Nearshore surface suspended sediment analysis method based on deep learning model

Abstract

The present application relates to the technical field of marine environment monitoring, in particular to a nearshore surface suspended sediment analysis method based on a deep learning model. The method comprises: collecting surface optical remote sensing data and acoustic backscattering data of a target water area; preprocessing and spatiotemporal aligning the collected surface optical remote sensing data and acoustic backscattering data to generate a fusion data set; inputting the fusion data set into a first deep learning model to predict and output a suspended sediment distribution field map of the target water area; and according to the suspended sediment distribution field map of the target water area, using a second deep learning model to predict the suspended sediment distribution of the target water area at a future time. The present application decouples the bubble and sediment signals through a "double physical constraint" physical information neural network architecture, so that the present application achieves higher inversion accuracy in high-dynamic and high-turbidity nearshore areas including the surf zone, and significantly improves the accuracy and spatiotemporal resolution of nearshore suspended sediment analysis.

Description

Technical Field

[0001] This invention relates to the field of marine environmental monitoring technology, specifically to a method for analyzing nearshore surface suspended sediment based on a deep learning model. Background Technology

[0002] Suspended sediment is a key hydrodynamic and water quality parameter in nearshore waters (including estuaries, deltas, and coastal areas). The dynamic changes and transport processes of suspended sediment are not only major drivers of geomorphic evolution (such as coastal erosion and delta development), but also have a profound impact on aquatic ecosystems and human economic activities.

[0003] In industrial applications, accurate monitoring of suspended sediment is crucial. Firstly, it serves as a core basis for port maintenance and channel dredging. Sedimentation in channels and harbor basins reduces navigable depth, threatening navigational safety, thus necessitating periodic dredging. Dredging operations are extremely costly; accurate monitoring of suspended sediment and prediction of sedimentation are prerequisites for optimizing dredging plans, reducing maintenance costs, and assessing the environmental impact of dredging. Furthermore, high concentrations of suspended sediment reduce water transparency, impair photosynthesis, stress sensitive marine ecosystems such as coral reefs and seagrass beds, and may adsorb and transport pollutants such as heavy metals. Therefore, suspended sediment monitoring is of great significance for marine environmental protection.

[0004] However, the nearshore zone is one of the most hydrodynamically active regions on Earth, where suspended sediment is subject to multiple and nonlinear influences from tides, waves (especially breaking waves), turbulence, river inputs, and human activities, exhibiting extremely high spatiotemporal variability. This makes accurate, high-resolution measurements of the SSC a significant technical challenge.

[0005] To address this challenge, existing technologies have developed three main categories of monitoring methods: in-situ sampling, acoustic proxy, and optical remote sensing.

[0006] In-situ sampling schemes primarily utilize physical pump sampling, optical backscattering sensors, and laser in-situ scattering and transillumination (LST) instruments. These instruments determine suspended sediment by directly measuring the physical or optical properties of water samples. However, these methods are essentially "point measurements." While they can provide high-precision time-series data at the deployment point, they cannot provide detailed spatial and temporal profiles of suspended sediment. They cannot capture large-scale spatial distribution characteristics (e.g., a complete sediment plume).

[0007] Acoustic proxies primarily utilize acoustic Doppler current profilers (ADCPs). These instruments indirectly estimate suspended sediment by emitting sound waves and measuring the intensity of the signals scattered back by particles (sediment) in the water (i.e., acoustic backscattering). The advantage of ADCPs is that they can simultaneously provide flow profiles and suspended sediment profiles for the entire water column. However, acoustic schemes struggle to distinguish between bubbles and sediment. Nearshore wave breaking instantaneously generates a massive number of microbubbles in the water; these bubbles are also highly efficient acoustic scatterers, and their acoustic signals significantly overlap with sediment signals (especially fine-grained sediment). Existing acoustic algorithms perform well only under non-wave-breaking conditions, failing completely under wave-breaking conditions because they cannot distinguish the scattered signals from sediment and bubbles.

[0008] Optical remote sensing schemes utilize optical sensors mounted on satellites or drones to capture the reflectivity of water surfaces and establish a model relating reflectivity to surface suspended sediment. However, the physical characteristics of optical remote sensing limit its ability to measure only a few centimeters to a few meters below the water surface (the depth of light penetration), and it cannot provide information on the vertical structure of the water column. Summary of the Invention

[0009] To address the shortcomings of existing technologies, this invention provides a method for analyzing nearshore surface suspended sediment based on a deep learning model. This invention utilizes a systematic fusion approach, leveraging the complementary advantages of different data sources (acoustic and optical) and the unique capabilities of different deep learning architectures to collaboratively solve the problems existing in current technologies.

