Water depth inversion method, medium and system based on satellite remote sensing image data

Through the embedded neural network model and fine-tuning mechanism, the optical characteristic parameters of the water body are extracted, and the problem that the water depth inversion accuracy is affected by the variation of the water body's optical characteristic is solved, and efficient and robust water depth distribution map generation is achieved to meet the practical application needs.

CN119007034AActive Publication Date: 2024-11-22QINGDAO GUOCEN HAIYAO INFORMATION TECH CO LTD

Patent Information

Application Number
CN202411479963.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-10-23
Publication Date
2024-11-22
Estimated Expiration
2044-10-23

AI Technical Summary

Technical Problem

The accuracy of the existing water depth inversion method based on satellite remote sensing images is greatly affected by the variation of the optical characteristics of the water body and is easily disturbed by various factors, resulting in unstable accuracy.

Method used

The embedded neural network model is adopted to extract the optical characteristic parameters of the water body through the multi-spectral satellite remote sensing image processed after atmospheric correction, and a fine-tuning training set is established based on the field sampling point data, and the pre-trained water depth inversion model is fine-tuned to generate a continuous water depth distribution. Graphs and ensure accuracy through spatial interpolation and smoothing processing.

Benefits of technology

It improves the robustness and adaptability of water depth inversion, significantly reduces application costs, and the generated water depth distribution map is highly consistent with the actual situation, meeting the continuity and coverage requirements of water depth data in the sea area.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119007034A_ABST
    Figure CN119007034A_ABST
Patent Text Reader

Abstract

The invention provides a water depth inversion method based on satellite remote sensing image data, a medium and a system, and belongs to the technical field of remote sensing image water depth measurement, and the method comprises the steps: firstly carrying out the atmospheric correction of a satellite image, dividing a sea area into grid units, extracting the water optical characteristic parameters of each unit, and calculating the water optical characteristic parameters of each unit; and carrying out classification by utilizing a pre-trained water body classification model. And then based on the water depth data of a part of on-site sampling points and the corresponding satellite image and water body classification, establishing a fine tuning training set, and performing fine tuning on the pre-trained satellite remote sensing water depth inversion model to obtain a water depth inversion model suitable for the sea area. And inputting the satellite image and the water body of each grid unit into a water depth inversion model in a classified manner to obtain water depth values of the grid units, and generating a continuous water depth distribution diagram of the whole sea area through spatial interpolation and smoothing processing. The method solves the technical problems that the precision of remote sensing image inversion is greatly influenced by the variation of the optical characteristics of the water body, is easily interfered by various factors, and is poor in precision.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of water depth measurement using remote sensing images, and in particular, relates to a water depth inversion method, medium and system based on satellite remote sensing image data. Background Art

[0002] Offshore water depth information is one of the key parameters in the fields of water resources management, mariculture, hydrological monitoring, etc. Traditional water depth measurement methods mainly rely on field measurements, usually using equipment such as depth sounders or sonars for single-point or cross-sectional measurements. This method has high measurement accuracy, but limited coverage, low efficiency, and requires a lot of field work, which poses safety risks. With the development of remote sensing technology, using satellite remote sensing image data to invert water depth has become a more efficient, widespread and safer measurement method.

[0003] At present, there are several common methods for water depth inversion based on satellite remote sensing images: 1. Empirical regression model method: Establish an empirical regression relationship between water depth and image band reflectivity or band ratio and other parameters, and obtain the water depth inversion model through fitting. This method is simple and easy, but it has poor applicability to different sea areas and requires a large amount of measured data for empirical correction.

[0004] 2. Semi-analytical model method: Based on the theory of water body optics, an analytical relationship between water depth and water body optical characteristic parameters such as reflectivity, absorption coefficient, scattering coefficient, etc. is established, and these water body optical parameters are inverted through remote sensing images to estimate the water depth. This method has a good theoretical basis, but requires accurate water body optical parameter input, is easily interfered by various factors, and has poor accuracy.

[0005] 3. Data-driven model method: Using machine learning or deep learning methods, the water depth inversion model is learned from a large amount of remote sensing image data and measured water depth data, and the water depth estimation result is directly output. This method has good applicability and generalization ability, but it requires a large amount of high-quality training data support and it is difficult to explain the internal mechanism of the model.

[0006] 4. Some other methods require more data sources or more modal remote sensing data. For example, the Chinese invention patent with publication number CN105651263B discloses a shallow sea water depth multi-source remote sensing fusion inversion method, which comprehensively utilizes the different responses of multiple remote sensing data sources to water depth information, expands the use range of remote sensing image spectral data bands and spectral information, mines the water depth data, and improves the inversion accuracy. This patent requires the expansion of the spectral data bands of remote sensing images, which has high requirements for remote sensing equipment.

[0007] In summary, although the existing water depth inversion methods can achieve good results in some cases, there are still the following major problems: the inversion accuracy based on remote sensing images is greatly affected by the variation of the optical properties of the water body, and is easily disturbed by various factors and has poor accuracy.

[0008] Therefore, there is an urgent need for a new water depth inversion method based on satellite remote sensing images, which can more robustly extract water optical characteristic parameters, have strong generalization ability and adaptability, and can generate a continuous water depth distribution map of the entire sea area to meet the needs of practical applications. Summary of the invention

[0009] In view of this, the present invention provides a water depth inversion method, medium and system based on satellite remote sensing image data, which can solve the technical problem that the accuracy of remote sensing image inversion is greatly affected by the variation of water body optical properties, is easily disturbed by various factors and has poor accuracy.

[0010] The present invention is achieved in that: A first aspect of the present invention provides a water depth inversion method based on satellite remote sensing image data, which comprises the following steps: S10, obtaining a multispectral satellite remote sensing image of the sea area to be measured and actual water depth values ​​of multiple field sampling points of the sea area to be measured, and performing atmospheric correction processing on the multispectral satellite remote sensing image; S20, dividing the sea area to be measured into grids to obtain a plurality of unit grid sea areas; S30, using spectral analysis methods, extracting water optical characteristic parameters of each unit grid sea area from the multi-spectral satellite remote sensing image after atmospheric correction, including water reflectivity, absorption coefficient, scattering coefficient and seaweed concentration; S40, inputting the water body optical characteristic parameters into a pre-trained water body optical classification model to obtain the water body classification of each unit grid sea area; S50, establishing a fine-tuning training set according to the actual water depth values ​​of the multiple field sampling points and part of the multispectral satellite remote sensing images and water body classification of the corresponding unit grid sea area; S60, using the fine-tuning training set to fine-tune the pre-trained satellite remote sensing image water depth inversion model to obtain a water depth inversion model suitable for the sea area to be measured; S70, inputting part of the multispectral satellite remote sensing images and water body classification data corresponding to each unit grid sea area into the water depth inversion model to obtain the water depth value of each unit grid sea area; S80, performing spatial interpolation and smoothing processing on the water depth value of each unit grid sea area, and generating a continuous water depth distribution map of the entire sea area to be measured; S90, using the actual water depth values ​​of the field sampling points to evaluate and verify the accuracy of the generated water depth distribution map. If the accuracy does not meet the requirements, repeat S60 to S90 until the accuracy meets the requirements.

