Coral reef beach sand evolution process twinborn intelligent simulation method and device

Abstract

The invention provides an intelligent twinborn simulation method and device for a coral reef beach sand evolution process, relates to the field of high-dimensional space-time big data and artificial intelligence, and solves the technical problem of continuous and spatial refined simulation of the coral reef beach sand evolution process. The method comprises the following steps: acquiring original data of process elements influencing beach sand evolution, wherein the selection of the process elements considers main natural and artificial regional environment characteristics driving beach sand evolution; preprocessing the original data, and constructing a beach sand state sequence training data set in a multi-dimensional space cube form; selecting a SwinLSTM deep learning framework which is suitable for beach sand evolution simulation and has a space-time dependence extraction capability, fusing a cross attention mechanism, designing a loss function for beach sand form change characteristics, constructing a coral reef beach sand evolution process twinborn intelligent simulation model, and completing training, verification and testing of the model by using a training data set; the simulation model is applied to realize approximate simulation of evolution process and terrain of coral reef beach sand in a historical or future specified time period.

Description

Technical Field

[0001] The present invention relates to the field of high-dimensional spatiotemporal big data and artificial intelligence technology, and more specifically to a twin intelligent simulation method and device for the evolution process of coral reef beach sand. Background Art

[0002] As a key landform type in the long-term evolution of coral reefs, offshore coral reef beach sand (offshore coral reef beach sand refers to the sandy tidal flats outside the coral reefs affected by natural and human environmental dynamics, as well as the sandbars that are the first landform type before the stable emergence of coral islands) is deeply affected by the natural and human environmental dynamics. These sandy tidal flats and sandbars in front of the stable emergence of coral islands are not only an important part of the coral reef ecosystem, but their evolution process is also a direct reflection of the health and stability of coral reefs under the background of global climate change. Therefore, the simulation of the evolution process of coral reef beach sand has far-reaching significance for the rational development and protection of coral reefs and the maintenance of the health of marine ecosystems.

[0003] With the rapid development of remote sensing technology, researchers have begun to extract the exposed parts of coral reef beach sand from long-term remote sensing images, analyze the temporal changes of their area and circumference, and calculate the movement of the center of mass and its trajectory to gain insight into the dynamic evolution of beach sand. This process mostly uses statistical analysis methods to summarize and reveal the inherent laws of beach sand evolution. However, the conclusions obtained by combining the long-term observations and professional knowledge of experts in the field of marine science have revealed the main factors driving the evolution of coral reef beach sand to a certain extent, and have conducted a qualitative analysis of its evolution process. However, due to the limitations of the availability of high-resolution images in coral reef areas and the impact of sea clouds and rain on the continuity of historical images, the continuity of the evolution process is insufficient, which reduces the accuracy of the evolution analysis. At the same time, the inference of future evolution is often based on theoretical inferences supplemented by the calculation conclusions of shape statistical analysis, which is difficult to meet the urgent need for spatially fine simulation of beach sand evolution.

[0004] At present, although numerical simulation models such as the two-dimensional planar coastal dynamics numerical model (X-Beach) based on the structured Fortran 77 / 90 architecture and the three-dimensional hydrodynamic-water quality model system (Delft 3D) have been applied to the evolution simulation of continental coastal beaches, their application on offshore coral reef beaches faces many challenges. These models rely on preset model structures and parameters and must meet strict assumptions. However, the evolution process of real coral reef beaches is extremely complex and has highly nonlinear characteristics, which does not fully conform to the preset assumptions. Once the actual situation does not match the model assumptions, or the data characteristics change, the numerical simulation model will be difficult to adjust automatically, resulting in deviations between the simulation results and the actual evolution, and the accuracy is greatly reduced. In addition, these models also require a large number of precise continuous field measurements to obtain the specified variables in the calculation formula, which is difficult to achieve on offshore coral reefs and has high measurement costs.

[0005] In contrast, deep learning methods have significant advantages in capturing the spatiotemporal dependency characteristics between data due to their powerful data-driven and knowledge extraction capabilities. Existing studies have combined convolutional neural networks (CNN) with recurrent neural networks (RNN) to successfully learn the spatiotemporal dependencies in spatiotemporal data; convolutional LSTM (Convolutional Long Short-Term Memory, ConvLSTM) further improves the simulation accuracy by extending the fully connected LSTM and replacing linear operations with convolution operations. Recently, the introduction of the Vision Transformer model has provided a new perspective for capturing global dependencies. On this basis, the SwinLSTM deep learning framework came into being, which combines the advantages of LSTM and Swin Transformer, can capture global and local spatiotemporal dependencies, and effectively reduce the computational burden. This framework has been successfully applied in fields such as the MNIST digital dynamic change dataset (MNIST is a proper noun representing a widely used handwritten digital dataset). However, its application in beach sand evolution simulation has rarely been studied.

[0006] In particular, the evolution process of coral reef beach sand is unique: multiple natural and human environmental dynamic factors are intertwined, and it is necessary to extract and calculate key spatiotemporal change information from a large amount of spatiotemporal data through the intelligent methods of geographic information system (GIS) and remote sensing to characterize the beach sand evolution process, and design modules to combine multi-dimensional spatiotemporal features and input them into the model, which is an important prerequisite for model simulation; coral reef beach sand has unique geographical characteristics, and it is necessary to select reasonable variables according to the characteristics to characterize the evolution process for more accurate simulation inference, such as the periodic prevailing monsoon makes the evolution of coral reef beach sand present a certain periodic law, and in the future scenario, human maritime shipping activities will affect the development and accumulation rate of coral reefs, and then affect their growth and evolution process; the outer changes of coral reef beach sand are relatively large, and the changes of the corresponding grids need to be more focused during simulation. Therefore, it is necessary to select a deep learning framework with strong spatiotemporal dependency extraction capability suitable for beach sand evolution simulation problems, and design special models and methods for these characteristics on its basis, so as to effectively solve the twin simulation problem of the evolution process of offshore coral reef beach sand.

