Shallow sea water depth hyperspectral inversion method considering characteristics of seabed

Abstract

The present application discloses a kind of shallow sea water depth hyperspectral inversion method considering bottom characteristics, comprising the following steps: shallow sea hyperspectral remote sensing image is obtained by combining water depth data in hyperspectral remote sensing image with spaceborne laser radar data;Spectral data of the bottom of shallow sea hyperspectral remote sensing image water edge line is sampled, and the spectral data of the same kind of bottom is averaged, then assigned to the spatial distribution data of bottom, to obtain bottom spectral image;Shallow sea hyperspectral remote sensing image and bottom spectral image in the coverage area of spaceborne laser radar data are used as input data of training sample, water depth data in spaceborne laser radar data is used as output data of training sample, shallow sea water depth inversion model is trained, and shallow sea water depth inversion is carried out using trained shallow sea water depth inversion model.The present application fully excavates and utilizes bottom characteristics to weaken shallow sea bottom heterogeneous interference, and provides a more accurate method for shallow sea water depth remote sensing inversion.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of shallow water depth inversion methods, in particular to a shallow water depth hyperspectral inversion method considering bottom characteristics. BACKGROUND

[0002] Shallow water depth data plays an important role in the study of marine and coastal areas, and is widely used in marine ecosystem protection, coastline monitoring, channel management, and seabed topography mapping. Traditional water depth measurement methods usually rely on shipborne echo sounders and airborne LiDAR technology (Pacheco, Horta et al. 2015, Kerr and Purkis 2018, Li, Tang et al. 2023). Echo sounders can provide high-precision water depth data, but their data acquisition range is limited, especially in large areas of shallow water (Chu, Cheng et al. 2023). While LiDAR technology can obtain water depth data with high spatial resolution, it is costly (Stumpf, Holderied et al. 2003). These methods are usually not suitable for large-scale, cost-effective water depth measurement, especially in remote or inaccessible areas. Compared with the above methods, satellite-derived bathymetry (SDB) is not limited by region, low cost and high efficiency, and is an effective method for large-scale water depth measurement (Lyzenga, Malinas et al. 2006).

[0003] Water depth inversion algorithms include empirical models, semi-theoretical and semi-empirical models, and theoretical models (Chu, Cheng et al. 2023). Among them, empirical models and semi-theoretical and semi-empirical models are widely used in water depth inversion of different types of water bodies due to their relatively simple calculation process and high accuracy (Nguyen, Liquet et al. 2021). In recent years, with the improvement of remote sensing technology and computing power, especially the development of machine learning algorithms, these data-driven methods have been increasingly applied in water depth inversion (Nagamani, Chauhan et al. 2012). Compared with traditional models, machine learning algorithms can better capture the complex nonlinear relationship between water body reflection signals and water depth, thereby improving the inversion accuracy and robustness (Wu, Mao et al. 2022, Chu, Cheng et al. 2023, Li, Tang et al. 2023).

[0004] Although the SDB water depth inversion technology has made significant progress, its inversion accuracy is often affected by the characteristics of the seabed (Philpot 1989, Stumpf, Holderied et al. 2003, McKinna, Fearns et al. 2015). When the water depth is fixed, the type of seabed affects the intensity of the signal received by the remote sensing sensor, which in turn affects the results of remote sensing water depth inversion (Reichstetter, Fearns et al. 2015, Arabi, Salama et al. 2020). Studies have shown that by classifying the study area into multiple seabed single and uniform sub-regions through seabed classification, and constructing a water depth model in each type of seabed area, the influence of seabed factors on the water depth model can be weakened to a certain extent, thereby improving the accuracy of water depth inversion (Nagamani, Chauhan et al. 2012, Ma, Tao et al. 2014, Cheng, Ma et al. 2021).

[0005] Current studies show that the introduction of spatial characteristics of seabed can improve the accuracy of water depth inversion, but the study of another important feature of seabed, i.e. spectral characteristics, is less. Seabed spectral information can effectively reflect the physical and chemical properties of seabed, and its performance in water will undergo complex interactions with water depth and water optical properties (Hochberg 2003, Jay, Guillaume et al. 2017, Liu, Deng et al. 2018, Arabi, Salama et al. 2020), affecting the results of water depth inversion.