[0011] On the basis of the above technical solution, the water depth inversion method based on satellite remote sensing image data of the present invention can also be improved as follows: Among them, the accuracy meets the requirements, specifically referring to: the root mean square error (RMSE) between the generated water depth distribution map and the actual water depth value of the field sampling point is less than a preset threshold, and the relative error is within the allowable range; the preset threshold is set based on experience.

[0012] The preset threshold value is 0.0375~0.05 by default.

[0013] The water optical classification model adopts an embedded neural network, including a feature extraction module, a water optical property equation module, an error correction lightweight neural network group, a perception network and a fusion network.

[0014] Furthermore, the feature extraction module is used to extract the spectral features of water bodies from multi-spectral satellite remote sensing images, the input is the multi-band reflectance data after atmospheric correction, the output is the spectral feature vector group of water bodies, and the structure is a multi-layer convolutional neural network; The water body optical characteristic equation module includes a water body reflectivity calculation equation, an absorption coefficient calculation equation, a scattering coefficient calculation equation, and a seaweed concentration calculation equation; The error correction lightweight neural network group includes a water body reflectivity calculation equation error correction network, an absorption coefficient calculation equation error correction network, a scattering coefficient calculation equation error correction network and a seaweed concentration calculation equation error correction network; The perception network is used to extract specific optical properties of water bodies, the input is a spectral feature vector group of water bodies, the output is specific optical property parameters, and the structure is a multi-layer perceptron; The fusion network is used to integrate the output results of each branch network. The input is the output features of each branch network, and the output is the final water body classification result. The structure is a fully connected neural network.

[0015] Furthermore, the water body reflectance calculation equation is used to calculate the reflectance of the water body, the input is the spectral feature vector group of the water body, and the output is the estimated water body reflectance; The absorption coefficient calculation equation is used to calculate the absorption coefficient of the water body, the input is the spectral feature vector group of the water body, and the output is the estimated absorption coefficient; The scattering coefficient calculation equation is used to calculate the scattering coefficient of the water body, the input is the spectral feature vector group of the water body, and the output is the estimated scattering coefficient; The seaweed concentration calculation equation is used to calculate the seaweed concentration of a water body, the input is a spectral feature vector group of the water body, and the output is the estimated seaweed concentration.

[0016] Furthermore, the water body reflectance calculation equation error correction network is used to correct the reflectance calculation error, the input is the spectral feature vector group and the calculation result of the water body reflectance calculation equation, the output is the corrected reflectance, and the structure is MobileNet; The absorption coefficient calculation equation error correction network is used to correct the error of absorption coefficient calculation, the input is the spectral feature vector group and the calculation result of the absorption coefficient calculation equation, the output is the corrected absorption coefficient, and the structure is MobileNet; The scattering coefficient calculation equation error correction network is used to correct the error of the scattering coefficient calculation, the input is the spectral feature vector group and the calculation result of the scattering coefficient calculation equation, the output is the corrected scattering coefficient, and the structure is MobileNet; The seaweed concentration calculation equation error correction network is used to correct the error of seaweed concentration calculation, the input is the spectral feature vector group and the calculation result of the seaweed concentration calculation equation, the output is the corrected seaweed concentration, and the structure is MobileNet.

[0017] The seaweed concentration calculation equation and the seaweed concentration calculation equation error correction network do not directly calculate the seaweed concentration, but are obtained through approximate calculation of spectral feature vectors. The seaweed concentration calculation result is a parameter with a relatively small influence on water depth measurement.

[0018] Furthermore, the steps of establishing a training data set for the water body optical classification model specifically include: collecting measured optical property data of various water body types and corresponding multispectral satellite remote sensing images, performing data preprocessing and labeling, and constructing a training sample set containing different water body types.

[0019] The steps of training the water body optical classification model specifically include: initializing model parameters, using a training sample set for forward propagation and back propagation, optimizing model parameters, and iterating training until the model converges.

[0020] Furthermore, the satellite remote sensing image water depth inversion model adopts a deep learning method, and its specific structure is: a multi-layer convolutional neural network is used for feature extraction, a fully connected layer is used for water depth estimation, and a residual connection is used to improve model performance; the steps for establishing a training data set for the satellite remote sensing image water depth inversion model specifically include: collecting measured water depth data of multiple sea areas and corresponding multispectral satellite remote sensing images, performing data preprocessing and alignment, and constructing a training sample set containing different water depth ranges.

[0021] The steps of training the satellite remote sensing image water depth inversion model specifically include: initializing model parameters, using a training sample set for forward propagation and back propagation, optimizing model parameters, using a validation set to evaluate model performance, and iterating training until the model converges.

[0022] The steps of fine-tuning the satellite remote sensing image water depth inversion model specifically include: freezing the first few layers of parameters of the pre-trained model and only fine-tuning the parameters of the last few layers; using the fine-tuning training set of the sea area to be tested for a small amount of iterative training; adjusting the learning rate and optimizer parameters; using the early stopping strategy to prevent overfitting. The default first few layers are the first 2 layers, and the default back base is the last 3 layers, which can be adjusted according to the actual fine-tuning effect.

[0023] Furthermore, the spectral feature vector group Contains the following feature vectors: 1. : The reflectance value vector of each band, with dimension ,in is the number of spectral bands; 2. : Adjacent band reflectivity ratio vector, dimension is ; 3. : reflectivity first-order derivative vector, dimension is ; 4. : reflectivity second-order derivative vector, dimension is ; 5. : The trough position feature vector, dimension is ,in is the number of main absorption valleys; 6. : Peak position feature vector, dimension is ,in is the number of main reflection peaks.

[0024] The water body reflectivity calculation equation is specifically expressed as follows: ; In the formula, The wavelength to be estimated The reflectivity of water at is the weight coefficient to be determined; is the error term, which follows a normal distribution .

[0025] The water absorption coefficient calculation equation is specifically expressed as follows: ; In the formula, The wavelength to be estimated The absorption coefficient at ; is the coefficient to be determined; For the The central wavelength of each band; For the The location of the absorption valley; is the standard deviation of the Gaussian function; is the error term, which follows a normal distribution .

[0026] The water body scattering coefficient calculation equation is specifically expressed as follows: ; In the formula, The wavelength to be estimated The scattering coefficient at ; is the coefficient to be determined; is the reference wavelength; is the scattering index; For the The position of the reflection peak; is the standard deviation of the Gaussian function; is the error term, which follows a normal distribution .

[0027] The seaweed concentration calculation equation is specifically expressed as follows: ; In the formula, for the algae concentration; is the coefficient to be determined; is the band combination weight coefficient; is the error term, which follows a normal distribution .

[0028] The parameter acquisition method is as follows: 1. The steps to obtain the weight coefficient are: Step 1: Build a training dataset ,in is the spectral feature vector group, is the corresponding true value; Step 2: Define the loss function ,in is the corresponding equation, is the regularization term; Step 3: Optimize using gradient descent: ,in is the learning rate.