[0010] A method for analyzing nearshore surface suspended sediment based on a deep learning model includes the following steps:

[0011] Acquire surface optical remote sensing data of the target water area;

[0012] Deploy an unmanned surface vessel equipped with acoustic sensors to the target waters and use the acoustic sensors to acquire acoustic backscatter data below the unmanned surface vessel's trajectory;

[0013] The acquired surface optical remote sensing data and acoustic backscatter data are preprocessed and spatiotemporally aligned to generate a fused dataset;

[0014] The fused dataset is input into the first deep learning model, and a high-resolution suspended sediment distribution field map of the target water area is output.

[0015] Based on the high-resolution suspended sediment distribution field map of the target water area, the distribution of suspended sediment in the target water area at future times is predicted using a second deep learning model.

[0016] Preferably, the acoustic backscattering data is multi-frequency acoustic backscattering data, which includes scattering intensity information at different acoustic frequencies.

[0017] Preferably, the first deep learning model employs a physical information neural network architecture, which is trained by minimizing a hybrid loss function, specifically:

[0018] ;

[0019] In the formula, For the total loss, For data loss, For advection-diffusion physical loss, For acoustic physical loss, and These are hyperparameters used to balance different loss terms;

[0020] The data loss is used to ensure the accuracy of the model where data is available, the advection-diffusion physics loss is used to force the suspended sediment concentration field output by the model to be physically reasonable, and the acoustic physics loss is used to force the suspended sediment concentration and bubble porosity output by the model to be acoustically reasonable.

[0021] Preferably, the specific method for determining the acoustic physical loss is as follows:

[0022] The physical information output by the neural network at the coordinate points with acoustic measurement data is obtained, along with the concentration of suspended sediment and the void ratio of air bubbles.

[0023] Based on the above suspended sediment concentration and bubble porosity, an acoustic backscattering intensity is calculated using a multi-frequency acoustic scattering model.

[0024] The mean square error between the acoustic backscattering intensity and the measured acoustic backscattering intensity is the acoustic physical loss.

[0025] Preferably, the second deep learning model is a spatiotemporal prediction network, which includes convolutional long short-term memory units for predicting the spatiotemporal distribution of suspended sediment concentration in future time steps.

[0026] Preferably, the spatiotemporal prediction network adopts an encoder-predictor architecture, and both the encoder and the predictor are constructed from the convolutional long short-term memory units. The encoder and the predictor are configured as follows:

[0027] Encoder: Receives a sequence of spatial distribution maps of suspended sediment, processes the sequence through multi-layer convolutional long short-term memory units, extracts and encodes the multi-scale spatiotemporal features of the sequence, and generates a set of hidden states;

[0028] Predictor: Using the hidden state generated by the encoder as the initial state, it runs iteratively for multiple time steps in the future. In each step, a convolutional long short-term memory unit is used to predict the hidden state of the next time step, and the spatial distribution map of suspended sediment in the future time step is reconstructed from the hidden state.

[0029] Preferably, the method further includes self-supervised pre-training of the encoder of the spatiotemporal prediction network using mask reconstruction technology, specifically:

[0030] Unlabeled remote sensing images from different time periods and regions are acquired as a pre-training dataset;

[0031] Randomly mask certain pixel blocks in remote sensing images;

[0032] The encoder of the spatiotemporal prediction network is trained to reconstruct the masked pixels using the surrounding visible pixels;

[0033] The loss function is the difference between the reconstructed pixel and the original pixel.

[0034] A nearshore surface suspended sediment analysis system based on a deep learning model includes a data acquisition module, a data preprocessing module, a data analysis module, and a visualization output module.

[0035] The data acquisition module is used to collect acoustic and optical data related to the target water area;

[0036] The data preprocessing module is used to preprocess and spatiotemporally align the acoustic and optical data acquired by the data acquisition module, so as to unify the data under a single framework.

[0037] The data analysis data is used to analyze the data processed by the data preprocessing module using a deep learning model;

[0038] The visualization output module is used to visualize the analysis results of the data analysis module for decision support.

[0039] Preferably, the data processing module includes an acoustic data processing unit and an image data processing unit;

[0040] The acoustic data processing unit is used to perform backscatter data correction and georegistration on the collected acoustic data.