[0006] How to mine and integrate the spatial characteristics of seabed and the spectral characteristics of seabed to further weaken the interference of seabed heterogeneity on remote sensing inversion of water depth and improve the inversion accuracy is still lacking effective solutions. SUMMARY

[0007] In view of the problems in the related art, the present application proposes a shallow water depth hyperspectral inversion method considering seabed characteristics to overcome the above technical problems existing in the prior art.

[0008] To this end, the specific technical solutions adopted by the present application are as follows: a shallow water depth hyperspectral inversion method considering seabed characteristics, comprising the following steps:

[0009] S1, obtain hyperspectral remote sensing image and spaceborne laser radar data and preprocess, extract seawater pixels in the hyperspectral remote sensing image, and combine the water depth data in the spaceborne laser radar data to remove the pixels with water depth greater than the water depth threshold H in the seawater pixels, to obtain a shallow hyperspectral remote sensing image;

[0010] S2, obtain the spatial distribution data of the bottom material based on the shallow sea hyperspectral remote sensing image supervised classification, sample the spectral data of the bottom material at the water edge line of the shallow sea hyperspectral remote sensing image, and obtain the mean value of the spectral data of the same type of bottom material to obtain the spectrum of each type of bottom material, and assign the spectrum of the bottom material to the spatial distribution data to obtain the bottom material spectrum image, realize the composition of the bottom material spectrum, and the bottom material at least includes three types: coral, coral sand and detritus;

[0011] S3, a shallow sea water depth inversion model is established, the input data of the training sample includes the shallow sea hyperspectral remote sensing image and the bottom material spectrum image of the shallow sea hyperspectral remote sensing image covered by the spaceborne laser radar data, and the output data of the model training sample is the water depth data in the spaceborne laser radar data; the model is trained using the training sample; the shallow sea hyperspectral remote sensing image obtained in step S1 and the bottom material spectrum image obtained in step S2 are input into the trained shallow sea water depth inversion model, and the shallow sea water depth inversion result is output.

[0012] The spectrum data of the bottom material type are connected with the shallow sea hyperspectral remote sensing image as the input of the shallow sea water depth inversion model, the bottom material characteristics are fully mined and utilized to weaken the interference of the sea bottom material isomerization, and a more accurate method is provided for shallow sea water depth remote sensing inversion.

[0013] Further, step S1 specifically includes the following steps:

[0014] S11, the ZY-01 hyperspectral remote sensing image is geometrically corrected, atmospherically corrected and flare corrected; and the cloud and land parts of the remote sensing image in the region are masked and treated, and the sea water pixels are retained;

[0015] S12, the sea surface and seabed photon point cloud data in the ICESat-2 ATL03 laser radar data are extracted respectively, and the seabed photon point cloud data is refracted corrected, the tidal height at the time of collecting the ZY-01 hyperspectral remote sensing image and the ICESat-2 laser radar data is obtained from the tidal value provided by the tide station near the study area, and the tidal correction is carried out;

[0016] S13, the point cloud water depth data and the hyperspectral image data of the study area are geographically registered, and the range with water depth greater than the water depth threshold H is extracted from the point cloud data, the image pixels corresponding to the range are deep sea pixels, the spectral characteristics of the deep sea pixels are extracted, and the hyperspectral image of the study area is segmented according to the spectral characteristics of the deep sea pixels to obtain the deep sea hyperspectral remote sensing image and the shallow sea hyperspectral remote sensing image.

[0017] Further, step S2 specifically includes the following steps:

[0018] S21, the preprocessed shallow sea hyperspectral remote sensing image is taken as input;

[0019] S22, construct a bottom classification sample criterion, collect different bottom samples in the study area, the bottom includes at least three types: coral, coral sand and clastic, and the maximum likelihood method is used for supervised classification in the study area to obtain the spatial distribution data of the bottom;

[0020] S23, the spectral data of the bottom near the water edge line of the shallow sea hyperspectral remote sensing image is sampled, and the average value of the spectral data of the same type of bottom is obtained, so that the spectrum of each type of bottom is obtained;

[0021] S24, based on the spatial distribution data of the bottom and the spectral data of the water body, the spectral data of the bottom is added to the corresponding spatial distribution data of the bottom, so that the spectral image of the bottom is obtained, and the partitioned bottom spectral composite is realized.