[0029] 2. The steps to obtain the standard deviation parameters are: Step 1: Calculate the Gaussian kernel function response: ; Step 2: Maximize the likelihood function: ; Step 3: Use cross validation to determine the optimal value.

[0030] 3. The steps to obtain the scattering index are: Step 1: Calculate wavelength dependence: ; Step 2: Estimate the power relationship: ; Step 3: Optimization by least squares fitting.

[0031] 4. The steps to obtain the band combination weight are: Step 1: Calculate the band correlation coefficient matrix: ; Step 2: Apply principal component analysis to reduce dimensionality: ,in is the principal component matrix; Step 3: Determine weights based on contribution rate: ,in is the characteristic value.

[0032] A second aspect of the present invention provides a computer-readable storage medium, in which program instructions are stored. When the program instructions are run in a computer, they are used to execute the above-mentioned water depth inversion method based on satellite remote sensing image data.

[0033] A third aspect of the present invention provides a water depth inversion system based on satellite remote sensing image data, which includes the above-mentioned computer-readable storage medium.

[0034] Compared with the prior art, the water depth inversion method, medium and system based on satellite remote sensing image data provided by the present invention have the following beneficial effects: 1. The advanced embedded neural network model can robustly extract the optical characteristic parameters of water bodies from remote sensing images, including reflectivity, absorption coefficient, scattering coefficient and seaweed concentration. The model has multiple built-in submodules, which are responsible for feature extraction, parameter calculation, error correction and other functions, which can effectively reduce the impact of water optical characteristic variation on inversion accuracy.

[0035] 2. A targeted model fine-tuning mechanism is introduced to optimize the pre-trained water depth inversion model using a small amount of measured water depth data, so that it can better adapt to the characteristics of the current sea area to be measured, thereby greatly improving the generalization ability of the model. This fine-tuning method based on transfer learning can significantly reduce application costs and improve inversion effects.

[0036] 3. After obtaining the estimated water depth of each grid cell, this method uses spatial interpolation and smoothing to generate a continuous water depth distribution map of the entire sea area to be measured. This method can not only reflect the trend of water depth changes within the sea area, but also better meet the requirements for continuity and coverage of water depth data in practical applications.

[0037] 4. In the accuracy evaluation and verification phase, this method comprehensively considers the two indicators of RMSE and relative error, sets a reasonable accuracy threshold, and ensures that the final generated water depth distribution map is consistent with the actual situation through iterative optimization. This mechanism not only ensures the reliability of the inversion results, but also helps to achieve a good match with actual application needs.

[0038] In general, the water depth inversion method based on satellite remote sensing images proposed in the present invention fully utilizes the advantages of machine learning and remote sensing technology, and achieves innovative breakthroughs in key links such as extracting water body optical properties, model generalization, continuous water depth distribution generation and precision control, effectively solving the main problems existing in the prior art, greatly improving the robustness and adaptability of water depth inversion, and solving the technical problem that the inversion accuracy relying on remote sensing images is greatly affected by the variation of water body optical properties, is easily disturbed by various factors and has poor accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 A flow chart of the method provided by the present invention; Figure 2 This is a comparison chart of spectral characteristics of different sea area types; Figure 3 It is a three-dimensional relationship diagram of water depth, seaweed concentration and reflectivity; Figure 4 This is a distribution map of water depth contours in offshore waters. DETAILED DESCRIPTION

[0040] In order to make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0041] like Figure 1 FIG. 1 is a flow chart of a method for water depth inversion based on satellite remote sensing image data provided by the present invention. The method comprises the following steps: S10, obtaining a multispectral satellite remote sensing image of the sea area to be measured and actual water depth values ​​of multiple field sampling points in the sea area to be measured, and performing atmospheric correction processing on the multispectral satellite remote sensing image; S20, dividing the sea area to be measured into grids to obtain a plurality of unit grid sea areas; S30, using spectral analysis methods, extracting water optical characteristic parameters of each unit grid sea area from the multi-spectral satellite remote sensing image after atmospheric correction, including water reflectivity, absorption coefficient, scattering coefficient and seaweed concentration; S40, inputting the optical characteristic parameters of the water body into a pre-trained water body optical classification model to obtain the water body classification of each unit grid sea area; S50, establishing a fine-tuning training set according to the actual water depth values ​​of multiple field sampling points and part of the multispectral satellite remote sensing images and water body classification of the corresponding unit grid sea area; S60, using a fine-tuning training set, fine-tuning a pre-trained satellite remote sensing image water depth inversion model to obtain a water depth inversion model suitable for the sea area to be measured; S70, inputting part of the multispectral satellite remote sensing images and water body classification data corresponding to each unit grid sea area into a water depth inversion model to obtain a water depth value of each unit grid sea area; S80, performing spatial interpolation and smoothing processing on the water depth value of each unit grid sea area, and generating a continuous water depth distribution map of the entire sea area to be measured; S90, using the actual water depth values ​​of the field sampling points to evaluate and verify the accuracy of the generated water depth distribution map. If the accuracy does not meet the requirements, repeat S60 to S90 until the accuracy meets the requirements.

[0042] The specific implementation methods of the above steps are described in detail below: The specific implementation of step S10 is as follows: first, obtain the multi-spectral satellite remote sensing image data of the sea area to be measured. While obtaining the multi-spectral satellite image data, the actual water depth values ​​of multiple field sampling points of the sea area to be measured should also be obtained. Since the satellite image data will be affected by atmospheric scattering and absorption during the transmission to the ground, it is necessary to perform atmospheric correction processing on the acquired multi-spectral satellite remote sensing image data. The main purpose of atmospheric correction is to remove the influence of the atmosphere on the image so that the image reflectivity value reflects the optical properties of the water body more realistically. The atmospheric correction processing can adopt methods such as the FLAASH model based on the radiation transfer theory or the QUAC model based on the dark target correction. Through the atmospheric correction processing, the multi-spectral satellite remote sensing image data reflecting the real optical properties of the sea area to be measured can be obtained.

[0043] The specific implementation of step S20 is as follows: the multi-spectral satellite remote sensing image data of the sea area to be measured is spatially gridded. Specifically, the entire sea area to be measured is divided into a plurality of unit grid sea areas. The purpose of this is to extract and analyze the optical characteristic parameters of each unit grid sea area independently in the future. The size of the grid can be appropriately adjusted according to the actual situation, for example, it can be set to This spatial grid division method can better reflect the differences in water depth distribution in different areas of the sea area to be measured.

[0044] The specific implementation of step S30 is as follows: using spectral analysis methods, extract the water optical characteristic parameters of each unit grid sea area from the multi-spectral satellite remote sensing image after atmospheric correction. Specifically, the following four water optical characteristic parameters need to be extracted: 1. Water reflectivity : Indicates that water is at a certain wavelength The reflectivity characteristic at . It is given by the formula Calculated, where are different spectral feature vectors, is the undetermined weight coefficient, is the error term.

[0045] 2. Absorption coefficient : Indicates that water is at a certain wavelength The absorption characteristics at the point. It is expressed by the formula Calculated, where are different spectral feature vectors, is the coefficient to be determined, is the wavelength parameter, is the standard deviation of the Gaussian function, is the error term.