[0041] The image data processing unit is used to perform radiometric correction, atmospheric correction, geometric correction, and synchronization on the acquired image data.

[0042] Preferably, the data analysis module includes a first analysis and processing unit and a second analysis and processing unit;

[0043] The first analysis and processing unit is used to receive the data output by the data preprocessing module and use the first deep learning model to predict the distribution field of suspended sediment in the target water area.

[0044] The second analysis and processing unit is used to predict the future spatiotemporal distribution of suspended sediment in the target water area based on the suspended sediment distribution field output by the first analysis and processing unit and using a second deep learning model.

[0045] Compared with the prior art, the advantages of this invention are:

[0046] By employing a physical information neural network architecture with "dual physical constraints" of advection-diffusion physical loss and acoustic physical loss, the decoupling of bubble and sediment signals is achieved. This enables the scheme to achieve higher inversion accuracy than single technologies in high-dynamic, high-turbidity nearshore areas, including the breakwater zone.

[0047] By deeply fusing optical remote sensing (high spatial coverage) and acoustic profiles (high vertical resolution), and utilizing the physical interpolation capability of physical information neural networks, a suspended sediment data field integrating temporal attributes was generated, overcoming the limitations of spatial sparsity of in-situ sampling and temporal sparsity (cloud occlusion) of remote sensing.

[0048] By introducing a physical information neural network, the advection-diffusion equation and acoustic scattering equation are incorporated as strong constraints into the loss function, ensuring that the evolution of suspended sediment (such as the diffusion of sediment plumes) follows the basic physical laws of fluid dynamics and sediment transport. The results are physically interpretable and reliable. Attached Figure Description

[0049] Figure 1 This is a flowchart of the nearshore surface suspended sediment analysis method based on a deep learning model proposed in this invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0051] Reference Figure 1 This invention provides a method for analyzing nearshore surface suspended sediment based on a deep learning model.

[0052] S1. Acquire surface optical remote sensing data of the target water area, deploy an unmanned surface vessel equipped with an acoustic sensor to the target water area, and use the acoustic sensor to acquire acoustic backscatter data below the unmanned surface vessel's trajectory, wherein:

[0053] The acoustic sensor mounted on the unmanned vessel is preferably an arrayed multi-frequency acoustic Doppler current profiler (ADCP). This sensor is used to acquire water flow velocity profile data and multi-frequency backscatter intensity profile data.

[0054] In a preferred embodiment, in-situ sensors may also be deployed: a laser in-situ scattering and transmission instrument (for measuring particle size and concentration), an optical backscattering sensor (for measuring turbidity), or an automated pumping sampler (for acquiring physical water samples for laboratory analysis). The data collected by these sensors, as in-situ truth values, can be used for model training and validation.

[0055] S2. Preprocess and spatiotemporally align the acquired multi-source data to unify data from different sources, with different spatiotemporal resolutions, and different physical characteristics under a single framework. Specifically:

[0056] For acoustic data, the following is performed:

[0057] Velocity data processing: The velocity data measured by ADCP is subjected to quality control and coordinate transformation to drive the hydrodynamic model.

[0058] Backscatter data correction: Multi-step correction is performed on multi-frequency backscatter intensity data: First, noise correction is performed (removal of instrument background noise); second, sound wave propagation loss correction is performed, which includes geometric attenuation (as distance increases) and acoustic absorption and attenuation caused by water and suspended sediment.

[0059] Geo-registration: Geospatial registration of ADCP vertical profile data with geographic coordinates (e.g., via Global Navigation Satellite System GNSS).

[0060] For image data, perform:

[0061] Radiometric and atmospheric correction: Radiometric calibration is performed on remote sensing images, and atmospheric correction is carried out to eliminate the effects caused by scattering and absorption by atmospheric molecules, aerosols and thin clouds, thereby obtaining accurate water surface reflectance.

[0062] Data missing handling: For regions with missing data, interpolation algorithms can be used for reconstruction.

[0063] Geometric correction and synchronization: Perform geometric correction to ensure that the pixel coordinates of the image are precisely aligned with the geographic coordinates of the acoustic sensor.

[0064] Finally, spatiotemporal fusion is performed:

[0065] First, a unified nearshore 4D (x,y,z,t) spatiotemporal grid is defined.

