CN-121298532-B - Near-shore surface layer suspended sediment analysis method based on deep learning model

CN121298532BCN 121298532 BCN121298532 BCN 121298532BCN-121298532-B

Abstract

The invention relates to the technical field of marine environment monitoring, in particular to a near-shore surface layer suspended sediment analysis method based on a deep learning model. The method comprises the steps of collecting surface optical remote sensing data and acoustic back scattering data of a target water area, preprocessing and time-space alignment of the collected surface optical remote sensing data and acoustic back scattering data to generate a fusion data set, inputting the fusion data set into a first deep learning model, predicting and outputting a suspended sediment distribution field diagram of the target water area, and predicting the suspended sediment distribution of the target water area at a future moment by utilizing a second deep learning model according to the suspended sediment distribution field diagram of the target water area. The invention finishes decoupling of the bubble and sediment signals through the physical information neural network architecture of double physical constraints, so that the invention realizes higher inversion precision in a high-dynamic and high-turbidity offshore area comprising the crushed wave bands, and remarkably improves the accuracy and space-time resolution of offshore suspended sediment analysis.

Inventors

CHEN PEIXIONG
ZHOU XIN
ZHANG MIAO

Assignees

自然资源部第二海洋研究所

Dates

Publication Date: 20260512
Application Date: 20251211

Claims (8)

1. The method for analyzing the suspended sediment on the near-shore surface layer based on the deep learning model is characterized by comprising the following steps of: Acquiring surface optical remote sensing data of a target water area; deploying an unmanned ship carrying an acoustic sensor to the target water area, and acquiring acoustic back scattering data below the unmanned ship track by using the acoustic sensor; Preprocessing and space-time alignment are carried out on the acquired surface optical remote sensing data and acoustic back scattering data, and a fusion data set is generated; inputting the fusion data set into a first deep learning model, and outputting to obtain a high-resolution suspended sediment distribution field diagram of a target water area; The first deep learning model adopts a physical information neural network architecture, the model is trained by minimizing a mixed loss function, and the mixed loss function is specifically: ; In the formula, In order to account for the total loss, The data is lost and, in response to the loss of data, In order to achieve a advection-diffusion physical loss, In order to be an acoustic physical loss, And Is a super parameter for balancing different loss terms; The data loss is used for ensuring the accuracy of the model in the place with the data, the advection-diffusion physical loss is used for forcing the suspended sediment concentration field output by the model to be physically reasonable, and the acoustic physical loss is used for forcing the suspended sediment concentration and the bubble void ratio output by the model to be acoustically reasonable; the specific determination mode of the acoustic physical loss is as follows: acquiring suspended sediment concentration and bubble void ratio output by the physical information neural network on a coordinate point with acoustic measurement data; Calculating the acoustic back scattering intensity by using a multi-frequency acoustic scattering model based on the suspended sediment concentration and the bubble void ratio; Calculating the mean square error between the acoustic back scattering intensity and the actually measured acoustic back scattering intensity, namely the acoustic physical loss; and predicting and obtaining the suspended sediment distribution of the target water area at the future moment by using a second deep learning model according to the high-resolution suspended sediment distribution field diagram of the target water area.
2. The method for analyzing suspended sediment on a near-shore surface layer based on a deep learning model according to claim 1, wherein the acoustic back-scattering data is multi-frequency acoustic back-scattering data, and the data comprises scattering intensity information of different acoustic frequencies.
3. The method for analyzing suspended sediment on a near-shore surface layer based on a deep learning model according to claim 1, wherein the second deep learning model is a space-time prediction network, and the space-time prediction network comprises a convolution long-short-term memory unit for predicting the space-time distribution of suspended sediment concentration of a future time step.
4. A method of analyzing suspended sediment in a near shore surface layer based on a deep learning model according to claim 3, wherein the space-time prediction network adopts an encoder-predictor architecture, and the encoder and predictor are both constructed by the convolution long-short term memory unit, and the encoder and predictor are configured to: the encoder is used for receiving the suspended sediment space distribution diagram sequence, processing the sequence through a plurality of layers of convolution long-short-term memory units, extracting and encoding multi-scale space-time characteristics of the sequence, and generating a group of hidden states; And the predictor is used for iteratively running in a plurality of time steps in the future by taking the hidden state generated by the encoder as an initial state, wherein each step uses a convolution long-short-term memory unit to predict the hidden state of the next time step, and a suspended sediment space distribution diagram of the future time step is reconstructed from the hidden state.
5. The method for analysis of suspended sediment in a near shore surface layer based on a deep learning model according to claim 4, further comprising self-supervised pre-training of the encoder of the spatio-temporal prediction network using a mask reconstruction technique, in particular: obtaining unlabeled remote sensing images from different time phases and different areas as a pre-training data set; randomly masking some pixel blocks of the remote sensing image; An encoder that trains the spatio-temporal prediction network reconstructs masked pixels using surrounding visible pixels; wherein the loss function is the difference between the reconstructed pixel and the original pixel.
6. A near-shore surface layer suspended sediment analysis system based on a deep learning model, which is used for realizing the near-shore surface layer suspended sediment analysis method based on the deep learning model as claimed in any one of claims 1-5, and is characterized by comprising a data acquisition module, a data preprocessing module, a data analysis module and a visual output module; the data acquisition module is used for acquiring acoustic related data and optical related data of a target water area by utilizing acoustic sensors and optical sensors deployed on the unmanned aerial vehicle and the unmanned ship; The data preprocessing module is used for preprocessing and space-time aligning the acoustic related data and the optical related data acquired by the data acquisition module so as to unify the data under one frame; the data analysis data are used for analyzing the data processed by the data preprocessing module by using a deep learning model; the visual output module is used for visually outputting the analysis result of the data analysis module for decision support.
7. The deep learning model-based near shore surface layer suspended sediment analysis system of claim 6, wherein the data processing module comprises an acoustic data processing unit and an image data processing unit; The acoustic data processing unit is used for carrying out backscattering data correction and geographic registration on the acquired acoustic data; The image data processing unit is used for carrying out radiation correction, atmosphere correction, geometric correction and synchronization on the acquired image data.
8. The deep learning model-based near shore surface layer suspended sediment analysis system of claim 6, wherein the data analysis module comprises a first analysis processing unit and a second analysis processing unit; The first analysis processing unit is used for receiving the data output by the data preprocessing module and predicting a suspended sediment distribution field of the output target water area by using a first deep learning model; the second analysis processing unit is used for predicting future space-time distribution of suspended sediment in the target water area by using a second deep learning model based on the suspended sediment distribution field output by the first analysis processing unit.