[0046] 3. Scattering coefficient : Indicates that water is at a certain wavelength The scattering characteristics at the location are given by the formula Calculated, where are different spectral feature vectors, is the coefficient to be determined, is the reference wavelength, is the scattering index, is the reflection peak position, is the standard deviation of the Gaussian function, is the error term.

[0047] 4. Seaweed concentration : represents the concentration of algae in water. It is expressed by the formula Calculated, where are different spectral feature vectors, is the coefficient to be determined, is the band combination weight, is the error term.

[0048] The calculation process of the above four water optical characteristic parameters involves a series of undetermined parameters, which can be optimized and solved by gradient descent method, maximum likelihood estimation, principal component analysis and other algorithms. After obtaining these four water optical characteristic parameters of each unit grid sea area, we can proceed to the next step of water body classification.

[0049] The specific implementation of step S40 is as follows: using the water optical characteristic parameters extracted above as input, inputting them into the pre-trained water optical classification model, thereby obtaining the water classification results of each unit grid sea area. The water optical classification model used here is a model based on an embedded neural network, which includes the following five main modules: 1. Feature extraction module: It uses a multi-layer convolutional neural network structure to extract the spectral feature vector group of water bodies from multispectral satellite remote sensing images. .

[0050] 2. Water optical characteristics equation module: contains 4 branches, which are used to calculate the reflectivity of water. , absorption coefficient , scattering coefficient and algae concentration These equations contain some unknown parameters that need to be solved by optimization algorithms.

[0051] 3. Error correction lightweight neural network group: It contains 4 sub-networks, which are used to correct the systematic errors in the calculation of water reflectivity, absorption coefficient, scattering coefficient and seaweed concentration, and adopts the MobileNet structure.

[0052] 4. Perception network: A multi-layer perceptron structure is used to extract specific optical characteristic parameters of water bodies.

[0053] 5. Fusion network: A fully connected neural network structure is used to integrate the output features of each branch network to give the final water body classification result.

[0054] The process of establishing the training data set for the water optical classification model includes: collecting the measured optical property data of various water types and the corresponding multispectral satellite remote sensing images, preprocessing and labeling the data, and constructing a training sample set containing different water types. The training process includes: initializing the model parameters, using the training sample set for forward propagation and back propagation, optimizing the model parameters, and iterating the training until the model converges.

[0055] The specific implementation method of step S50 is as follows: using the actual water depth values ​​of the multiple field sampling points obtained above, as well as the partial multispectral satellite remote sensing images and water body classification results of the corresponding unit grid sea area, a fine-tuning training set is constructed. This fine-tuning training set will be used to fine-tune and optimize the pre-trained satellite remote sensing image water depth inversion model in the future, so that it can better adapt to the characteristics of the current sea area to be tested. The process of establishing the fine-tuning training set includes: firstly matching the location information of the field sampling points with the corresponding unit grid sea area, then extracting partial multispectral satellite remote sensing image data and water body classification results of these unit grid sea areas, and finally combining these data with the actual water depth values ​​of the field sampling points to form a fine-tuning training set.

[0056] The specific implementation of step S60 is as follows: using the fine-tuning training set constructed above, fine-tuning and optimizing the pre-trained satellite remote sensing image water depth inversion model, thereby obtaining a water depth inversion model suitable for the current sea area to be measured. The satellite remote sensing image water depth inversion model used here adopts a deep learning method, which includes the following main modules: 1. Multi-layer convolutional neural network: used to extract water features from multispectral satellite remote sensing images.

[0057] 2. Fully connected layer: used to estimate water depth based on the extracted features.

[0058] 3. Residual connection: used to improve the performance of the model.

[0059] The process of constructing the training data set for the water depth inversion model includes: collecting measured water depth data of multiple sea areas and corresponding multispectral satellite remote sensing images, performing data preprocessing and registration, and constructing a training sample set containing different water depth ranges. The training process includes: initializing model parameters, using the training sample set for forward propagation and back propagation, optimizing model parameters, using the validation set to evaluate model performance, and iterating training until the model converges.

[0060] When fine-tuning and optimizing the pre-trained model, the specific steps are as follows: first, freeze the first two layers of the pre-trained model parameters, and only fine-tune the last three layers of parameters; then use the previously constructed fine-tuning training set for a small number of iterative training; adjust the learning rate and optimizer parameters at the same time; and finally use the early stopping strategy to prevent overfitting. Through this fine-tuning method, the feature extraction capability of the pre-trained model can be fully utilized, while customized optimization can be performed according to the characteristics of the current sea area to be tested, thereby obtaining a more accurate water depth inversion model.

[0061] The specific implementation of step S70 is as follows: the water depth inversion model obtained by the above optimization is applied to each unit grid sea area, and the partial multispectral satellite remote sensing image data and water body classification results corresponding to the unit grid sea area are input to obtain the water depth value of the unit grid sea area. The multispectral satellite remote sensing image data input here has undergone the above atmospheric correction processing, and the water body classification results come from step S40. By inputting all the unit grid sea areas of the entire sea area to be tested into the water depth inversion model one by one, the water depth estimation value of each unit grid sea area can be obtained.

[0062] The specific implementation of step S80 is as follows: perform spatial interpolation and smoothing on the estimated water depth of each unit grid sea area obtained in the previous step, thereby generating a continuous water depth distribution map of the entire sea area to be measured. The specific operations are as follows: 1. Spatial interpolation: Using the Kriging interpolation algorithm, spatial interpolation is performed in the entire sea area to be measured based on the known unit grid water depth values ​​to obtain a continuous water depth distribution.

[0063] 2. Smoothing: Gaussian filtering and smoothing the interpolated water depth distribution map can effectively suppress the problem of discontinuous water depth distribution caused by data discreteness.

[0064] Through the above-mentioned spatial interpolation and smoothing processing, a continuous water depth distribution map of the entire sea area to be measured can be obtained, providing necessary data support for subsequent accuracy evaluation and verification.

[0065] The specific implementation of step S90 is as follows: using the actual water depth values ​​of the field sampling points obtained previously, the generated water depth distribution map is evaluated and verified for accuracy. The specific operations are as follows: 1. Calculate the root mean square error (RMSE) between the generated water depth distribution map and the actual water depth values ​​at the field sampling points.

[0066] 2. Calculate the relative error between the generated water depth distribution map and the actual water depth value at the field sampling point.

[0067] 3. Compare the calculated RMSE and relative error with the preset accuracy threshold.

[0068] The preset accuracy threshold is set based on experience. Generally, RMSE is less than 0.0375~0.05, and the relative error can meet the requirements within the allowable range.

[0069] If the current accuracy does not meet the requirements, it is necessary to repeat steps S60 to S90 to further fine-tune and optimize the water depth inversion model until the accuracy meets the expectations. Through this iterative optimization method, it can be ensured that the final generated water depth distribution map is highly consistent with the actual situation.