[0066] Next, all preprocessed data undergoes spatiotemporal matching. This includes aligning the image data and ADCP data with the in-situ samples (“ground truth”) in both time and space.

[0067] Then, these spatiotemporally sparse and asynchronous data points are mapped onto this unified grid through interpolation or data assimilation to form a fused, sparse 4D data cube, which serves as the input for subsequent deep learning models.

[0068] S3. Input the generated data cube (as input to physical constraints) and the in-situ truth value (as input to data constraints) into the first deep learning model (preferably using a physical information neural network architecture in this embodiment). Train the model so that its output (suspended sediment concentration and bubble porosity) simultaneously satisfies both data constraints and physical constraints (fluid dynamics and acoustics).

[0069] The network core of the first deep learning model can be a fully connected network or a U-Net to achieve spatial field mapping.

[0070] The network receives the generated 4D data cube, along with the corresponding hydrodynamic parameters (from ADCP measurements or external hydrodynamic models) at the coordinate points within the data cube, and the corresponding multi-frequency acoustic backscattering intensity. It outputs the suspended sediment concentration and bubble porosity at each coordinate point.

[0071] The training process of this network is achieved by minimizing a hybrid loss function. This hybrid loss function is specifically:

[0072] ;

[0073] In the formula, For the total loss, For data loss, For advection-diffusion physical loss, For acoustic physical loss, and These are hyperparameters used to balance different loss terms.

[0074] Data loss: The mean square error between the network output's predicted suspended sediment concentration and the true value is calculated at all sparse measurement points with in-situ "true values". This loss term ensures that the model is accurate where data is available.

[0075] Advection-diffusion physical loss: The residuals of the advection-diffusion equation for suspended sediment concentration are calculated at a large number of "configuration points" (where no ground truth data is available) randomly sampled throughout the entire 4D data cube. This loss term forces the network to learn the physically (hydrodynamically) possible suspended sediment concentration field.

[0076] Acoustic physical loss: At all coordinate points with multi-frequency acoustic measurements (from ADCP), the physical information neural network outputs the predicted suspended sediment concentration and bubble porosity.

[0077] These two predicted values ​​are input into a known multi-frequency acoustic scattering model. The model calculates the acoustic backscattering intensity based on sediment concentration, particle size (which can be assumed or output by a physical information neural network), and bubble porosity.

[0078] The mean square error between the predicted acoustic intensity and the actual acoustic intensity measured by ADCP is calculated, which is the acoustic physical loss.

[0079] This loss term forces the network output of suspended sediment concentration and bubble porosity to be acoustically reasonable, meaning they must be able to “explain” the observed multi-frequency acoustic signals.

[0080] By simultaneously minimizing data loss, advection-diffusion physical loss, and acoustic physical loss, the physical information neural network architecture of this scheme can find a field that simultaneously satisfies three constraints: (suspended sediment concentration, bubble porosity); fits sparse measured data; follows fluid dynamics; and interprets the observed multi-frequency acoustic signals. This solves the problem of acoustic signals becoming ineffective due to bubble contamination in existing technologies.

[0081] S4. The suspended sediment distribution field output from S3 (e.g., data from the past 72 hours) is used as the input sequence and fed into the second deep learning network. This network predicts the spatiotemporal distribution of suspended sediment in the future (e.g., the next 24 hours) by learning historical evolution patterns.

[0082] For the second deep learning network, this embodiment preferably employs a hybrid architecture of U-Net and ConvLSTM. This architecture combines the advantages of U-Net in spatial feature extraction with the advantages of ConvLSTM in spatiotemporal sequence modeling.

[0083] The input to this network comes from the sequence of suspended sediment spatial distribution maps over the past N time steps, output from step S3. In this network:

[0084] U-Net Encoder: At each historical time step k, the suspended sediment concentration map is fed into the encoder (downsampling) part of U-Net. The encoder extracts the multi-scale spatial features of the suspended sediment concentration distribution at that moment (e.g., the shape, edges, and concentration gradient of the sediment plume) through a series of convolution and pooling operations.

[0085] ConvLSTM: The spatial feature map (a high-dimensional tensor sequence) extracted by the encoder is fed into one or more ConvLSTM layers. The core advantage of ConvLSTM lies in its convolutional gating units, which enable it to capture the temporal evolution patterns of these spatial features (e.g., how plumes spread and settle with tides). The final output of ConvLSTM is its prediction of the hidden state at future time steps, which encodes predictions of future spatial features.