[0007] The expert experience inference method based on traditional statistics relies on the continuity and availability of image data, the conclusions are not quantitative enough and it is difficult to meet the requirements of spatial fine-grained simulation. The numerical simulation method requires a large amount of measured data that fits the model formula variables and is difficult to adapt to complex nonlinear beach sand evolution conditions. For the geographical process of coral reef beach sand evolution, the existing deep learning methods have insufficient feature extraction capabilities for complex environmental dynamic mechanisms (which need to be connected with remote sensing big data cloud computing and GIS methods), fail to accurately characterize the unique characteristics of geographical objects such as their periodicity and the impact of human activities, and have limited ability to capture local areas of drastic changes.

[0008] In view of this, we urgently need an innovative twin intelligent simulation method for the evolution process of offshore coral reef beach sands, so as to conduct a more accurate twin simulation of the evolution process of coral reef beach sands in history and future specified time periods, and provide strong support for the ecological protection and sustainable development of coral reefs. Summary of the invention

[0009] In view of the above problems, the present invention provides a twin intelligent simulation method and device for the evolution process of coral reef beaches. The method integrates the cross attention mechanism, uses the SwinLSTM framework that combines Swin Transformer and LSTM, builds a deep learning model, designs a loss function that focuses on grids with large changes, realizes the twin intelligent simulation of the evolution process of offshore coral reef beaches, and simulates the probability of coral reef beach exposure and approximate terrain in a specified time period.

[0010] Coral reef beach sand refers to the sandy tidal flats outside the coral reef that are affected by natural and human environmental dynamics, as well as the sandbars that serve as a landform type before the stable exposed coral islands.

[0011] Swin Transformer refers to a visual feature extraction model based on the Transformer architecture, which is specifically used to process two-dimensional spatial data such as images. LSTM (Long Short-Term Memory Network) is a recurrent neural network variant for sequence data processing that can model temporal dependencies. SwinLSTM refers to a hybrid neural network architecture that combines the spatial modeling capabilities of Swin Transformer and the temporal modeling mechanism of LSTM, and is suitable for cross-modal data processing scenarios that contain both spatial information and temporal changes.

[0012] According to a first aspect of the present invention, a twin intelligent simulation method for the evolution process of coral reef beach sand is provided, comprising: obtaining original data of process elements affecting beach sand evolution, wherein the process elements include the main natural and artificial regional environmental characteristics driving beach sand evolution and the exposed morphological characteristics of the beach sand itself; preprocessing the original data to construct a beach sand state sequence training data set in the form of a multi-dimensional space cube; selecting a SwinLSTM deep learning framework with spatiotemporal dependency extraction capability suitable for beach sand evolution simulation, integrating a cross-attention mechanism, designing a loss function for the beach sand morphological change characteristics, and constructing a twin intelligent simulation model for the evolution process of coral reef beach sand; using the training data set to complete the training, verification and testing of the twin intelligent simulation model for the evolution process of coral reef beach sand, and obtaining a target twin intelligent simulation model for the evolution process of coral reef beach sand; applying the target twin intelligent simulation model for the evolution process of coral reef beach sand to realize an approximate simulation of the evolution process and terrain of coral reef beach sand in a specified period of time in the past or in the future; wherein the twin intelligent simulation model for the evolution process of coral reef beach sand includes a beach sand feature extraction module, a spatiotemporal dependency capture module and a twin simulation module, wherein the beach sand feature extraction module adopts a cross-attention mechanism; the spatiotemporal dependency capture module is composed of a Swin It consists of two parts, Transformer and LSTM. The SwinTransformer part is used to vertically learn global spatial dependencies; the LSTM part is used to capture long-term and short-term temporal dependencies by horizontally updating the information of cell states and hidden states.

[0013] In some exemplary embodiments, the target coral reef beach sand evolution twin intelligent simulation model is applied to realize the approximate simulation of the evolution process and terrain of the coral reef beach sand in the past and the future specified time period, including: inputting the data to be simulated into the target coral reef beach sand evolution twin intelligent simulation model; utilizing the beach sand feature extraction module in the target coral reef beach sand evolution twin intelligent simulation model, and adopting the cross attention mechanism to obtain the initial input of the coral reef beach sand state on each time slice; dividing the initial input into small data blocks of uniform size and non-overlapping, flattening the small data blocks in order from left to right and from top to bottom, and inputting The spatial-temporal dependency capture module is input; the global spatial dependency of the initial input is obtained by using the SwinTransformer part in the spatial-temporal dependency capture module; the long-term and short-term time dependencies of the initial input are obtained by using the LSTM part in the spatial-temporal dependency capture module; based on the spatial dependency and time dependency, the twin simulation module is used to twin the exposure probability of the coral reef beach sand at a specified time of evolution; and based on the exposure probability, the evolution process and topography of the coral reef beach sand in the history or future specified time period are approximated, wherein the data type of the data to be simulated is the same as that of the training data in the training data set.

[0014] In some exemplary embodiments, if the input data to be simulated is a data set of beach sand state elements that characterizes the evolution of coral reef beach sand in a certain period of history, the result of the coral reef beach sand evolution simulation is the evolution process of the specified historical period; if the input data to be simulated is a data set of beach sand state elements that characterizes the evolution of coral reef beach sand in a recent period of time, the result of the coral reef beach sand evolution simulation is the evolution process of the specified future period.

[0015] In some exemplary embodiments, the original data includes remote sensing images, ship automatic identification system data, natural environment dynamic factor data, and monsoon period attributes; the beach sand state factor data set includes beach sand state factor data organized in the form of spatial cubes under multiple time slices, covering endogenous variables and exogenous variables, wherein endogenous variables include exposure probability; and exogenous variables include natural environment dynamic factor data, artificial environment dynamic factor data, extreme weather data such as storm surges, and monsoon period attributes.

[0016] In some exemplary embodiments, a cross-attention mechanism is used to obtain the initial input of the coral reef beach sand state at each time slice, using endogenous variables as queries and exogenous variables as keys and values.