[0022] The shallow sea water depth inversion model is any kind of machine learning model.

[0023] As preferred, the shallow sea water depth inversion model selects an LGBM (Light Gradient Boosting Machine) water depth inversion model or an RF (Random Forest) water depth inversion model.

[0024] Wherein, LGBM is a high-efficiency machine learning framework based on gradient boosting tree (GBT) algorithm. It is developed by Microsoft Asia Research Institute, and it is particularly suitable for processing large data sets and high-dimensional features. LGBM adopts a histogram-based decision tree algorithm and a depth-constrained leaf growth strategy, thereby significantly speeding up the training speed and reducing the memory consumption. By integrating multiple decision trees, LGBM can capture the complex nonlinear relationship between water depth and hyperspectral features, thereby providing more accurate depth estimation.

[0025] Wherein, RF is an ensemble learning algorithm based on decision tree method. It constructs many decision trees in the training process, and outputs the majority of their predictions for classification tasks and the average value for regression tasks. By using bootstrap sampling and random feature selection, RF reduces the risk of overfitting and enhances the generalization ability of the model.

[0026] According to another aspect of the present application, a system for implementing the above-mentioned shallow sea water depth hyperspectral inversion method considering the characteristics of the bottom is provided, which comprises a data preprocessing module, a bottom spatial and spectral feature extraction module and an algorithm establishment module.

[0027] Wherein, the data preprocessing module is used for obtaining and preprocessing the hyperspectral remote sensing image and the spaceborne laser radar data, extracting the seawater pixels in the hyperspectral remote sensing image, and combining the water depth data in the spaceborne laser radar data to remove the pixels with water depth greater than the water depth threshold H in the seawater pixels, so as to obtain the shallow sea hyperspectral remote sensing image.

[0028] The substrate space and spectral feature extraction module obtains the spatial distribution data of the substrate based on the supervised classification of the shallow sea hyperspectral remote sensing image, performs spectral data sampling on the substrate at the water edge line of the shallow sea hyperspectral remote sensing image, and obtains the spectrum of each type of substrate by averaging the spectral data of the same type of substrate, and the spectrum of the substrate is assigned to the spatial distribution data to obtain the substrate spectral image, and the substrate spectrum is combined, wherein the substrate at least includes three types: coral, coral sand and clastic;

[0029] The algorithm establishment module is used for establishing a shallow sea water depth inversion model, the input data of the training sample includes the shallow sea hyperspectral remote sensing image and the substrate spectral image of the shallow sea covered by the spaceborne laser radar data, and the output data of the model training sample is the water depth data in the spaceborne laser radar data; the model is trained by using the training sample; the shallow sea hyperspectral remote sensing image obtained in step S1 and the substrate spectral image obtained in step S2 are input into the trained shallow sea water depth inversion model, and the shallow sea water depth inversion result is output.

[0030] The beneficial effects of the present application are:

[0031] 1、The present application proposes a shallow sea water depth hyperspectral inversion considering the substrate characteristics, fully excavates and utilizes the substrate characteristics to weaken the interference of the seabed substrate isomerization, and provides a more accurate method for shallow sea water depth remote sensing inversion.

[0032] 2、The present application proposes a strategy of extracting substrate spatial features based on remote sensing image supervised classification and extracting substrate spectral features based on image water edge line, which improves the water depth inversion accuracy.

[0033] 3、The present application is based on the commonly used LGBM machine learning algorithm and RF machine learning algorithm, the algorithm uses the same scene as the classic algorithm, which is conducive to the popularization and application of the method, and provides technical support for island reef construction, navigation safety, ecological protection and other applications. BRIEF DESCRIPTION OF DRAWINGS

[0034] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0035] Figure 1 It is a flowchart of a shallow sea water depth hyperspectral inversion considering substrate characteristics according to an embodiment of the present application;

[0036] Figure 2 It is a general diagram of the research area in the example;

[0037] Figure 3is a track plot of ICESat-2 point cloud in the study area in the example;

[0038] Figure 4 is a seabed classification result plot in the study area in the example;

[0039] Figure 5 is a shallow water depth inversion result plot in the study area in the example;

[0040] Figure 6 is a system block diagram of a shallow water depth hyperspectral inversion system considering seabed characteristics according to an embodiment of the present application.