[0086] U-Net Decoder: The predicted output of ConvLSTM is fed into the decoder (upsampling) part of U-Net. The decoder reconstructs the abstract, low-resolution future feature map into a high-resolution spatial distribution map of suspended sediment concentration at future time steps through a series of deconvolutions and skip connections.

[0087] This scheme first performs spatial dimensionality reduction and feature extraction using U-Net, then feeds these highly abstract features into ConvLSTM for temporal modeling, and finally restores the spatial resolution using the U-Net decoder. This cascaded "encoding-prediction-decoding" architecture has stronger spatiotemporal nonlinear modeling capabilities than using ConvLSTM alone to process the original pixels, and can more accurately predict the complex dynamic evolution of suspended sediment concentration fields.

[0088] In a preferred embodiment, the method further includes: performing self-supervised pre-training on the optical feature extraction network (e.g., the encoder of U-Net) in the scheme before initiating the main model training.

[0089] The pre-training dataset can be unlabeled remote sensing images from different time periods and regions. Specifically, pre-training can use mask reconstruction techniques:

[0090] Randomly obscure certain pixel blocks in the remote sensing image (turning them gray).

[0091] The optical feature extraction network is trained to reconstruct masked pixels using surrounding visible pixels (i.e., spectral and spatial context).

[0092] The loss function is the difference between the reconstructed pixel and the original pixel. Through this process, the optical feature extraction network is forced to learn general, robust features of water optics and spatial structure.

[0093] S5. Transform the outputs of steps S3 and S4 into decision support products for end users.

[0094] Specifically, it can be used for nearshore management such as port maintenance and dredging. Ports and waterways require regular dredging to maintain navigable depth, but dredging operations are costly and have environmental impacts. Traditional periodic dredging or responsive dredging leads to significant waste.

[0095] This approach can be used to first generate a high-precision historical siltation map of the waterway using the first deep learning model described above, identifying siltation hotspots. Then, the second deep learning module described above, based on historical data and hydrodynamic forecasts, generates a siltation prediction map for the next week or month.

[0096] Based on this forecast map, port managers have shifted from "passive responsive dredging" to "proactive predictive dredging," scheduling dredging operations only in areas and at times where siltation is predicted to exceed a threshold, significantly reducing costs.

[0097] The nearshore surface suspended sediment analysis system based on deep learning models is used to implement the aforementioned nearshore surface suspended sediment analysis method based on deep learning models. It includes a data acquisition module, a data preprocessing module, a data analysis module, and a visualization output module.

[0098] The data acquisition module is used to collect acoustic and optical data related to the target water area.

[0099] The data preprocessing module is used to preprocess and spatiotemporally align the acoustic and optical data acquired by the data acquisition module, so as to unify the data under a single framework.

[0100] The data analysis data is used to analyze the data processed by the data preprocessing module using a deep learning model.

[0101] The visualization output module is used to visualize the analysis results of the data analysis module for decision support.

[0102] The data processing module includes an acoustic data processing unit and an image data processing unit.

[0103] The acoustic data processing unit is used to perform backscatter data correction and georegistration on the acquired acoustic data.

[0104] The image data processing unit is used to perform radiometric correction, atmospheric correction, geometric correction, and synchronization on the acquired image data.

[0105] The data analysis module includes a first analysis and processing unit and a second analysis and processing unit.

[0106] The first analysis and processing unit is used to receive the data output by the data preprocessing module and use the first deep learning model to predict the distribution field of suspended sediment in the target water area.

[0107] The second analysis and processing unit is used to predict the future spatiotemporal distribution of suspended sediment in the target water area based on the suspended sediment distribution field output by the first analysis and processing unit and using a second deep learning model.

