neural network; (d) support vector regression (SVR )).

Fig. 3 is a schematic diagram of the correlation between the measured depth and the water depth estimation results based on satellite image data and multiple adaptive regression splines (MARS) for multiple training sets.

Fig. 4 is the result graph of RMSE values under different water depth ranges by different methods.

Figure 5 is a diagram of bathymetric map estimation based on multiple training sets of satellite image data and multiple adaptive regression splines (MARS).

Figure 6 is an image of the measured water depth in the Nanshan Port area of Sanya.

Detailed ways

The present invention will be further explained below in conjunction with the accompanying drawings and embodiments.

As shown in Figure 1, the embodiment of the present invention provides a multi-spectral satellite remote sensing depth-sounding method based on the stacking integrated model, including the following steps:

S101. Acquire multispectral satellite images of the water area to be measured and corresponding measured sample data sets, and perform preprocessing on the multispectral satellite images; divide the measured sample data sets into training sets and verification sets;

Preprocessing of multispectral satellite imagery, including: radiometric correction, geometric registration, and flare removal.

Radiation correction: The storage format of remote sensing images is dimensionless and does not have any physical meaning, so it must be converted into surface reflectance or radiance through radiation correction before it can be used continuously. Considering that the research area is mainly the ocean, the surface is flat, and the radiation distortion caused by terrain factors is not considered. Therefore, the radiometric correction of the images in the study area mainly includes two parts: radiometric calibration and atmospheric correction. Use ENVI software to complete image processing in this patent.

1) Radiation calibration

Radiation calibration is the process of converting the voltage or digital value (DN, Digital Number, also called digital quantization value) recorded by the sensor into absolute radiance. Radiation calibration can be divided into relative radiation calibration and absolute radiation calibration. The purpose of relative radiation calibration is to eliminate the inconsistency of the response of each detection element of the sensor to the target object. Relative radiometric calibration needs to deal with the response difference of each target in the same imaging range, and also deal with the response difference between the original DN values of each scanning zone after imaging. The GF-6 image uses the following formula to perform relative radiometric calibration target.

In the formula, q(λ): represents the gray value after relative radiation correction, the gray value before p(λ) processing, the bias of the dark target when A(λ), and the distance between objects when B(λ) relative gain.

2) Atmospheric correction

The radiance at the top of the atmosphere has been obtained through radiometric calibration, and atmospheric correction is the process of converting the radiance at the top of the atmosphere to the surface radiance. The main purpose of atmospheric correction is to weaken the atmospheric scattering and reflection energy received by the satellite sensor, and to obtain as much real reflectivity information of the surface as possible.

Geometric registration: Geometric correction is to correct the geometric error caused by the coordinate displacement on the remote sensing image. The collinear equation strictly describes the imaging relationship of the image in theory, and it is the most accurate to use it for geometric correction. However, for some high-resolution commercial remote sensing satellites, such as WorldView, IKONOS and other sensors, the information of the sensors is confidential, and the satellite orbit of the imaging method is not public. Without knowing the imaging method and orbit parameters, geometric correction cannot pass Strict geometric model is completed. The rational function model (Rational Function Model, RFM) can obtain a new type of sensor imaging model that is nearly consistent with the accuracy of the strict imaging model. It is a generalized mathematical correction model that combines the image point coordinates (r, c) The coordinates (X, Y, Z) of the ground points are related by ratio polynomials. The RPC model formula is:

In the formula: P(X,Y,Z) is a polynomial function, and the specific expression is:

In the formula: a ₀ , a ₁ , a ₂ ..., a ₁₉ are coefficients of rational functions.

Flare Elimination: When the shallow sea area is affected by wind and waves, sunlight will be reflected by Fresnel on the rough sea surface, and white flares will be generated in the remote sensing image, which will have a certain impact on the retrieval accuracy of shallow sea depth. According to the theory proposed by Lyzenga et al., water bodies have strong absorption characteristics in the near-infrared band. It can be considered that the radiance in the near-infrared band is only composed of atmospheric scattering and solar flares. On the image after atmospheric correction, it is only affected by solar flares. . Select N sample points in the deep water area that does not contain water body information on the image, and the covariance ρ _ij between the visible light band i and the near-infrared band j is expressed as:

Among them, L _in represents the radiance value of the nth sample point in the band i, and L _jn represents the radiance value of the nth sample point in the near-infrared band.