[0017] In some exemplary embodiments, preprocessing the original data to construct a beach sand state sequence training data set in the form of a multi-dimensional space cube includes: extracting the original data of indirect elements from the original remote sensing images and ship automatic identification system data, wherein the indirect element data include the probability of beach sand exposure, shipping intensity and ship density, shoreline artificialization rate and typhoon intensity; setting a grid unit framework of a specified spatial resolution and geographic coordinate system, and unifying the indirect element data and the natural environment dynamic element data that can be directly obtained under the grid unit framework; according to the calculation target of the indirect elements, using the geographic information system spatial analysis calculation method and remote sensing big data cloud batch processing technology to complete batch calculations and obtain the final calculation results of the indirect elements. Use one-hot encoding to convert the monsoon period attributes into vectors to obtain the monsoon period attributes represented by the vectors; the final calculation results of the indirect factors, the natural environment dynamic factors that can be directly obtained, and the monsoon period attributes represented by the vectors are integrated to form beach sand status element data in the form of a multidimensional space cube, and each time slice includes a multidimensional space cube; wherein the beach sand status element data set contains beach sand status element data organized in the form of space cubes under multiple time slices.

[0018] In some exemplary embodiments, using a training data set to complete the training, verification, and testing of a twin intelligent simulation model of the coral reef beach sand evolution process includes: dividing the training data set into a training set, a verification set, and a test set according to a preset ratio, wherein each sample in the training set, the verification set, and the test set includes a multidimensional space cube on T+L time slices, wherein T represents the number of front time slices, which is used to extract the spatiotemporal features characterizing the beach sand evolution process, and L represents the number of rear time slices, which is required through T The goal is to simulate the evolution of beach sand in time slices, T is numbered from 1, that is, the front time slices are numbered as: 1, 2, ..., T-1, T; L is numbered from T+1; that is, the rear time slices are numbered as: T+1, T+2, ..., L-1, L; T and L are both positive integers; based on the state element data of coral reef beach sand organized in the form of spatial cubes on T time slices in the training set, the twin intelligent simulation model of the coral reef beach sand evolution process is used to extract the beach sand evolution characteristics, and the simulation results on the T+1-T+L time slices are obtained; based on the simulation results and sample labels, the loss value is calculated using the loss function; back propagation is performed based on the loss value to update the model weights, and the simulation results are obtained in the middle The twin intelligent simulation model of the evolution process of intermediate coral reef beaches is constructed; the early stopping mechanism is adopted to calculate the loss value of the twin intelligent simulation model of the evolution process of intermediate coral reef beaches using the validation set data, until the loss value of the validation set changes within the preset range, the training is stopped, and the model with the highest accuracy on the validation set during the training process is saved to obtain the trained twin intelligent simulation model of the evolution process of coral reef beaches, so as to avoid overfitting; the preset range refers to setting the number of training rounds to N. When the loss value of the validation set is still rising after N rounds, the training is no longer continued, and the model with the highest accuracy on the validation set is saved, where N is a positive integer; the sample label is the actual exposure probability grid of the coral reef beaches at the T+1-T+L time slice.

[0019] In some exemplary embodiments, it also includes: in multiple rounds of training, automatically adjusting the global hyperparameters of the learning rate through a particle swarm optimization algorithm, using the Adam algorithm to update local network weights, automatically training multiple twin intelligent simulation models of the coral reef beach sand evolution process, and selecting the twin intelligent simulation model of the coral reef beach sand evolution process with the smallest loss value on the validation set as the target twin intelligent simulation model of the coral reef beach sand evolution process.

[0020] In some exemplary embodiments, the loss function is:

[0021]

[0022] Among them, Loss is the loss value, is the true value on grid (i, j), is the simulated value on grid (i,j), is the weight adjustment value on grid (i, j), and K is the total number of grids in the beach sand simulation area;

[0023] Adjust as follows:

[0024] Calculate {X for each grid over the entire previous time series 1 , X 2 ,…, X T Cumulative change C ij :

[0025]

[0026] The high-frequency changing beach sand boundary area will produce a larger accumulation value.

[0027] Normalize the accumulated value to the preset weight range [δ,γ]:

[0028]

[0029] Through time-series dynamic analysis, the sensitive areas of beach sand morphology change are automatically identified, including erosion / deposition boundaries, and higher loss weights are assigned to the sensitive areas, so that the model training focuses on the spatial locations where the terrain fluctuates violently.

[0030] According to a second aspect of the present invention, a twin intelligent simulation device for the evolution process of coral reef beach sand is provided, including: an acquisition module for acquiring raw data of process factors affecting beach sand evolution, wherein the process factors include the main natural and artificial regional environmental characteristics that drive beach sand evolution and the exposed morphological characteristics of the beach sand itself; a preprocessing module for preprocessing the raw data and constructing a beach sand state sequence training data set in the form of a multi-dimensional space cube; a model construction module for selecting a SwinLSTM deep learning framework with spatiotemporal dependency extraction capability suitable for beach sand evolution simulation, integrating a cross-attention mechanism, designing a loss function for beach sand morphological change characteristics, and constructing a coral reef beach sand state sequence training data set; and a model construction module for selecting a SwinLSTM deep learning framework with spatiotemporal dependency extraction capability suitable for beach sand evolution simulation, integrating a cross-attention mechanism, and designing a loss function for beach sand morphological change characteristics. A twin intelligent simulation model of the coral reef beach sand evolution process; a model training module, used to complete the training, verification and testing of the twin intelligent simulation model of the coral reef beach sand evolution process using the training data set, and obtain the target coral reef beach sand evolution process twin intelligent simulation model; a model application module, used to apply the target coral reef beach sand evolution process twin intelligent simulation model to achieve an approximate simulation of the evolution process and topography of the coral reef beach sand in history or in a specified time period in the future; wherein, the twin intelligent simulation model of the coral reef beach sand evolution process includes a beach sand feature extraction module, a spatiotemporal dependency capture module and a twin simulation module, wherein the beach sand feature extraction module adopts a cross-attention mechanism; the spatiotemporal dependency capture module consists of two parts, Swin Transformer and LSTM, the Swin Transformer part is used to vertically learn global spatial dependencies; the LSTM part is used to capture long-term and short-term temporal dependencies by horizontally updating the information of cell states and hidden states. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The above contents and other objects, features and advantages of the present invention will become more apparent through the following description of the embodiments of the present invention with reference to the accompanying drawings, in which:

[0032] Figure 1 A flowchart of a twin intelligent simulation method for the coral reef beach sand evolution process according to an embodiment of the present invention is schematically shown.

[0033] Figure 2 The flowchart of preprocessing raw data according to an embodiment of the present invention is schematically shown.