[0041] In the figure: 1, data preprocessing module; 2, seabed space and spectral feature extraction module; 3, algorithm establishment module. DETAILED DESCRIPTION

[0042] To further illustrate the embodiments, the present application provides drawings which are part of the disclosure of the present application, mainly used to illustrate the embodiments, and can be used to explain the operating principle of the embodiments in conjunction with the related description of the specification. Those skilled in the art should be able to understand other possible implementations and advantages of the present application by referring to these contents. The components in the figures are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0043] The present application will be further described in conjunction with the drawings and specific embodiments, as Figure 1 is a flow chart of the shallow water depth hyperspectral inversion method considering seabed characteristics of the present application. The method includes the following steps:

[0044] S1, obtain hyperspectral remote sensing image and spaceborne laser radar data and perform preprocessing, extract seawater pixels in the hyperspectral remote sensing image, and combine the water depth data in the spaceborne laser radar data to remove the pixels with water depth greater than the water depth threshold H in the seawater pixels, to obtain a shallow water hyperspectral remote sensing image.

[0045] This embodiment takes a certain sea area island as the experimental area (as Figure 2 shown), the remote sensing image data is ZY-01 hyperspectral satellite image, the image acquisition time is January 13, 2023, 17:04 (UTC), the wavelength of ZY-01 is 450nm-700nm, there are 36 bands, and the spatial resolution is 30m. The water depth data is ICESat-2 ATL03, and the refraction correction is performed on the extracted effective seabed photon points to convert the datum plane of the point cloud water depth data to the local average sea level (as Figure 3 shown). Figure 3 In the figure, the black straight line segment is the coverage area of the spaceborne laser radar data, and it can be seen that the water depth data in the point cloud data is local data and cannot cover the entire test area.

[0046] The step S1 specifically comprises the following steps:

[0047] S11, using professional remote sensing processing software ENVI to perform geometric correction, atmospheric correction and flare correction on the ZY-01 hyperspectral remote sensing image; and performing mask processing on the cloud and land part of the remote sensing image in the region, and retaining the seawater pixels;

[0048] S12, using PhotonLabeler software to extract sea surface and seabed photon point cloud data in ICESat-2 ATL03 laser radar data respectively, and performing refraction correction on the seabed photon point cloud data, obtaining the tidal height at the time of collecting the ZY-01 hyperspectral remote sensing image and the ICESat-2 laser radar data from the tidal value provided by the tide station near the study area, and carrying out tidal correction.

[0049] S13, geographically registering the point cloud water depth data and the hyperspectral image data of the study area, extracting the range with water depth greater than 20m from the point cloud data, taking the image pixels corresponding to the range as deep sea pixels, extracting the spectral characteristics of the deep sea pixels, and segmenting the hyperspectral image of the study area according to the spectral characteristics of the deep sea pixels to obtain deep sea hyperspectral remote sensing image and shallow sea hyperspectral remote sensing image. In addition, the deep sea range can also be obtained through the chart data, and the shallow sea hyperspectral remote sensing image can be obtained after masking the hyperspectral image data.

[0050] S2, obtaining the spatial distribution data of the bottom based on the supervised classification of the shallow sea hyperspectral remote sensing image, sampling the spectral data of the bottom near the waterline of the shallow sea hyperspectral remote sensing image, and averaging the spectral data of the same type of bottom to obtain the spectrum of each type of bottom, and assigning the spectrum of the bottom to the spatial distribution data to obtain the bottom spectral image, realizing the composite of the bottom spectrum, wherein the bottom includes at least three types: coral, coral sand and detritus. Step S2 specifically comprises the following steps:

[0051] S21, taking the preprocessed hyperspectral remote sensing image as input;