The theory of Lyzenga et al. expresses the formula of flare removal as:

是样本点在近红外波段的平均辐射亮度。Among them: L _i 'is the radiance of band i after removing solar flares, L _i is the radiance value of band i polluted by flares, L _j is the radiance value of near-infrared band,

is the average radiance of the sample point in the near-infrared band.

S102. Establish a stacking integrated model, the stacking integrated model has two layers, including a base learning layer of the first layer and a generalization layer of the second layer, and the base learning layer includes a plurality of basic learners;

The basic learning layer includes four water depth inversion machine learning models: random forest, support vector machine, GBDT and neural network; the generalization layer is the MARS multiple adaptive regression spline algorithm.

Machine learning is a way to improve the ability of models through prior information. Specifically, for a given task and performance metric, use prior information to improve the initial model in a computational way to obtain an improved model with better performance. A prediction model is constructed by using several labeled sample data, so that the prediction output of the prediction model conforms to the real value as much as possible. We call this prediction model a regression model. The sample points used to build the model are called training samples, and the sample points used to check the model results are called inspection samples. The essence of using machine learning is to find a mapping from input data to output data and use this mapping as a predictive model. In the present invention, five machine learning models are used before and after:

1) Random Forest Model (RandomForest, RF)

The random forest model is an integrated supervised learning method. In its algorithm, multiple prediction models are generated at the same time, and the prediction results of each model are comprehensively analyzed to improve the prediction accuracy. The random forest algorithm is designed to sample samples and variables to generate a large number of decision trees, conduct self-service sampling for each tree, and use out-of-bag sample data for error estimation. When generating a decision tree, the variables are randomly selected, so the random forest will not overfit as the number of trees increases. The random forest algorithm can still have an efficient learning rate in the case of large data sets, can calculate the relative importance of variables, and has interpretability for the results. Random forest is the evolution of the idea of Bagging. Random forest adds random selection of features on the basis of random sampling of Bagging samples, and its basic idea does not depart from the category of Bagging.

2) Support Vector Machine (Support Vector Machine, SVM)

Support vector machine (Support Vector Machine, SVM) is a kind of generalized linear classifier (generalized linear classifier) for binary classification of data according to supervised learning. For the linear case, regression is performed directly by the decision function. For nonlinear cases, linear regression is realized by constructing a decision function in a high-dimensional space, which is suitable for constructing a multidimensional small sample regression model. In this experiment, the kernel method was chosen for nonlinear classification.

The regression function of the SVR model is:

Note: φ(x) is a nonlinear mapping function, w is the weight vector of the hyperplane, and b is the bias vector. In order to improve the calculation speed of the test samples, the Lagrangian multiplier is introduced, and its function is established, the formula is as follows:

Note: a _i is the introduced Lagrangian coefficient; C is the penalty factor, a _i and a are the optimal solution vectors, and the obtained a _i and a are brought into the equation to solve the parameter regression coefficient w and the intercept item b. Finally, the regression function can be obtained.

Support vector regression mainly replaces nonlinear functions by increasing the dimension and constructing linear functions, so as to realize linear regression, which is the application of support vector machine functions in the direction of regression.

3) GBDT model (GradientBoosting Decision Tree, GBDT)

The GBDT algorithm, also known as MART (Multiple Addictive Regression Tree), is a type of integrated learning algorithm that generates a strong learner by combining many weak learners. Each calculation is to reduce the residual error of the last time, so each prediction function of GBDT must adopt a sequence, which is generated sequentially in a serial manner, and the latter model parameter needs the result of the last round of the model. This is in line with Boosting's idea of repeatedly reducing the residual to improve the effect of the model.

Its specific working principle is to input the training samples of the model {(x ₁ ,y ₁ ),(x ₂ ,y ₂ ),…,(x _n ,y _n )}, and the statistical loss function is L(y,f( x)), the number of iterations is n=1, 2, 3..., N.