[0070] In general, the water depth inversion method based on satellite remote sensing image data proposed in the present invention mainly includes the following key steps: 1. Obtain multispectral satellite remote sensing images and field sampling water depth data of the sea area to be measured, and perform atmospheric correction processing.

[0071] 2. Divide the sea area into grids and extract the water optical characteristic parameters of each unit grid sea area.

[0072] 3. Use the pre-trained water optical classification model to classify the unit grid sea area.

[0073] 4. Construct a fine-tuning training set to perform targeted optimization on the pre-trained depth inversion model.

[0074] 5. Apply the optimized water depth inversion model to estimate the water depth value of each unit grid sea area.

[0075] 6. Perform spatial interpolation and smoothing on the unit grid water depth values ​​to generate a continuous water depth distribution map.

[0076] 7. Use the water depth data from field sampling points to evaluate and verify the accuracy of the generated water depth distribution map, and perform iterative optimization when necessary.

[0077] The entire method makes full use of the spatial coverage advantage of satellite remote sensing data and combines it with field measured data to fine-tune the model, thereby obtaining more accurate water depth distribution information and providing valuable data support for applications such as sea area monitoring and mariculture.

[0078] A second aspect of the present invention provides a computer-readable storage medium, in which program instructions are stored. When the program instructions are run in a computer, they are used to execute the above-mentioned water depth inversion method based on satellite remote sensing image data.

[0079] A third aspect of the present invention provides a water depth inversion system based on satellite remote sensing image data, which includes the above-mentioned computer-readable storage medium.

[0080] Specifically, the principle of the present invention is to make full use of the rich information contained in remote sensing image data, combine it with the on-site measured water depth data, and realize high-precision inversion of water depth through advanced machine learning models.

[0081] First, the present invention uses atmospheric correction technology to remove atmospheric interference on the signal from the original satellite remote sensing image and extracts reflectivity data reflecting the real optical characteristics of the water body. Then, the entire sea area to be measured is divided into multiple grid units, and four key water body optical characteristic parameters are extracted for each grid unit sea area: reflectivity, absorption coefficient, scattering coefficient and seaweed concentration. These water body optical characteristic parameters are the key inputs for water depth inversion and can fully describe the optical characteristics of the water body.

[0082] In order to robustly extract these water optical parameters, the present invention adopts a water optical classification model based on embedded neural network. The model includes multiple functional modules such as feature extraction, parameter calculation, and error correction. It can effectively extract the optical characteristics of water bodies from remote sensing images and calculate the required optical parameters using built-in physical models. At the same time, the model also has built-in lightweight neural network components, which can perform adaptive error correction on the calculation results, enhancing the adaptability to the variation of water optical characteristics.

[0083] After obtaining the water optical parameters of each grid unit, the present invention further uses a small amount of measured water depth data to perform targeted fine-tuning and optimization on the pre-trained water depth inversion model. The water depth inversion model adopts a deep learning method, including a multi-layer convolutional neural network for feature extraction, a fully connected layer for water depth estimation, and a residual connection for performance improvement. Through this fine-tuning mechanism based on transfer learning, the advantages of the pre-trained model can be fully utilized, and at the same time, it can be customized and optimized according to the characteristics of the current sea area to be tested, greatly improving the generalization ability of the model.

[0084] Finally, the present invention uses spatial interpolation and smoothing to convert the estimated water depth of each grid cell into a continuous water depth distribution map of the entire sea area. This can not only better reflect the trend of water depth changes within the sea area, but also meet the requirements for continuity and coverage of water depth data in practical applications. In the accuracy evaluation and verification stage, this method comprehensively considers the two indicators of RMSE and relative error, and ensures that the final result is consistent with the actual situation through iterative optimization.

[0085] In general, the present invention makes full use of the rich information in remote sensing images, combines advanced machine learning technology, and achieves innovation in key links such as water body optical parameter extraction, model fine-tuning optimization, and continuous water depth distribution generation, solving the main problems existing in the prior art. This method conforms to the physical mechanism and application requirements of water depth inversion and has strong adaptability and robustness.

[0086] A specific embodiment 1 of the method provided by the first aspect of the present invention is provided below. The specific implementation of each step in this embodiment 1 is described in detail as follows: The specific implementation of step S10 is as follows: First, obtain multispectral remote sensing image data of the sea area to be measured ,in represents wavelength, Indicates the two-dimensional spatial coordinates. At the same time, the actual water depth values ​​of multiple field sampling points in the sea area to be measured are obtained. ,in is the number of sampling points. Since remote sensing image data will be affected by atmospheric scattering and absorption during transmission, it is necessary to perform atmospheric correction on the acquired multispectral remote sensing image data. The purpose of atmospheric correction is to remove the influence of the atmosphere on the image so that the image reflectance value It can more realistically reflect the optical properties of water bodies. Atmospheric correction can be done by using the FLAASH model based on radiation transfer theory or the QUAC model based on dark target correction. Specifically, the FLAASH model uses the following formula for atmospheric correction: ; in, is the radiance detected by the sensor, is the atmospheric radial radiance, is the surface diffuse radiance, is the atmospheric transmittance. Through atmospheric correction processing, multispectral remote sensing image data reflecting the real optical characteristics of the sea area to be measured can be obtained. .

[0087] The specific implementation of step S20 is as follows: the multispectral remote sensing image data of the sea area to be measured is obtained Specifically, the entire sea area to be tested is divided into grid unit sea area, the size of each grid unit is set to The purpose of this is to extract and analyze the optical characteristic parameters of each grid unit sea area independently. Appropriate adjustments can be made according to actual conditions, for example, This spatial grid division method can better reflect the differences in water depth distribution in different areas of the sea area to be measured.

[0088] The specific implementation of step S30 is as follows: using a spectral analysis method to obtain the multi-spectral remote sensing image after atmospheric correction Extract the four water optical characteristic parameters of each grid unit sea area: 1. Water reflectivity : ; in, represents the reflectivity value vector of each band, represents the reflectivity ratio vector of adjacent bands, represents the trough position feature vector, represents the peak position feature vector, is the undetermined weight coefficient, is the error term, which follows a normal distribution .

[0089] 2. Absorption coefficient : ; in, represents the first-order derivative vector of reflectivity, For the The central wavelength of the band, For the The location of the absorption valley, is the standard deviation of the Gaussian function, is the coefficient to be determined, is the error term, which follows a normal distribution .

[0090] 3. Scattering coefficient : ; in, represents the reflectivity second-order derivative vector, is the reference wavelength, is the scattering index, For the The position of the reflection peak, is the standard deviation of the Gaussian function, is the coefficient to be determined, is the error term, which follows a normal distribution .

[0091] 4. Seaweed concentration : ; in, is the band combination weight coefficient, is the coefficient to be determined, is the error term, which follows a normal distribution .

[0092] The calculation process of the above four water optical characteristic parameters involves a series of undetermined parameters, which can be optimized and solved by the following methods: 1. Obtaining weight coefficient: Step 1: Build a training dataset ,in is the spectral feature vector group, is the corresponding true value; Step 2: Define the loss function ,in is the corresponding equation, is the regularization term; Step 3: Optimize using gradient descent: ,in is the learning rate.