[0034] Figure 3 The schematic diagram shows the probability of beach sand exposure in the same area under different monsoon period attributes.

[0035] Figure 4 The schematic diagram shows the architecture overview of the twin intelligent simulation model of the coral reef beach sand evolution process according to an embodiment of the present invention.

[0036] Figure 5 The schematic diagram shows the probability of beach sand exposure in a certain area in units of months.

[0037] Figure 6 A twin intelligent simulation device for the coral reef beach sand evolution process according to an embodiment of the present invention is schematically shown. DETAILED DESCRIPTION

[0038] Below, embodiments of the present invention will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the present invention. In the following detailed description, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of embodiments of the present invention. However, it is apparent that one or more embodiments may also be implemented without these specific details. In addition, in the following description, descriptions of known structures and technologies are omitted to avoid unnecessary confusion of concepts of the present invention.

[0039] The terms used herein are only for describing specific embodiments and are not intended to limit the present invention. The terms "comprise", "include", etc. used herein indicate the existence of features, steps, operations and / or components, but do not exclude the existence or addition of one or more other features, steps, operations or components.

[0040] All terms (including technical and scientific terms) used herein have the meanings commonly understood by those skilled in the art unless otherwise defined. It should be noted that the terms used herein should be interpreted as having a meaning consistent with the context of this specification and should not be interpreted in an idealized or overly rigid manner.

[0041] When using expressions such as "at least one of A, B, and C, etc.", they should generally be interpreted according to the meaning of the expression commonly understood by those skilled in the art (for example, "a system having at least one of A, B, and C" should include but is not limited to a system having A alone, B alone, C alone, A and B, A and C, B and C, and / or A, B, C, etc.).

[0042] Figure 1 A flowchart of a twin intelligent simulation method for the coral reef beach sand evolution process according to an embodiment of the present invention is schematically shown.

[0043] like Figure 1 As shown, a twin intelligent simulation method for the coral reef beach sand evolution process according to an embodiment of the present invention includes steps S110 to S150.

[0044] In step S110, original data of process factors affecting beach sand evolution are obtained, wherein the process factors include the main natural and artificial regional environmental characteristics driving beach sand evolution and the exposure morphological characteristics of the beach sand itself.

[0045] Optionally, the original data includes remote sensing images, automatic ship identification system data, natural environment dynamic factor data and monsoon period attributes.

[0046] In step S120, the original data is preprocessed to construct a beach sand state sequence training data set in the form of a multi-dimensional space cube.

[0047] In some exemplary embodiments, the beach sand state sequence training data set includes endogenous variables and exogenous variables, wherein the endogenous variables include exposure probability; and the exogenous variables include natural environment dynamic element data, man-made environment dynamic element data, extreme weather data such as storm surges, and monsoon period attributes.

[0048] The exposure probability characterizes the exposure state of beach sand at each spatial position on the time slice and is a basic endogenous variable. The dynamic factors of the natural environment mainly include ocean hydrodynamic factors, sea breeze, extreme weather disturbances and periodic monsoons. Offshore coral reefs evolve mainly under the driving force of the seawater dynamics of ocean currents, waves and tides and sea breezes, which transport coral sand and affect the exposure and morphology of coral reef beach sand. Typhoons are the main extreme weather at sea that affects the evolution of beach sand; areas where coral reefs are located often have obvious periodic monsoon characteristics, which makes the evolution of coral reefs also periodic. In addition, due to the intervention of human activities, the physical morphology of the coral reef beach sand coastline has been significantly changed. Human activities such as shipping will also affect the health of the coral reefs themselves, change the accumulation rate, and thus affect the evolution of beach sand.

[0049] In an embodiment of the present invention, the state sequence training data set of the main process of coral reef beach sand evolution selects exposure probability data characterizing the actual exposure state of beach sand, basic element data of natural environment dynamics (sea breeze, ocean current, ocean wave, tide level), extreme weather data such as storm surge (long-term typhoon track frequency), monsoon period attributes (northeast monsoon period, southwest monsoon period or transition period), and artificial environment dynamic element data (shipping intensity, ship density, shoreline artificialization rate) for fusion construction.

[0050] In some exemplary embodiments, step S120 includes steps S121 to S125, see Figure 2 .

[0051] In step S121, the original data of indirect elements are extracted from the original remote sensing images and the ship automatic identification system data, wherein the indirect element data include the probability of beach sand exposure, shipping intensity and ship density, shoreline artificialization rate and typhoon intensity. After the original data of the indirect elements are extracted, the final calculation results of the indirect elements are obtained using the geographic information system (GIS) spatial analysis calculation method and the remote sensing cloud computing platform. The GIS spatial analysis calculation method includes kernel density calculation, raster operation, raster reclassification, etc.

[0052] For example, to obtain exposure probability data, a multi-band remote sensing image set with a spatial resolution of 10m*10m of the target coral reef beach sand is obtained on a cloud server in units of months, and cloud cover screening conditions and spatial range (a rectangle containing all possible exposure locations in the long time series of beach sand) are set. For a single-view image, if each pixel in the image is an exposed beach sand, it is assigned a value of 0, otherwise it is assigned a value of 1. The exposure probability grid of the month is obtained based on the pixel-by-pixel statistics of all exposed binary grids of the month. The specific calculation formula for the exposure probability data P is as follows:

[0053] (1)

[0054] Among them, n is the number of computable grids that meet the cloud cover conditions in a month, and p is the value assigned to the grid of a single image.

[0055] Using ship AIS data, the number of ships in a unit grid is calculated on a monthly basis, and the natural break method is used to reclassify the results to reflect the shipping intensity.

[0056] Using ship AIS data, the spatial analysis method of kernel density is used to calculate the ship density.

[0057] If the beach sand is a sandy tidal flat outside the coral reef affected by natural and human environmental dynamics, the islands and reefs where it is located may have artificial shore sections. Calculate the artificiality rate of the coral reef coastline where it is located. Extract the artificial coastline and the entire coral reef coastline and calculate their lengths respectively. The specific calculation formula for the coastline artificiality rate V is as follows:

[0058] (2)

[0059] in, is the total length of the artificial coastline of the coral islands and reefs where the beach sand is located. It is the total length of the coral island reef coastline where the beach sand is located.