[0052] S22, constructing a bottom classification sample criterion, collecting different bottom samples in the study area, in this embodiment, the bottom includes three types: coral, coral sand and detritus, and using the maximum likelihood method to carry out supervised classification in the study area to obtain the spatial distribution data of the bottom;

[0053] S23, sampling the spectral data of the bottom near the waterline of the shallow sea hyperspectral remote sensing image and averaging the spectral data of the same type of bottom to obtain the spectrum of each type of bottom;

[0054] S24, based on the spatial distribution data of the bottom and the spectral data of the water body of the bottom, adding the spectral data of the bottom to the corresponding spatial distribution data of the bottom to obtain the spectral image of the bottom, and realizing the composite of the spectral data of the partitioned bottom.

[0055] According to the tidal table, the height of the tide at the time of image acquisition of the experimental area is 0.9 m (reference nearby tide station), the point cloud water depth of the experimental area has been corrected to the instantaneous sea surface water depth at the time of image acquisition, and the water depth of 0-20 m is extracted as the experimental water depth data of the application. The ZY-01 image pixel mean water depth is selected as the water depth, that is, the mean value of the measured water depth data in a single pixel is taken as the water depth value of the pixel.

[0056] In this embodiment, three types of typical bottom samples in the study area are collected, including corals, coral sand and clastic. The maximum likelihood method is used for bottom classification of the island, and the classification result is shown in Figure 4 . Among them, the land area has been masked.

[0057] S3, a shallow sea water depth inversion model is established, the input data of the training sample includes shallow sea hyperspectral remote sensing image and bottom spectrum image of the area covered by the spaceborne laser radar data, and the output data of the model training sample is the water depth data in the spaceborne laser radar data; the model is trained using the training sample; the shallow sea hyperspectral remote sensing image obtained in step S1 and the bottom spectrum image obtained in step S2 are input into the trained shallow sea water depth inversion model, and the shallow sea water depth inversion result is output.

[0058] Among them, the shallow sea water depth inversion model can be any machine learning model. The samples can be divided into training samples and test samples according to the proportion, the training samples are used for training the model, and the test samples are used for evaluating the trained model. In this embodiment, LGBM water depth inversion model considering bottom characteristics and RF water depth inversion model considering bottom characteristics are selected for learning and classification.

[0059] In specific application, the root mean square error (Root Mean Square Error, RMSE), the mean absolute error (the Mean Absolute Error, MAE) and the coefficient of determination (the coefficient of determination, R 2 ) are used to quantitatively evaluate the accuracy of the water depth inversion result output by the water depth inversion model. The smaller the RMSE and MAE, the higher the accuracy of the water depth inversion. The calculation formula of the quantitative index is as follows:

[0060]

[0061] n represents the number of water depth points; H estimation is the inverted water depth; H lidar is the water depth extracted from ICESat-2; and are the average value of the water depth extracted from ICESat-2 and the average value of the inverted water depth, respectively.

[0062] The water depth inversion results of the classic method and the substrate feature considering method proposed in the application are compared, and each method includes two groups of results of LGBM and RF algorithm. The LGBM algorithm considering the substrate feature is as shown in (a), and the RF algorithm considering the substrate feature is as shown in (b). The reef disc contour is clear, and the underwater topographic feature is smooth and less noisy in the figure. The water depth inversion accuracy is shown in Table 1: Figure 5 Figure 5

[0063] Table 1: Evaluation table of shallow sea water depth inversion accuracy

[0064]

[0065] As can be seen from Table 1, compared with the classic LGBM and RF algorithm, the RMSE, MAE error of the LGBM and RF algorithm considering the substrate feature is the lowest, and the correlation coefficient R 2 of the verification sample is the highest. The high spectral water depth inversion method considering the substrate feature proposed in the application has excellent water depth inversion performance.

[0066] As can be seen from the table, the RMSE, MAE and R 2 of the water depth inversion result of the classic LGBM algorithm are 1.30 m, 0.67 m and 0.83 respectively.

[0067] The RMSE, MAE and R 2 of the water depth inversion result of the LGBM algorithm considering the substrate feature are 1.05 m, 0.55 m and 0.89 respectively.

[0068] The RMSE, MAE and R 2 of the water depth inversion result of the classic RF algorithm are 1.29 m, 0.64 m and 0.83 respectively.