Compute residuals for each sample i=1,2,...,I:

Use the residual error obtained in the previous step as the new true value of the sample, and use ( _xi , T _in ) as the training data of the next tree, so as to update the regression to obtain a new regression tree. For each leaf node of the regression tree, the range is E _kn , k=1,2,...K calculates the loss function, and obtains the optimal fitting value:

By updating the learner:

Get the final learner:

4) Neural Network (Back-Propagation Network, BP)

BP neural network (Back-Propagation Network) is a multi-layer feed-forward neural network, and its error is trained according to the reverse propagation algorithm. During the training process, the BP neural network can continuously modify the network weights and thresholds, so that the error function decreases along the opposite direction of the gradient direction, so that the output value is close to the expectation. This paper chooses MLP (Multilayer Perceptron) neural network is a common ANN (Artificial Neuro Network artificial neural network) algorithm, which consists of an input layer, an output layer and one or more hidden layers. All neurons in MLP are similar, each neuron has several input (connection to the previous layer) neuron and output (connection to the next layer) neuron, the neuron will pass the same value to the neuron connected to it multiple output neurons.

5) Multiple Adaptive Regression Splines (MARS)

The Multiple Adaptive Regression Splines (MARS) algorithm is a nonparametric multiple regression method using adaptive selection splines. Although linearity is assumed in this approach, a model can be built with coefficients that depend on the predictor variables.

S103. Input the preprocessed multispectral satellite image and the measured sample data set into the stacking integrated model for training, and obtain the overall water depth information of the water area to be measured;

The training process is as follows:

(1) input the multi-spectral satellite images and the training set data after preprocessing to a plurality of basic learners of the basic learning layer to carry out water depth inversion machine learning, and the result is used as the output of the basic learning layer;

(2) using the verification set data to verify and evaluate the output of the base learning layer;

(3) Using the output of the basic learning layer as the training data set of the generalization layer for training, and using the verification set data for verification and evaluation to obtain the overall water depth information of the water area to be tested.

The present invention uses the "SCIKIT-LEARN" package in the python language, and uses the training sample points corresponding to the selected band as a data set, which is input into four kinds of water depth inversion machine learning of random forest, support vector machine, GBDT and neural network model, which is used as the base learner output in stack generalization.

Then use the stacking module in the "SCIKIT-LEARN" package to input the first layer results as a data set into the MARS multiple adaptive regression spline algorithm in the "py-earth" package, and train it again as a meta-learner to obtain final output. The accuracy of the model is compared and compared with the verification set, and the optimized water depth inversion model is obtained.

S104. Perform quality assessment and bathymetry mapping for the overall water depth information of the water area to be measured.

Taking the Nanshan Port area of Sanya as an example, the depth estimation based on multiple datasets is better than that based on a single dataset, the RMSE value decreases and the ^R2 value increases. As shown in Figure 2, in the experiment, when using a training data set, the regression effect of the random forest algorithm is the best, while the regression result of the neural network is relatively poor. As shown in Figure 3, the ensemble results using multiple datasets outperform other individual data models. As shown in Figure 4, even though the overall trend of the secondary integrated Stacking model is very close to the random forest model, the Stacking model has better data performance than the random forest in the range of 4-9.5m, below 0.5m and 9.5 The random forest model performs better in the range above m.

Compared with the measured data, the satellite topographic map can better reflect the change trend of water depth, but there are some small errors in the details. Comparing Figure 4 with Figure 5, the water depth trends are very close. In addition, in Figure 4, there is a relatively deep water area on the northwest side of the pier, which is presumed to be a waterway, which is also reflected in the inverted Figure 5, but the depth is lower than the actual measurement. The water near the pier is relatively deep, and there are obvious traces of excavation and dredging, which are also reflected in the map.

In summary, the method combining multiple training datasets and machine learning methods is effective in estimating water depth from satellite imagery, and the method does outperform methods based on a single training dataset. Due to the nonparametric nature of the machine learning method, it can successfully retrieve water depths from observed satellite imagery with relatively high coherence and consistency compared to depths estimated by acoustic methods. In future work, the accuracy of water depth estimation can be improved by the atmospheric correction algorithm of high-resolution remote sensing images. These algorithms can be applied with ensemble learning to deal with coastal turbid water at different depths.

According to another embodiment, there is also provided a multi-spectral satellite remote sensing depth-sounding device based on the stacking integrated model, including:

Preprocessing module: used to obtain the multispectral satellite image of the water area to be measured and the corresponding measured sample data set, and perform preprocessing on the multispectral satellite image; divide the measured sample data set into a training set and a verification set;

Integrated model modeling module: set up stacking integrated model, described stacking integrated model comprises the basic learning layer of the first layer and the generalization layer of the second layer, and described basic learning layer comprises a plurality of basic learners;

Water depth calculation module: used for inputting the preprocessed multispectral satellite image and the measured sample data set into the stacking integrated model to obtain the water depth information of the water area to be measured;

Evaluation and mapping module: used for quality assessment and bathymetric mapping of the overall water depth information of the water area to be measured.