[0093] 2. Obtaining standard deviation parameters: Step 1: Calculate the Gaussian kernel function response: ; Step 2: Maximize the likelihood function: ; Step 3: Use cross validation to determine the optimal value.

[0094] 3. Scattering index acquisition: Step 1: Calculate wavelength dependence: ; Step 2: Estimate the power relationship: ; Step 3: Optimization by least squares fitting.

[0095] 4. Band combination weight acquisition: Step 1: Calculate the band correlation coefficient matrix: ; Step 2: Apply principal component analysis to reduce dimensionality: ,in is the principal component matrix; Step 3: Determine weights based on contribution rate: ,in is the characteristic value.

[0096] Through the above method, the four water optical characteristic parameters of each grid unit sea area can be obtained, providing necessary input for subsequent water body classification.

[0097] The specific implementation of step S40 is as follows: using the four water optical characteristic parameters extracted above as input, inputting them into a pre-trained water optical classification model based on an embedded neural network, thereby obtaining the water classification results of each grid unit sea area. The water optical classification model includes the following five main modules: 1. Feature extraction module: It uses a multi-layer convolutional neural network structure to extract the spectral feature vector group of water bodies from multispectral remote sensing images. .

[0098] 2. Water optical characteristics equation module: contains 4 branches, which are used to calculate the reflectivity of water. , absorption coefficient , scattering coefficient and algae concentration These equations contain some unknown parameters that need to be solved by optimization algorithms.

[0099] 3. Error correction lightweight neural network group: It contains 4 sub-networks, which are used to correct the systematic errors in the calculation of water reflectivity, absorption coefficient, scattering coefficient and seaweed concentration, and adopts the MobileNet structure.

[0100] 4. Perception network: A multi-layer perceptron structure is used to extract specific optical characteristic parameters of water bodies.

[0101] 5. Fusion network: A fully connected neural network structure is used to integrate the output features of each branch network to give the final water body classification result.

[0102] The process of establishing the training data set for the water optical classification model includes: collecting the measured optical property data of various water types and the corresponding multispectral remote sensing images, preprocessing and labeling the data, and constructing a training sample set containing different water types. The training process includes: initializing the model parameters, using the training sample set for forward propagation and back propagation, optimizing the model parameters, and iterating the training until the model converges.

[0103] The specific implementation of step S50 is as follows: using the actual water depth values ​​of multiple field sampling points obtained previously , and some multispectral remote sensing images of the corresponding grid unit sea area and water body classification results , build a fine-tuning training set This fine-tuning training set will be used to subsequently optimize the pre-trained satellite remote sensing image water depth inversion model to make it better adapt to the characteristics of the current sea area to be tested. The construction process of the fine-tuning training set includes: first matching the location information of the field sampling points with the corresponding grid unit sea area, and then extracting some multispectral remote sensing image data of these grid unit sea areas. and water body classification results Finally, these data are compared with the actual water depth values ​​of the field sampling points. Composing fine-tuning training set .

[0104] The specific implementation of step S60 is as follows: Use the fine-tuning training set constructed above , optimize the pre-trained satellite remote sensing image water depth inversion model, so as to obtain a water depth inversion model suitable for the current sea area to be measured. The satellite remote sensing image water depth inversion model used here adopts a deep learning method, which contains the following main modules: 1. Multi-layer convolutional neural network: used to extract information from multispectral remote sensing images Extract the characteristics of water bodies.

[0105] 2. Fully connected layer: used to perform water depth analysis based on the extracted features estimate.

[0106] 3. Residual connection: used to improve the performance of the model.

[0107] The loss function of the model can be defined as: ; in, The model predicts The water depth value of each grid cell.

[0108] The process of constructing the training data set for the water depth inversion model includes: collecting measured water depth data from multiple sea areas And the corresponding multispectral remote sensing image , perform data preprocessing and registration, and construct a training sample set containing different water depth ranges The training process includes: initializing model parameters, using training sample sets Perform forward propagation and back propagation, optimize model parameters, use the validation set to evaluate model performance, and iterate training until the model converges.

[0109] When fine-tuning and optimizing the pre-trained model, the specific steps are as follows: 1. Freeze the pre-trained model Layer parameters, only fine-tuned Layer parameters; 2. Use the fine-tuning training set constructed earlier Perform training with a small number of iterations; 3. Adjust the learning rate and optimizer parameters; 4. Use early stopping strategy to prevent overfitting.

[0110] Through this fine-tuning method, we can fully utilize the feature extraction capabilities of the pre-trained model, while at the same time customizing and optimizing it according to the characteristics of the current sea area to be tested, thereby obtaining a more accurate water depth inversion model.

[0111] The specific implementation of step S70 is as follows: the water depth inversion model obtained by the above optimization is applied to each grid unit sea area, and the partial multispectral remote sensing image data corresponding to the grid unit sea area is input. and water body classification results , you can get the water depth value of the sea area of ​​the grid unit The multispectral remote sensing image data input here After the aforementioned atmospheric correction, the water body classification results By inputting all the grid unit sea areas of the entire sea area to be measured into the water depth inversion model one by one, the water depth estimation value of each grid unit sea area can be obtained. .

[0112] The specific implementation of step S80 is as follows: the estimated water depth of each grid unit sea area obtained in the previous step Perform spatial interpolation and smoothing to generate a continuous water depth distribution map of the entire sea area to be measured The specific operations are as follows: 1. Spatial interpolation: Using the Kriging interpolation algorithm, according to the known grid unit water depth value Perform spatial interpolation in the entire sea area to obtain a continuous water depth distribution The Kriging interpolation algorithm can be expressed as: ; in, is the interpolation weight coefficient, which can be determined by solving the Kriging equations.

[0113] 2. Smoothing: Interpolation of the water depth distribution map Gaussian filtering can effectively suppress the problem of discontinuous water depth distribution caused by data discreteness. Gaussian filtering can be expressed as: ; in, is the standard deviation of the Gaussian kernel function.

[0114] Through the above spatial interpolation and smoothing processing, a continuous water depth distribution map of the entire sea area to be measured can be obtained. , providing necessary data support for subsequent accuracy evaluation and verification.

[0115] The specific implementation of step S90 is as follows: using the actual water depth value of the field sampling point obtained previously , the generated water depth distribution map Perform accuracy assessment and verification. The specific operations are as follows: 1. Calculated water depth distribution map Compared with the actual water depth at the field sampling point The root mean square error (RMSE) between ; 2. Calculated water depth distribution map Compared with the actual water depth at the field sampling point The relative error between: ; 3. The calculated RMSE and relative error Compare with the preset accuracy threshold. The preset accuracy threshold is generally RMSE less than 0.0375~0.05, and the relative error is within the allowable range.

[0116] If the current accuracy does not meet the requirements, it is necessary to repeat steps S60 to S90 to further fine-tune and optimize the water depth inversion model until the accuracy meets the expectations. Through this iterative optimization method, the final generated water depth distribution map can be ensured. It is highly consistent with the actual situation.