[0069] The RMSE, MAE and R 2 of the water depth inversion result of the RF algorithm considering the substrate feature are 1.05 m, 0.52 m and 0.89 respectively.

[0070] According to another embodiment of the application, a shallow sea water depth hyperspectral inversion system considering the substrate feature is provided, which comprises a data preprocessing module 1, a substrate space and spectral feature extraction module 2 and an algorithm establishment module 3.

[0071] The data preprocessing module 1 is used for acquiring and preprocessing hyperspectral remote sensing images and spaceborne laser radar data, extracting seawater pixels in the hyperspectral remote sensing images, and removing the pixels with water depth greater than a water depth threshold H in the seawater pixels in combination with the water depth data in the spaceborne laser radar data to obtain shallow sea hyperspectral remote sensing images.

[0072] ​​The substrate space and spectral feature extraction module 2 obtains the spatial distribution data of the substrate based on the supervised classification of the shallow sea hyperspectral remote sensing image, performs spectral data sampling on the substrate at the water edge line of the shallow sea hyperspectral remote sensing image, and obtains the spectrum of each type of substrate by averaging the spectral data of the same type of substrate, and assigns the spectrum of the substrate to the spatial distribution data to obtain a substrate spectral image, and realizes the combination of the substrate spectrum, wherein the substrate at least includes three types: coral, coral sand and clastic;

[0073] The algorithm establishment module 3 is used for establishing a shallow sea water depth inversion model, the input data of the training sample includes the shallow sea hyperspectral remote sensing image and the substrate spectral image of the shallow sea hyperspectral remote sensing image covered by the spaceborne laser radar data, and the output data of the model training sample is the water depth data in the spaceborne laser radar data; the model is trained using the training sample; the shallow sea hyperspectral remote sensing image obtained in step S1 and the substrate spectral image obtained in step S2 are input into the trained shallow sea water depth inversion model, and the shallow sea water depth inversion result is output.

[0074] In summary, by means of the above technical solutions of the present application, the present application proposes a shallow sea water depth hyperspectral inversion method considering substrate features, which fully excavates and utilizes substrate features to weaken the interference of seabed substrate isomerization, and provides a more accurate method for shallow sea water depth remote sensing inversion; the present application proposes a strategy of extracting substrate spatial features based on remote sensing image supervised classification and extracting substrate spectral features based on image water edge line, which excavates and utilizes substrate spatial features and substrate spectral features, and improves the water depth inversion accuracy; the present application is based on the commonly used LGBM machine learning algorithm and RF machine learning algorithm, the algorithm uses the same scene as the classic algorithm, which is conducive to the popularization and application of the present application method, and provides technical support for island reef construction, navigation safety, ecological protection and other applications.

[0075] The above only describes the preferred embodiments of the present application and does not limit the present application, and any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for shallow sea water depth hyperspectral inversion considering the characteristics of the seabed, comprising the following steps: S1. Obtain hyperspectral remote sensing images and spaceborne laser radar data and preprocess them, extract seawater pixels in the hyperspectral remote sensing images, and combine the water depth data in the spaceborne laser radar data to remove pixels with water depth greater than a water depth threshold H in the seawater pixels, thereby obtaining shallow sea hyperspectral remote sensing images; S2. Obtain spatial distribution data of the seabed based on supervised classification of the shallow sea hyperspectral remote sensing images, sample the spectral data of the seabed near the waterline of the shallow sea hyperspectral remote sensing images, and average the spectral data of the same type of seabed, thereby obtaining the spectrum of each type of seabed, assigning the spectrum of the seabed to the spatial distribution data, and obtaining a seabed spectral image, thereby realizing the composition of the seabed spectrum, wherein the seabed includes at least three types: coral, coral sand, and detritus; S3. Establish a shallow sea water depth inversion model, the input data of the training sample includes the shallow sea hyperspectral remote sensing images and the seabed spectral image of the area covered by the spaceborne laser radar data, and the output data of the model training sample is the water depth data in the spaceborne laser radar data; train the model using the training sample; input the shallow sea hyperspectral remote sensing images obtained in step S1 and the seabed spectral image obtained in step S2 into the trained shallow sea water depth inversion model, and output the shallow sea water depth inversion result.