According to an embodiment of another aspect, a device is also provided, including a memory, a processor, and a computer program stored in the memory and executable on the memory. When the processor executes the computer program, the above-mentioned stacking integration model-based Multispectral satellite remote sensing bathymetry method.

According to yet another embodiment, there is also provided a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed in a computer, the computer is made to execute the above-mentioned multispectral satellite remote sensing based on the stacking integration model Sounding method.

Those skilled in the art should be aware that, in the above one or more examples, the functions described in the embodiments of this specification may be implemented by hardware, software, firmware or any combination thereof. When implemented in software, the functions may be stored on or transmitted over as one or more instructions or code on a computer-readable medium.

Although the specific implementations of the present invention have been described above, those skilled in the art should understand that these are only examples, and various changes or changes can be made to these implementations without departing from the principle and essence of the present invention. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning or limited experiments on the basis of the prior art shall be within the scope of protection determined by the claims Inside.

Claims (8)

1. a multispectral satellite remote sensing sounding method based on stacking integration model, is characterized in that, comprises the steps: (1) Obtain the multispectral satellite image of the water area to be measured and the corresponding measured sample data set, carry out preprocessing for the multispectral satellite image; divide the measured sample data set into a training set and a verification set; (2) set up the stacking integration model, the stacking integration model has two layers, including the base learning layer of the first layer and the generalization layer of the second layer, and the base learning layer includes a plurality of basic learners; (3) input the multispectral satellite image and the measured sample data set into the stacking integrated model after preprocessing to obtain the overall water depth information of the water area to be measured; (4) Carry out quality assessment and bathymetry mapping for the overall water depth information of the water area to be measured. 2. The multispectral satellite remote sensing depth sounding method based on the stacking integrated model according to claim 1, wherein the preprocessing for the multispectral satellite image comprises: radiation correction, geometric registration and flare elimination. 3. the multispectral satellite remote sensing sounding method based on stacking integrated model according to claim 1, is characterized in that, a plurality of base learners of described basic learning layer comprise: random forest, support vector machine, GBDT and neural network Four machine learning models for bathymetry inversion. 4. the multispectral satellite remote sensing sounding method based on stacking integrated model according to claim 1, is characterized in that, the learner of described generalization layer is MARS multiple adaptive regression spline algorithm. 5. the multispectral satellite remote sensing sounding method based on stacking integrated model according to claim 1, is characterized in that, the training method of described stacking integrated model comprises: (2.1) Input the multi-spectral satellite images and the training set data after preprocessing into a plurality of basic learners of the basic learning layer to perform water depth inversion machine learning, and the result is used as the output of the basic learning layer; (2.2) using the verification set data to verify and evaluate the output of the base learning layer; (2.3) Using the output of the basic learning layer as the training data set of the generalization layer for training, and using the verification set data for verification and evaluation, and obtaining the overall water depth information of the water area to be tested. 6. A multispectral satellite remote sensing depth-sounding device based on the stacking integrated model, characterized in that, it is applied to the multispectral satellite remote sensing depth-sounding method based on the stacking integrated model described in any one of claims 1-5, comprising: Preprocessing module: used to obtain the multispectral satellite image of the water area to be measured and the corresponding measured sample data set, and perform preprocessing on the multispectral satellite image; divide the measured sample data set into a training set and a verification set; Integrated model modeling module: for setting up stacking integrated model, described stacking integrated model comprises the basic learning layer of the first layer and the generalization layer of the second layer, and described basic learning layer comprises a plurality of basic learners; Water depth calculation module: used for inputting the preprocessed multispectral satellite image and the measured sample data set into the stacking integrated model to obtain the water depth information of the water area to be measured; Assessment and mapping module: used for quality assessment and bathymetric mapping of the overall water depth information of the water area to be measured. 7. A device, comprising a memory, a processor, and a computer program stored in the memory and capable of being executed on the memory, characterized in that: when the processor executes the computer program, the computer program according to any one of claims 1-5 is implemented. The multispectral satellite remote sensing bathymetry method based on the stacking integration model described above. 8. A computer-readable storage medium, on which a computer program is stored, characterized in that: when the computer program is executed in a computer, the computer is made to perform the stacking-based integration described in any one of claims 1-5. Modeling methods for multispectral satellite remote sensing bathymetry.

Images