[0117] In general, the water depth inversion method based on satellite remote sensing image data proposed in the present invention mainly includes the following key steps: 1. Obtain multispectral remote sensing image data of the sea area to be measured and field sampling water depth data , and atmospheric correction is performed.

[0118] 2. Divide the sea area into grids and extract four water optical characteristic parameters of each grid unit sea area: water reflectivity , absorption coefficient , scattering coefficient and algae concentration .

[0119] 3. Use the pre-trained water optical classification model to classify the unit grid sea area.

[0120] 4. Construct fine-tuning training set ,Perform targeted optimization on the pre-trained water depth inversion model.

[0121] 5. Apply the optimized water depth inversion model to estimate the water depth value of each grid unit sea area .

[0122] 6. Water depth value for unit grid Perform spatial interpolation and smoothing to generate a continuous water depth distribution map .

[0123] 7. Use the water depth data of field sampling points The generated water depth distribution map Conduct accuracy assessment and verification, and perform iterative optimization when necessary.

[0124] The entire method makes full use of the spatial coverage advantage of satellite remote sensing data and combines it with field measured data to fine-tune the model, thereby obtaining more accurate water depth distribution information and providing valuable data support for applications such as sea area monitoring and mariculture.

[0125] In order to better understand and implement the present invention, Example 2 of a specific application scenario of the present invention is provided below. The scenario of Example 2 is as follows: the environment of a certain offshore waters is relatively complex, and detailed measurement of the water depth of the offshore waters is required. The measurement team uses the present invention to conduct a water depth survey.

[0126] First, the survey team collected multispectral satellite remote sensing image data of the offshore area. Specifically, they obtained multispectral images from the stars, with a resolution of 10 meters and 13 bands, covering the spectral range from visible light to near-infrared. At the same time, they also collected water depth measurement data from 35 field sampling points for subsequent model training and verification. For the collected satellite image data, the researchers first performed atmospheric correction processing, using the FLAASH model to remove the influence of the atmosphere on the image. The corrected reflectance image is shown in Table 1: Table 1. Reflectance data of multispectral satellite remote sensing images of offshore waters Wavelength (nm) 1 2 3 4 5 6 7 8 9 10 11 12 13 R1 0.033 0.027 0.035 0.041 0.048 0.054 0.058 0.062 0.059 0.053 0.049 0.047 0.045 R2 0.029 0.024 0.032 0.039 0.045 0.051 0.055 0.059 0.056 0.050 0.046 0.044 0.042 … … … … … … … … … … … … … … R500 0.041 0.035 0.043 0.049 0.055 0.062 0.065 0.069 0.065 0.059 0.055 0.053 0.051 Among them, R1~R500 represent the reflectivity data of 500 different locations in the offshore waters.

[0127] Next, the survey team divided the offshore area into spatial grids, For each grid cell, they extracted four key water optical characteristic parameters: water reflectivity, ; Absorption coefficient ; Scattering coefficient ; Algae concentration ; The measurement team solved the undetermined parameters involved in these calculation formulas using the aforementioned optimization algorithm. The specific results are shown in Table 2: Table 2. Water optical characteristic parameters of grid cells in the offshore area

[0128] After obtaining the water optical characteristic parameters of each grid cell, the measurement team input them into a pre-trained water optical classification model based on an embedded neural network to obtain the water classification results for each grid cell.

[0129] Using this water optical classification model, the survey team classified 2,000 grid cells in the offshore area. Next, the survey team used the 35 field water depth sampling points collected to build a fine-tuning training set to optimize the pre-trained satellite remote sensing image water depth inversion model. The inversion model uses a deep learning method, including multi-layer convolutional networks, fully connected layers, and residual connections. By performing a small amount of iterative training on the fine-tuning training set and appropriately adjusting the learning rate and optimizer parameters, a water depth inversion model that is more suitable for the current offshore area was finally obtained.

[0130] Figure 2 :Comparison of spectral characteristics of different sea area types. This is a line graph showing the reflectance characteristics of three different water types (clean water, eutrophic water and turbid water) at 13 different wavelengths. Different line types and markers are used in the graph to distinguish the water types. The horizontal axis represents wavelength (nm) and the vertical axis represents reflectance. It can be clearly seen from the graph that the reflectance of turbid water is generally higher than that of the other two water bodies, while the reflectance of clean water is the lowest, which is related to the content of suspended matter and algae in the water.

[0131] Figure 3 :Water depth-seaweed concentration-reflectivity three-dimensional relationship diagram, this is a three-dimensional surface diagram that shows the complex relationship between water depth, seaweed concentration and reflectivity. The X-axis represents water depth (0-20m), the Y-axis represents seaweed concentration (0-10mg / m³), and the Z-axis represents reflectivity. Color mapping is used in the figure to enhance the visualization effect. It can be clearly seen that with the increase of water depth and the change of seaweed concentration, the reflectivity shows a nonlinear change characteristic.

[0132] Using the optimized water depth inversion model, the survey team estimated the water depth of 2,000 grid cells in the offshore area and obtained the water depth value of each grid cell. In order to generate a continuous water depth distribution map of the entire offshore area, they used the Kriging interpolation algorithm to spatially interpolate the unit water depth value and used Gaussian filtering for smoothing.

[0133] Figure 4 :The contour distribution map of water depth in offshore waters, which is a composite map combining contour lines and scattered points, shows the water depth distribution characteristics of offshore waters. The red triangles in the figure represent the locations of 35 sampling points, and the contour lines and fill colors in the background represent the continuous water depth distribution obtained by interpolation. The RdYlBu color scheme is used in the figure, with blue areas representing deeper waters and red areas representing shallower waters. It can be clearly seen that the water depth in the central area of ​​the offshore waters is larger, while the water depth near the shore is smaller.

[0134] Finally, the survey team used the 35 field water depth sampling points collected to evaluate and verify the accuracy of the generated water depth distribution map. Specifically, they calculated the root mean square error (RMSE) and relative error between the generated water depth distribution map and the measured water depth value. The results are shown in Table 3: Table 3. Accuracy evaluation of water depth distribution map index Numeric RMSE 0.42 m Relative error 7.8% It can be seen from Table 3 that the RMSE between the water depth distribution map generated by the method of the present invention and the measured water depth value is 0.42 meters, and the relative error is 7.8%, both of which meet the preset accuracy requirements (RMSE < 0.5 meters, relative error < 10%). This shows that the water depth inversion method of the present invention can obtain relatively accurate water depth distribution information of the entire offshore area, providing reliable data support for subsequent applications.

[0135] In summary, in this embodiment, the surveying team successfully applied the water depth inversion method based on satellite remote sensing images proposed in the present invention to conduct a water depth survey of a certain inland offshore sea area. Through key steps such as atmospheric correction, water body optical property extraction, water body classification, model fine-tuning optimization, and continuous water depth distribution generation, a high-precision water depth distribution map of the entire offshore sea area was finally obtained, meeting the actual application needs of the local government. This method gives full play to the advantages of remote sensing technology, combines advanced machine learning algorithms, solves the main problems existing in existing water depth inversion technology, and provides reliable data support for water resources management, marine aquaculture and other fields, with broad application prospects.