2. The method according to claim 1, wherein, Step S1 specifically comprises the following steps: S11. Geometric correction, atmospheric correction, and flare correction are performed on the ZY-01 hyperspectral remote sensing images; and the cloud and land parts of the remote sensing images in the region are masked and processed, and the seawater pixels are retained; S12. The sea surface and seabed photon point cloud data in the ICESat-2 ATL03 laser radar data are extracted respectively, and the seabed photon point cloud data is corrected for refraction; the tidal height at the time of acquisition of the ZY-01 hyperspectral remote sensing images and the ICESat-2 laser radar data is obtained from the tidal values provided by the tide stations near the study area, and tidal correction is carried out; S13. Geographical registration is performed on the point cloud water depth data and the hyperspectral image data of the study area, the range of water depth greater than the water depth threshold H is extracted from the point cloud data, the image pixels corresponding to the range are deep sea pixels, the spectral features of the deep sea pixels are extracted, and the hyperspectral image of the study area is segmented according to the spectral features of the deep sea pixels, thereby obtaining the deep sea hyperspectral remote sensing image and the shallow sea hyperspectral remote sensing image. 3.The shallow water depth hyperspectral retrieval method considering the characteristics of seabed according to claim 1, wherein, Step S2 specifically comprises the following steps: S21. The preprocessed shallow sea hyperspectral remote sensing image is used as input; S22. Construct seabed classification sample criteria, collect different seabed samples in the study area, and the seabed includes at least three types: coral, coral sand, and detritus; use the maximum likelihood method to carry out supervised classification in the study area, thereby obtaining seabed spatial distribution data; S23. Sample the spectral data of the seabed near the waterline of the shallow sea hyperspectral remote sensing image and average the spectral data of the same type of seabed, thereby obtaining the spectrum of each type of seabed; S24. Based on the seabed spatial distribution data and the seabed water spectral data, add the seabed spectrum to the corresponding seabed spatial distribution data, thereby obtaining the spectral image of the seabed and realizing the composition of the seabed spectrum.

4. The method of claim 1, wherein, The shallow sea water depth inversion model is any machine learning model.

5. The method of claim 1, wherein, The shallow sea water depth inversion model is an LGBM water depth inversion model or an RF water depth inversion model.

6. The method of claim 1, wherein, The water depth threshold H is 20 meters.

7. A system for shallow water depth hyperspectral inversion considering seabed characteristics, for implementing the method for shallow water depth hyperspectral inversion considering seabed characteristics according to any one of claims 1-6, characterized in that, The system comprises a data preprocessing module (1), a bottom space and spectral feature extraction module (2), and an algorithm establishing module (3). The data preprocessing module (1) is configured to acquire and preprocess hyperspectral remote sensing images and spaceborne laser radar data, extract seawater pixels in the hyperspectral remote sensing images, remove pixels with water depth greater than a water depth threshold H from the seawater pixels in combination with water depth data in the spaceborne laser radar data, and obtain shallow sea hyperspectral remote sensing images. The bottom space and spectral feature extraction module (2) is configured to obtain spatial distribution data of the bottom based on supervised classification of the shallow sea hyperspectral remote sensing images, sample spectral data of the bottom at a waterline of the shallow sea hyperspectral remote sensing images, average spectral data of the same type of bottom, obtain spectral data of each type of bottom, assign the spectral data of the bottom to the spatial distribution data, obtain a bottom spectral image, and realize spectral data composition of the bottom, wherein the bottom comprises at least three types: coral, coral sand, and clastic. The algorithm establishing module (3) is configured to establish a shallow sea water depth inversion model, wherein input data of a training sample comprises shallow sea hyperspectral remote sensing images and bottom spectral images of a region covered by spaceborne laser radar data, and output data of the model training sample is water depth data in the spaceborne laser radar data; the model is trained using the training sample; and the shallow sea hyperspectral remote sensing images obtained in step S1 and the bottom spectral images obtained in step S2 are input into the trained shallow sea water depth inversion model to output shallow sea water depth inversion results.

Images