[0060] If the beach sand is a stable sandbar of the previous landform type exposed in front of the coral island, the artificialization rate of the shoreline is 0.

[0061] Using typhoon track data, the kernel density spatial analysis method was used to conduct density cluster analysis on typhoon track points on a monthly basis to obtain the analysis results of the typhoon intensity in the area.

[0062] In step S122, a grid unit framework of a specified spatial resolution and geographic coordinate system is set, and the indirect element data and the natural environment dynamic element data that can be directly obtained are unified under the grid unit framework.

[0063] In the embodiment of the present invention, the indirect element data includes the element data calculated in step S121, and the spatial range of all data is consistent with the exposure probability calculation range, which is the designated coral reef beach sand simulation range.

[0064] In step S123, according to the calculation target of the indirect elements, the spatial analysis calculation method of the geographic information system and the cloud batch processing technology of remote sensing big data are used to complete the batch calculation and obtain the final calculation result of the indirect elements.

[0065] In step S124, the monsoon period attribute is converted into a vector using one-hot encoding to obtain the monsoon period attribute represented by the vector.

[0066] Figure 3 The schematic diagram shows the probability of beach sand exposure in the same area under different monsoon attributes, where Figure 3 (a) is the transitional non-monsoon period; Figure 3 (b) is the southwest monsoon period; Figure 3 (c) is the transitional non-monsoon period and Figure 3 (d) is the northeast monsoon period. Figure 3 From (a), (b), (c) and (d) in the figure, we can see that the probability of beach sand exposure varies under different monsoon period attributes. Therefore, it is necessary to consider the monsoon period attribute as an important influencing factor and use one-hot encoding to convert the monsoon period attribute into a vector to obtain the monsoon period attribute represented by the vector.

[0067] In step S125, the final calculation results of the indirect factors, the natural environment dynamic factors that can be directly obtained, and the monsoon period attributes represented by the vector are integrated to form beach sand status element data in the form of a multidimensional space cube. Each time slice includes a multidimensional space cube that represents the state of the beach sand at that time. The beach sand status element data set contains beach sand status element data organized in the form of space cubes under multiple time slices.

[0068] Each processed coral reef beach sand evolution process state element data is used as a channel. If a certain element has a single value in a designated beach sand area of ​​the entire simulation, the raster value in the entire spatial range will take this value. In general, each beach sand state sequence training data set includes multiple samples, each sample is composed of T+L beach sand state element data, and the element data represents the state of the coral reef beach sand itself and the natural and artificial geographical environment elements that drive the evolution of the process on the time slice, forming a multidimensional space cube.

[0069] return Figure 1In step S130, the SwinLSTM deep learning framework with the ability to extract spatiotemporal dependencies suitable for beach sand evolution simulation is selected, the cross-attention mechanism is integrated, and the loss function is designed according to the characteristics of beach sand morphological changes to construct a twin intelligent simulation model of coral reef beach sand evolution process. For the architecture of the twin intelligent simulation model of coral reef beach sand evolution process, please refer to Figure 4 ,like Figure 4 As shown, the twin intelligent simulation model of the coral reef beach sand evolution process includes a beach sand feature extraction module, a spatiotemporal dependency capture module and a twin simulation module. The beach sand feature extraction module adopts a cross-attention mechanism; the spatiotemporal dependency capture module consists of two parts, Swin Transformer and LSTM. The Swin Transformer part is used to vertically learn global spatial dependencies; the LSTM part is used to capture long-term and short-term time dependencies by horizontally updating the information of cell states and hidden states.

[0070] In step S140, the training data set is used to complete the training, verification and testing of the twin intelligent simulation model of the coral reef beach sand evolution process, and the twin intelligent simulation model of the target coral reef beach sand evolution process is obtained.

[0071] In some exemplary embodiments, step S140 includes steps S141 to S145.

[0072] In step S141, the training data set is divided into a training set, a validation set and a test set according to a preset ratio, wherein each sample in the training set, the validation set and the test set includes a multidimensional space cube on T+L time slices, wherein T represents the number of front time slices, which is used to extract the spatiotemporal features characterizing the beach sand evolution process, L represents the number of rear time slices, which is the target that needs to be simulated through the beach sand evolution process of T time slices, T is numbered from 1, that is, the front time slices are numbered: 1, 2, ..., T-1, T; L is numbered from T+1; that is, the rear time slices are numbered: T+1, T+2, ..., L-1, L; T and L are both positive integers.

[0073] In step S142, based on the state element data of coral reef beach sand organized in the form of spatial cubes on T time slices in the training set, the twin intelligent simulation model of the coral reef beach sand evolution process is used to extract the beach sand evolution characteristics to obtain the simulation results on the T+1-T+L time slices.

[0074] In step S143, based on the simulation results and the sample labels, the loss value is calculated using a loss function, wherein the sample labels are the probability grids of the actual exposure of the coral reef beach sand in the T+1-T+Lth time slice.

[0075] For example, the loss function is expressed as:

[0076] (3)

[0077] Among them, Loss is the loss value, is the true value on grid (i, j), is the simulated value on grid (i,j), is the weight adjustment value on grid (i, j), and K is the total number of grids in the beach sand simulation area;

[0078] Adjust as follows:

[0079] Calculate {X for each grid over the entire previous time series 1 , X 2 ,…, X T Cumulative change C ij :

[0080] (4)

[0081] The high-frequency changing beach sand boundary area will produce a larger cumulative value.

[0082] Normalize the accumulated value to the preset weight range [δ,γ]:

[0083] (5)

[0084] in, is the weight adjustment value on the grid (i, j), δ is the minimum value in the preset weight range, γ is the maximum value in the preset weight range, C ij For each grid {X 1 , X 2 ,…, X T Cumulative change, min (C) is the cumulative change of all grids in the beach sand simulation area. ij The minimum value of C in all grids in the beach sand simulation area. ij The maximum value of .

[0085] Through time series dynamic analysis, the sensitive areas of beach sand morphology change are automatically identified, including erosion / deposition boundaries, and higher loss weights are assigned to the sensitive areas, so that the model training focuses on the spatial locations where the terrain fluctuates violently. In step S144, back propagation is performed based on the loss value to update the model weights and obtain a twin intelligent simulation model of the evolution process of the intermediate coral reef beach sand.