Description

Near-shore surface layer suspended sediment analysis method based on deep learning model Technical Field The invention relates to the technical field of marine environment monitoring, in particular to a near-shore surface layer suspended sediment analysis method based on a deep learning model. Background Suspended sediment is a critical hydrodynamic and water quality parameter in offshore bodies of water, including estuaries, delta and coastal areas. Dynamic change and transportation process of suspended sediment are not only main driving forces for landform evolution (such as coastal erosion and delta development), but also have profound effects on aquatic ecosystems and human economic activities. In industrial application, the accurate monitoring of suspended sediment is important. First, it is the core basis for port maintenance and channel dredging. The sediment accumulation in the channel and harbor pool can reduce the navigation water depth and threaten the navigation safety, so periodic dredging is needed. The cost of dredging operations is extremely high, and accurate suspended sediment monitoring and silting prediction are prerequisites for optimizing the dredging plan, reducing maintenance costs and assessing environmental impact during dredging. In addition, the high-concentration suspended sediment can reduce the transparency of the water body, influence photosynthesis, cause pressure on sensitive marine ecosystems such as coral reefs, seaweed beds and the like, and possibly adsorb and transport pollutants such as heavy metals and the like. Therefore, suspended sediment monitoring is of great significance to marine environmental protection. However, the coastal zone is one of the most hydrodynamic regions of the earth, and suspended sediment is subject to multiple and nonlinear effects of tides, waves (especially detritus), turbulence, river inputs, and human activity, exhibiting extremely high space-time variability. This makes accurate, high resolution measurement of SSC a significant technical challenge. To address this challenge, the prior art developed three general types of monitoring methods, in-situ sampling, acoustic proxy, and optical telemetry. In-situ sampling schemes mainly use physical pump sampling, optical backscatter sensors, and laser in-situ scattering and transmission instruments. These instruments determine suspended sediment by directly measuring the physical or optical properties of the water sample. But these methods are essentially "point measurements". While they can provide high-precision time series data at the deployment site, they do not provide detailed spatial and temporal profiles of suspended sediment. They cannot capture a wide range of spatially distributed features (e.g., a complete silt plume). Acoustic proxy solutions mainly use Acoustic Doppler Current Profiler (ADCP). These instruments indirectly infer suspended sediment by emitting sound waves and measuring the signal strength (i.e., acoustic backscatter) of particles (sediment) scattered back from the body of water. An advantage of ADCP is that it provides both a water flow profile and a suspended sediment profile of the entire water column. But it is difficult to distinguish between bubbles and silt by acoustic means. Near shore wave breaking can instantaneously produce massive microbubbles in the body of water, which are also highly efficient acoustic scatterers that produce acoustic signals that severely overlap with sediment signals (particularly fine particle sediment). Existing acoustic algorithms perform well only under non-detritus conditions, but fail completely under detritus conditions because they cannot distinguish sediment from the scattered signal of bubbles. The optical remote sensing scheme utilizes an optical sensor carried by a satellite or an unmanned aerial vehicle to capture the reflectivity of the surface of a water body and establishes a relation model of the reflectivity and surface layer suspended sediment. However, the physical characteristics of optical remote sensing determine that it can only measure "surface layer" suspended sediment (light penetration depth) of several centimeters to several meters below the water surface, and cannot provide vertical structural information of the water column. Disclosure of Invention In order to solve the defects in the prior art, the invention provides a near-shore surface layer suspended sediment analysis method based on a deep learning model. The invention solves the problems existing in the prior art cooperatively by a systematic fusion method by utilizing the complementary advantages of different data sources (acoustics and optics) and the unique capabilities of different deep learning architectures. A near shore surface layer suspended sediment analysis method based on a deep learning model comprises the following steps: Acquiring surface optical remote sensing data of a target water area; deploying an unmanned ship carrying an acou