[0136] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A water depth inversion method based on satellite remote sensing image data, characterized in that: The following steps are involved: S10, obtaining a multispectral satellite remote sensing image of the sea area to be measured and actual water depth values ​​of multiple field sampling points of the sea area to be measured, and performing atmospheric correction processing on the multispectral satellite remote sensing image; S20, dividing the sea area to be measured into grids to obtain a plurality of unit grid sea areas; S30, using spectral analysis methods, extracting water optical characteristic parameters of each unit grid sea area from the multi-spectral satellite remote sensing image after atmospheric correction, including water reflectivity, absorption coefficient, scattering coefficient and seaweed concentration; S40, inputting the water body optical characteristic parameters into a pre-trained water body optical classification model to obtain the water body classification of each unit grid sea area; S50, establishing a fine-tuning training set according to the actual water depth values ​​of the multiple field sampling points and part of the multispectral satellite remote sensing images and water body classification of the corresponding unit grid sea area; S60, using the fine-tuning training set to fine-tune the pre-trained satellite remote sensing image water depth inversion model to obtain a water depth inversion model suitable for the sea area to be measured; S70, inputting part of the multispectral satellite remote sensing images and water body classification data corresponding to each unit grid sea area into the water depth inversion model to obtain the water depth value of each unit grid sea area; S80, performing spatial interpolation and smoothing processing on the water depth value of each unit grid sea area, and generating a continuous water depth distribution map of the entire sea area to be measured; S90, using the actual water depth values ​​of the field sampling points to evaluate and verify the accuracy of the generated water depth distribution map. If the accuracy does not meet the requirements, repeat S60 to S90 until the accuracy meets the requirements.

2. The method for water depth inversion based on satellite remote sensing image data according to claim 1, characterized in that: The accuracy meets the requirements, which specifically means that the root mean square error (RMSE) between the generated water depth distribution map and the actual water depth value of the field sampling point is less than a preset threshold, and the relative error is within the allowable range; the preset threshold is set based on experience.

3. The method for water depth inversion based on satellite remote sensing image data according to claim 1, characterized in that: The water optical classification model adopts an embedded neural network. It includes feature extraction module, water optical property equation module, error correction lightweight neural network group, perception network and fusion network.

4. The method for water depth inversion based on satellite remote sensing image data according to claim 3 is characterized in that: The feature extraction module is used to extract the spectral features of water bodies from multi-spectral satellite remote sensing images. The input is the multi-band reflectance data after atmospheric correction, and the output is the spectral feature vector group of water bodies. The structure is a multi-layer convolutional neural network. The water body optical characteristic equation module includes a water body reflectivity calculation equation, an absorption coefficient calculation equation, a scattering coefficient calculation equation, and a seaweed concentration calculation equation; The error correction lightweight neural network group includes a water body reflectivity calculation equation error correction network, an absorption coefficient calculation equation error correction network, a scattering coefficient calculation equation error correction network and a seaweed concentration calculation equation error correction network; The perception network is used to extract specific optical properties of water bodies, the input is a spectral feature vector group of water bodies, the output is specific optical property parameters, and the structure is a multi-layer perceptron; The fusion network is used to integrate the output results of each branch network. The input is the output features of each branch network, and the output is the final water body classification result. The structure is a fully connected neural network.

5. The method for water depth inversion based on satellite remote sensing image data according to claim 4 is characterized in that: The water body reflectance calculation equation is used to calculate the reflectance of the water body, the input is the spectral feature vector group of the water body, and the output is the estimated water body reflectance; The absorption coefficient calculation equation is used to calculate the absorption coefficient of the water body, the input is the spectral feature vector group of the water body, and the output is the estimated absorption coefficient; The scattering coefficient calculation equation is used to calculate the scattering coefficient of the water body, the input is the spectral feature vector group of the water body, and the output is the estimated scattering coefficient; The seaweed concentration calculation equation is used to calculate the seaweed concentration of a water body, the input is a spectral feature vector group of the water body, and the output is the estimated seaweed concentration.

6. The method for water depth inversion based on satellite remote sensing image data according to claim 5, characterized in that: The water body reflectance calculation equation error correction network is used to correct the reflectance calculation error, the input is the spectral feature vector group and the calculation result of the water body reflectance calculation equation, the output is the corrected reflectance, and the structure is MobileNet; The absorption coefficient calculation equation error correction network is used to correct the error of absorption coefficient calculation, the input is the spectral feature vector group and the calculation result of the absorption coefficient calculation equation, the output is the corrected absorption coefficient, and the structure is MobileNet; The scattering coefficient calculation equation error correction network is used to correct the error of the scattering coefficient calculation, the input is the spectral feature vector group and the calculation result of the scattering coefficient calculation equation, the output is the corrected scattering coefficient, and the structure is MobileNet; The seaweed concentration calculation equation error correction network is used to correct the error of seaweed concentration calculation, the input is the spectral feature vector group and the calculation result of the seaweed concentration calculation equation, the output is the corrected seaweed concentration, and the structure is MobileNet.

7. The method for water depth inversion based on satellite remote sensing image data according to claim 6, characterized in that: The steps of establishing the training data set of the water body optical classification model specifically include: collecting measured optical characteristic data of various water body types and corresponding multispectral satellite remote sensing images, performing data preprocessing and annotation, and constructing a training sample set containing different water body types.

8. The method for water depth inversion based on satellite remote sensing image data according to claim 7, characterized in that: The satellite remote sensing image water depth inversion model adopts a deep learning method, and its specific structure is: a multi-layer convolutional neural network is used for feature extraction, a fully connected layer is used for water depth estimation, and a residual connection is used to improve model performance; the steps for establishing a training data set for the satellite remote sensing image water depth inversion model specifically include: collecting measured water depth data of multiple sea areas and corresponding multispectral satellite remote sensing images, performing data preprocessing and alignment, and constructing a training sample set containing different water depth ranges.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores program instructions, and when the program instructions are executed in a computer, they are used to execute the water depth inversion method based on satellite remote sensing image data as described in any one of claims 1 to 8.

10. A water depth inversion system based on satellite remote sensing image data, characterized in that: A computer-readable storage medium comprising the computer-readable storage medium of claim 9.

Citation Information

Patent Citations

  • Multi-source Remote Sensing Fusion Inversion Method for Shallow Sea Depth

    CN105651263B

  • Permeable band ratio factor-based water depth inversion method

    CN102176001A

  • Multispectral remote sensing water depth inversion method and device based on semi-analytic model

    CN105930304A

  • Marine sediment testing system and method based on deep learning

    CN112162979A

  • Water depth inversion method based on improved four-waveband remote sensing image QAA algorithm water quality inversion result

    CN113793374A

Cited By

  • Shallow sea water depth inversion method and system based on multi-angle optical radiation transmission model

    CN120008564A

  • SAR remote sensing water body extraction method, device and equipment based on backscattering characteristics

    CN120374648A

  • Optical remote sensing shallow sea water depth inversion method and device

    CN120766141A

  • CDOM absorption coefficient inversion method and system based on liquid neural network

    CN121954868A