[0086] In step S145, an early stopping mechanism is adopted to calculate the loss value of the twin intelligent simulation model of the intermediate coral reef beach sand evolution process using the validation set data until the loss value of the validation set changes within a preset range. The training is then stopped and the model with the highest accuracy on the validation set during the training process is saved to obtain the trained twin intelligent simulation model of the coral reef beach sand evolution process to avoid overfitting.

[0087] On the basis of the above method, in order to further improve the effect of the model, the training method can also include automatically adjusting the global hyperparameters of the learning rate through the particle swarm optimization algorithm in multiple rounds of training, using the Adam algorithm to update the local weights of the network, automatically training multiple twin intelligent simulation models of the coral reef beach sand evolution process, and selecting the twin intelligent simulation model of the coral reef beach sand evolution process with the smallest loss value on the verification set as the target twin intelligent simulation model of the coral reef beach sand evolution process.

[0088] return Figure 1 In step S150, the twin intelligent simulation model of the target coral reef beach sand evolution process is applied to achieve an approximate simulation of the evolution process and topography of the coral reef beach sand in the history or future specified time period.

[0089] In some exemplary embodiments, step S150 includes steps S151 to S157.

[0090] In step S151, the data to be simulated is input into the twin intelligent simulation model of the target coral reef beach sand evolution process, wherein the data type of the data to be simulated is the same as that of the training data in the training data set.

[0091] In step S152, the beach sand feature extraction module in the twin intelligent simulation model of the target coral reef beach sand evolution process is used to obtain the initial input of the coral reef beach sand state at each time slice using the cross-attention mechanism.

[0092] In the embodiment of the present invention, the data of the exposure probability layer are endogenous variables, and the remaining variable layers: natural environment dynamic factor data (sea breeze, current, waves, tide level), artificial environment dynamic factor data (shipping intensity, ship density, shoreline artificialization rate), storm surge and other extreme weather data (long-term typhoon track frequency), monsoon period attributes (northeast monsoon period, southwest monsoon period or transition period) are exogenous variables. Endogenous variables are used as queries, and exogenous variables are used as keys and values ​​to establish a connection between the two types of variables. Using the cross-attention mechanism, the initial input of the coral reef beach sand state on the t time slice is obtained based on the following formula:

[0093] (6)

[0094] in, is the value of the endogenous variable at time slice t, is the value of the exogenous variable at time slice t, represents the multi-headed cross-attention between endogenous and exogenous variables.

[0095] Assume that the input time series length of the model is T, and the beach sand state factor dataset within the time length T consists of T time slices. The value on each time slice is calculated by the above formula: The length of the output time series is L, and the output value is the probability of beach sand exposure at time L. Data at T moments are input, and the values ​​at L moments are output.

[0096] In step S153, the initial input is divided into evenly sized and non-overlapping data blocks, and the data blocks are flattened in order from left to right and from top to bottom to input into the spatiotemporal dependency capture module.

[0097] For example, within the time series length T, at each time slice t, the beach sand feature extraction module calculates The data is divided into non-overlapping small blocks, each of which is a*a in size. The data is flattened from left to right and from top to bottom, and input into the spatiotemporal dependency capture module in an embedded manner.

[0098] In step S154, the global spatial dependency of the initial input is obtained by using the Swin Transformer part in the spatiotemporal dependency capture module.

[0099] In step S155, the LSTM part in the spatiotemporal dependency capture module is used to obtain the long-term and short-term temporal dependencies of the initial input.

[0100] In step S156, based on the spatial dependency and the temporal dependency, the twin simulation module is used to twin the exposure probability of the coral reef beach sand evolution at a specified time.

[0101] In step S157, based on the exposure probability, an approximate simulation of the evolution process and topography of the coral reef beach sand in the history or future specified time period is achieved.

[0102] In step S156 and step S157, different simulation results are obtained based on different time periods of input data. The simulation results are divided into two parts: one is the evolution process of a coral reef beach sand in a specified period of the twin historical period, and the other is the prediction of the evolution process of coral reef beach sand in the future based on the evolution of a coral reef beach sand in the recent period. The model can input the beach sand state element data representing the evolution of coral reef beach sand within a known time T, predict the evolution of coral reef beach sand within the time period (T+1~T+L), and characterize it with the probability of coral reef exposure.

[0103] Simulate the evolution of coral reef beach sand during a specified period in the historical period: input the beach sand state element data that characterizes the evolution of coral reef beach sand within the historical period T, and simulate the beach sand evolution results (monthly exposure probability) within the next continuous time L.

[0104] Predict the evolution of coral reef beach sand over a period of time in the future: input the beach sand state element data that characterizes the evolution of coral reef beach sand in the most recent time T, and use the model to predict the beach sand evolution results (monthly scale exposure probability) in the future time L.

[0105] The model result is the probability of coral reef beach sand exposure in units of months, grid by grid, and position by position within the twin time period L. Isoprobability lines can be obtained based on the exposure probability. Based on the relationship between the initial beach sand simulation area terrain data, the exposure probability map, and the simulated exposure probability map, the terrain between the full probability line (p=1) and the zero probability line (p=0) of the coral reef beach sand is approximately simulated. (The coral reef area between the full probability line (p=1) and the zero probability line (p=0) will be submerged at high tide and exposed at low tide, which is an approximate intertidal zone.) Both of the above two types of evolution processes can obtain approximate terrain simulation results. The probability of beach sand exposure in a certain area in units of months is as follows: Figure 5 shown.

[0106] The embodiment of the present invention effectively extracts the main element features that affect the evolution process of coral reef beach sand as the input of the deep learning model, and fully considers the characteristics of coral reef beach sand, such as the driving force of natural and human factors, and the periodic change characteristics caused by the monsoon period.

[0107] Aiming at the problem of simulating the evolution of the entire coral reef beach containing spatial information on the grid scale within a specified continuous period of history and a specified period of time in the future, which is difficult to be solved by existing tidal flat evolution analysis methods, this method integrates the cross-attention mechanism, the SwinLSTM method combining Swin Transformer and LSTM, constructs a deep learning model, designs a loss function that focuses on grids with large changes, realizes the twin intelligent simulation of the evolution process of offshore coral reef beaches, and simulates the probability of coral reef beach exposure and approximate terrain in a specified time period.