[0108] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0109] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for analyzing nearshore surface suspended sediment based on a deep learning model, characterized in that, Includes the following steps: Acquire surface optical remote sensing data of the target water area; Deploy an unmanned surface vessel equipped with acoustic sensors to the target waters and use the acoustic sensors to acquire acoustic backscatter data below the unmanned surface vessel's trajectory; The acquired surface optical remote sensing data and acoustic backscatter data are preprocessed and spatiotemporally aligned to generate a fused dataset; The fused dataset is input into the first deep learning model, and a high-resolution suspended sediment distribution field map of the target water area is output. The first deep learning model employs a physical information neural network architecture. This model is trained by minimizing a hybrid loss function, which is specifically: ； In the formula, For the total loss, For data loss, For advection-diffusion physical loss, For acoustic physical loss, and These are hyperparameters used to balance different loss terms; The data loss is used to ensure the accuracy of the model where data is available; the advection-diffusion physical loss is used to force the suspended sediment concentration field output by the model to be physically reasonable; and the acoustic physical loss is used to force the suspended sediment concentration and bubble porosity output by the model to be acoustically and physically reasonable. The specific method for determining the acoustic physical loss is as follows: The physical information output by the neural network at the coordinate points with acoustic measurement data is obtained, along with the concentration of suspended sediment and the void ratio of air bubbles. Based on the above suspended sediment concentration and bubble porosity, an acoustic backscattering intensity is calculated using a multi-frequency acoustic scattering model. The mean square error between the acoustic backscattering intensity and the measured acoustic backscattering intensity is the acoustic physical loss. Based on the high-resolution suspended sediment distribution field map of the target water area, the distribution of suspended sediment in the target water area at future times is predicted using a second deep learning model.

2. The method for analyzing nearshore surface suspended sediment based on a deep learning model according to claim 1, characterized in that, The acoustic backscattering data is multi-frequency acoustic backscattering data, which includes information on the scattering intensity at different acoustic frequencies.

3. The method for analyzing nearshore surface suspended sediment based on a deep learning model according to claim 1, characterized in that, The second deep learning model is a spatiotemporal prediction network, which includes convolutional long short-term memory units for predicting the spatiotemporal distribution of suspended sediment concentration in future time steps.

4. The nearshore surface suspended sediment analysis method based on a deep learning model according to claim 3, characterized in that, The spatiotemporal prediction network adopts an encoder-predictor architecture, and both the encoder and predictor are constructed from the convolutional long short-term memory units. The encoder and predictor are configured as follows: Encoder: Receives a sequence of spatial distribution maps of suspended sediment, processes the sequence through multi-layer convolutional long short-term memory units, extracts and encodes the multi-scale spatiotemporal features of the sequence, and generates a set of hidden states; Predictor: Using the hidden state generated by the encoder as the initial state, it runs iteratively for multiple time steps in the future. In each step, a convolutional long short-term memory unit is used to predict the hidden state of the next time step, and the spatial distribution map of suspended sediment in the future time step is reconstructed from the hidden state.

5. The nearshore surface suspended sediment analysis method based on a deep learning model according to claim 4, characterized in that, It also includes using mask reconstruction technology to perform self-supervised pre-training on the encoder of the spatiotemporal prediction network, specifically: Unlabeled remote sensing images from different time periods and regions are acquired as a pre-training dataset; Randomly mask certain pixel blocks in remote sensing images; The encoder of the spatiotemporal prediction network is trained to reconstruct the masked pixels using the surrounding visible pixels; The loss function is the difference between the reconstructed pixel and the original pixel.

6. A nearshore surface suspended sediment analysis system based on a deep learning model, used to implement the nearshore surface suspended sediment analysis method based on a deep learning model as described in any one of claims 1-5, characterized in that, It includes a data acquisition module, a data preprocessing module, a data analysis module, and a visualization output module; The data acquisition module is used to collect acoustic and optical data of the target water area using acoustic and optical sensors deployed on drones and unmanned vessels. The data preprocessing module is used to preprocess and spatiotemporally align the acoustic and optical data acquired by the data acquisition module, so as to unify the data under a single framework. The data analysis data is used to analyze the data processed by the data preprocessing module using a deep learning model; The visualization output module is used to visualize the analysis results of the data analysis module for decision support.

7. The nearshore surface suspended sediment analysis system based on a deep learning model according to claim 6, characterized in that, The data processing module includes an acoustic data processing unit and an image data processing unit; The acoustic data processing unit is used to perform backscatter data correction and georegistration on the collected acoustic data. The image data processing unit is used to perform radiometric correction, atmospheric correction, geometric correction, and synchronization on the acquired image data.

8. The nearshore surface suspended sediment analysis system based on a deep learning model according to claim 6, characterized in that, The data analysis module includes a first analysis and processing unit and a second analysis and processing unit. The first analysis and processing unit is used to receive the data output by the data preprocessing module and use the first deep learning model to predict the distribution field of suspended sediment in the target water area. The second analysis and processing unit is used to predict the future spatiotemporal distribution of suspended sediment in the target water area based on the suspended sediment distribution field output by the first analysis and processing unit and using a second deep learning model.

Images