[0108] Figure 6 A twin intelligent simulation device for the coral reef beach sand evolution process according to an embodiment of the present invention is schematically shown.

[0109] like Figure 6 As shown, the twin intelligent simulation device 800 of the coral reef beach sand evolution process of this embodiment includes an acquisition module 810, a preprocessing module 820, a model building module 830, a model training module 840 and a model application module 850.

[0110] The acquisition module 810 is used to acquire the original data of the process factors that affect the evolution of beach sand, wherein the process factors include the main natural and artificial regional environmental characteristics that drive the evolution of beach sand and the exposure morphological characteristics of the beach sand itself.

[0111] The preprocessing module 820 is used to preprocess the original data and construct a beach sand state sequence training data set in the form of a multi-dimensional space cube.

[0112] The model building module 830 is used to select the SwinLSTM deep learning framework with the ability to extract spatiotemporal dependencies suitable for beach sand evolution simulation, integrate the cross-attention mechanism, design the loss function based on the beach sand morphological change characteristics, and build a twin intelligent simulation model of the coral reef beach sand evolution process.

[0113] The model training module 840 is used to complete the training, verification and testing of the twin intelligent simulation model of the coral reef beach sand evolution process using the training data set to obtain the twin intelligent simulation model of the target coral reef beach sand evolution process.

[0114] The model application module 850 is used to apply the twin intelligent simulation model of the target coral reef beach sand evolution process to achieve an approximate simulation of the evolution process and topography of the coral reef beach sand in the past or in a specified time period in the future.

[0115] Among them, the twin intelligent simulation model of the coral reef beach sand evolution process includes a beach sand feature extraction module, a spatiotemporal dependency capture module and a twin simulation module. Among them, the beach sand feature extraction module adopts a cross-attention mechanism; the spatiotemporal dependency capture module consists of two parts, Swin Transformer and LSTM. The Swin Transformer part is used to vertically learn global spatial dependencies; the LSTM part is used to capture long-term and short-term time dependencies by horizontally updating the information of cell states and hidden states.

[0116] According to an embodiment of the present invention, any multiple modules among the acquisition module 810, the preprocessing module 820, the model building module 830, the model training module 840 and the model application module 850 can be combined into one module for implementation, or any one of the modules can be split into multiple modules. Alternatively, at least part of the functions of one or more of these modules can be combined with at least part of the functions of other modules and implemented in one module.

[0117] It will be appreciated by those skilled in the art that the features described in the various embodiments of the present invention may be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in the present invention. In particular, without departing from the spirit and teachings of the present invention, the features described in the various embodiments of the present invention may be combined and / or combined in various ways. All of these combinations and / or combinations fall within the scope of the present invention.

Claims

1. A twin intelligent simulation method for the evolution of coral reef beach sand, characterized in that: The method comprises: Obtaining the original data of the process factors that affect the evolution of beach sand, wherein the process factors include the main natural and artificial regional environmental characteristics that drive the evolution of beach sand and the exposure morphological characteristics of the beach sand itself; Preprocessing the raw data to construct a beach sand state sequence training data set in the form of a multi-dimensional space cube; We selected the SwinLSTM deep learning framework with the ability to extract spatiotemporal dependencies suitable for beach sand evolution simulation, integrated the cross-attention mechanism, designed a loss function based on the characteristics of beach sand morphological changes, and constructed a twin intelligent simulation model for the coral reef beach sand evolution process; The training data set is used to complete the training, verification and testing of the twin intelligent simulation model of the coral reef beach sand evolution process, and the twin intelligent simulation model of the target coral reef beach sand evolution process is obtained; Apply the twin intelligent simulation model of the target coral reef beach sand evolution process to achieve an approximate simulation of the evolution process and topography of the coral reef beach sand in the past or in a specified time period in the future; Among them, the twin intelligent simulation model of the coral reef beach sand evolution process includes a beach sand feature extraction module, a spatiotemporal dependency capture module and a twin simulation module. Among them, the beach sand feature extraction module adopts a cross-attention mechanism; the spatiotemporal dependency capture module consists of two parts, Swin Transformer and LSTM. The Swin Transformer part is used to vertically learn global spatial dependencies; the LSTM part is used to capture long-term and short-term time dependencies by horizontally updating the information of cell states and hidden states.

2. The method according to claim 1, characterized in that The application of the twin intelligent simulation model of the target coral reef beach sand evolution process to achieve the approximate simulation of the evolution process and topography of the coral reef beach sand in the past and future specified time periods includes: Input the data to be simulated into the twin intelligent simulation model of the target coral reef beach sand evolution process; The sand feature extraction module in the twin intelligent simulation model of the target coral reef beach sand evolution process is used to obtain the initial input of the coral reef beach sand state at each time slice by adopting the cross-attention mechanism; The initial input is divided into small data blocks of uniform size and non-overlapping, and the small data blocks are flattened in a left-to-right and top-to-bottom order to be input into a spatiotemporal dependency capture module; Using the Swin Transformer part in the spatiotemporal dependency capture module, the global spatial dependency of the initial input is obtained; Utilizing the LSTM part in the spatiotemporal dependency capturing module, the long-term and short-term temporal dependencies of the initial input are obtained; Based on the spatial dependency and the temporal dependency, using the twin simulation module, twinning the probability of coral reef beach sand evolution at a specified time; and Based on the exposure probability, the evolution process and topography of coral reef beach sand in history or in a specified time period in the future can be approximately simulated. The data to be simulated is of the same data type as the training data in the training data set.

3. The method according to claim 2, characterized in that If the input data to be simulated is a data set of beach sand state elements representing the evolution of coral reef beach sand in a certain period of history, the result of the coral reef beach sand evolution simulation is the evolution process of the specified historical period; If the input data to be simulated is a data set of beach sand state elements that characterizes the evolution of coral reef beach sand in a recent period of time, the result of the coral reef beach sand evolution simulation is the evolution process of a specified period of time in the future.

4. The method according to claim 3, characterized in that The original data include remote sensing images, ship automatic identification system data, natural environment dynamic factor data and monsoon period attributes; The beach sand state element dataset contains beach sand state element data organized in the form of spatial cubes under multiple time slices, covering endogenous variables and exogenous variables, wherein the endogenous variables include exposure probability; and the exogenous variables include natural environment dynamic element data, artificial environment dynamic element data, extreme weather data such as storm surges, and monsoon period attributes.

5. The method according to claim 4, characterized in that The cross-attention mechanism is used to obtain the initial input of the coral reef beach sand state at each time slice, using endogenous variables as queries and exogenous variables as keys and values.

6. The method according to claim 4, characterized in that The preprocessing of the raw data to construct a beach sand state sequence training data set in the form of a multi-dimensional space cube includes: Extracting the original data of indirect factors from the original remote sensing images and the ship automatic identification system data, wherein the indirect factor data include the probability of beach sand exposure, shipping intensity and ship density, shoreline artificialization rate and typhoon intensity; Setting a grid unit framework of a specified spatial resolution and geographic coordinate system, and unifying the indirect element data and the natural environment dynamic element data that can be directly obtained into the grid unit framework; According to the calculation target of indirect factors, the spatial analysis calculation method of geographic information system and cloud batch processing technology of remote sensing big data are used to complete batch calculation and obtain the final calculation results of indirect factors; Use one-hot encoding to convert the monsoon period attribute into a vector, and obtain the monsoon period attribute represented by the vector; The final calculation results of the indirect factors, the natural environment dynamic factors that can be directly obtained, and the monsoon period attributes represented by the vector are integrated to form beach sand state factor data in the form of a multidimensional space cube, and each time slice includes a multidimensional space cube; The beach sand state element data set includes beach sand state element data organized in the form of space cubes under multiple time slices.

7. The method according to claim 1, characterized in that The training, verification and testing of the twin intelligent simulation model of the coral reef beach sand evolution process using the training data set includes: The training data set is divided into a training set, a validation set and a test set according to a preset ratio, wherein each sample in the training set, the validation set and the test set includes a multidimensional space cube on T+L time slices, wherein T represents the number of front time slices, which is used to extract the spatiotemporal features characterizing the beach sand evolution process, L represents the number of rear time slices, which is the target to be simulated by the beach sand evolution process of T time slices, T is numbered from 1, that is, the front time slices are numbered as: 1, 2, ..., T-1, T; L is numbered from T+1; that is, the rear time slices are numbered as: T+1, T+2, ..., L-1, L; T and L are both positive integers; Based on the state element data of coral reef beach sand organized in the form of spatial cubes on T time slices in the training set, the twin intelligent simulation model of coral reef beach sand evolution process is used to extract beach sand evolution characteristics to obtain simulation results on T+1-T+L time slices; Based on the simulation results and sample labels, using a loss function, calculating a loss value; Back propagation is performed based on the loss value to update the model weights to obtain a twin intelligent simulation model of the intermediate coral reef beach sand evolution process; An early stopping mechanism is adopted to calculate the loss value of the twin intelligent simulation model of the evolution process of the intermediate coral reef beach sand using the validation set data until the loss value of the validation set changes within the preset range. The training is stopped and the model with the highest accuracy on the validation set during the training process is saved to obtain the trained twin intelligent simulation model of the evolution process of the coral reef beach sand to avoid overfitting. The preset range means that the number of training rounds is set to N. If the loss value of the validation set is still increasing after N rounds, training will not be continued and the model with the highest accuracy on the validation set will be saved. N is a positive integer. The sample labels are the actual exposure probability grids of coral reef beach sands in the T+1-T+L time slices.

8. The method according to claim 7, characterized in that The method also includes: in multiple rounds of training, automatically adjusting the global hyperparameter of the learning rate through a particle swarm optimization algorithm, using the Adam algorithm to update the local weights of the network, automatically training multiple twin intelligent simulation models of the coral reef beach sand evolution process, and selecting the twin intelligent simulation model of the coral reef beach sand evolution process with the smallest loss value on the verification set as the target twin intelligent simulation model of the coral reef beach sand evolution process.

9. The method according to claim 7, characterized in that: The loss function is: Among them, Loss is the loss value, is the true value on grid (i, j), is the simulated value on grid (i,j), is the weight adjustment value on grid (i, j), and K is the total number of grids in the beach sand simulation area; Adjust as follows: Calculate {X for each grid over the entire previous time series 1 , X 2 ,…, X T Cumulative change: The high-frequency changing beach sand boundary area will produce a larger cumulative value. Normalize the accumulated value to the preset weight range [δ,γ]: Through time-series dynamic analysis, the sensitive areas of beach sand morphology change are automatically identified, including erosion / deposition boundaries, and higher loss weights are assigned to the sensitive areas, so that the model training focuses on the spatial locations where the terrain fluctuates violently.

10. A twin intelligent simulation device for the evolution of coral reef beach sand, characterized in that: The device comprises: An acquisition module is used to acquire the original data of the process factors affecting the beach sand evolution, wherein the process factors include the main natural and artificial regional environmental characteristics driving the beach sand evolution and the exposure morphological characteristics of the beach sand itself; A preprocessing module, used for preprocessing the raw data to construct a beach sand state sequence training data set in the form of a multi-dimensional space cube; The model building module is used to select the SwinLSTM deep learning framework with the ability to extract spatiotemporal dependencies suitable for beach sand evolution simulation, integrate the cross-attention mechanism, design the loss function based on the characteristics of beach sand morphological changes, and build a twin intelligent simulation model of coral reef beach sand evolution process; A model training module is used to complete the training, verification and testing of the twin intelligent simulation model of the coral reef beach sand evolution process using the training data set to obtain the twin intelligent simulation model of the target coral reef beach sand evolution process; A model application module, used to apply the twin intelligent simulation model of the target coral reef beach sand evolution process to achieve an approximate simulation of the evolution process and topography of the coral reef beach sand in the past or in a specified time period in the future; Among them, the twin intelligent simulation model of the coral reef beach sand evolution process includes a beach sand feature extraction module, a spatiotemporal dependency capture module and a twin simulation module. Among them, the beach sand feature extraction module adopts a cross-attention mechanism; the spatiotemporal dependency capture module consists of two parts, Swin Transformer and LSTM. The Swin Transformer part is used to vertically learn global spatial dependencies; the LSTM part is used to capture long-term and short-term time dependencies by horizontally updating the information of cell states and hidden states.

Images