A coastline dynamic monitoring and early warning method and system

Abstract

The application provides a coastline dynamic monitoring and early warning method and system, multi-modal original data of a coastline is collected, and then a multi-modal data set aligned in a tidal phase is obtained; the multi-modal data set is unified in a spatial reference to obtain a multi-modal space-time data set jointly aligned in a tidal phase and a spatial coordinate; the multi-modal space-time data set is decomposed into a driving component of a tide, an erosion component of a storm, an action component of a wave and a disturbance component of human activities, and then multi-modal space-time features of a decoupled multi-modal data set are obtained by respectively extracting deep features; the multi-modal space-time features are modeled in a space-time heterogeneous manner to generate an evolution feature tensor of a unified space-time reference of a coastline, and coastline change monitoring and early warning is performed based on the evolution feature tensor. The scheme of the application can realize multi-modal monitoring data tidal phase self-adaptive alignment and precise decoupling of ocean dynamic process physical interpretability for coastline change space-time collaborative monitoring and early warning.

Description

Technical Field

[0001] This application relates to the field of big data analysis technology, and more specifically, to a method and system for dynamic monitoring and early warning of coastlines. Background Technology

[0002] As a crucial transitional zone for land-sea interaction, the coastline's dynamic evolution is influenced by the coupling of multiple ocean dynamic processes, including tidal periodicity, sudden storm impacts, continuous wave action, and human disturbance. Accurately monitoring the patterns of coastline changes and providing timely warnings of potential erosion risks are of great significance for disaster prevention and mitigation, ecological protection, and resource development in coastal areas. With the rapid development of multi-source observation technologies such as space-based remote sensing, airborne drones, shore-based video, sea-based buoys, and seabed fiber optic sensing, coastline monitoring has evolved from a single method to a three-dimensional "sky, land, and sea" perception model, providing a technological foundation for comprehensively capturing information on the multiple physical attributes of the coastal zone.

[0003] Current coastline dynamic monitoring technologies typically employ methods such as satellite image geometric registration or UAV aerial surveying to acquire coastal spatial data, and then analyze coastline change trends using numerical models or statistical methods. However, existing methods suffer from several drawbacks at the data processing level. The lack of a unified spatiotemporal reference for multimodal data, particularly the tidal phase difference between satellite transit times and UAV aerial photography times, leads to inconsistent definitions of the "waterline," making it difficult to accurately align heterogeneous data at the same tidal evolution stage. At the signal decomposition level, traditional filtering methods cannot physically separate aliased tidal, storm, wave, and human activity signals, resulting in a lack of interpretability of the dynamic processes in feature extraction results. At the spatial positioning level, the scarcity of ground control points in nearshore areas makes it difficult to establish a unified, high-precision spatial reference frame between UAVs and shore-based equipment, and the cumulative geometric correction errors affect the accuracy of change analysis. Finally, at the predictive modeling level, existing models struggle to simultaneously capture the collaborative evolution patterns between different spatial locations along the coastline and the long- and short-term temporal dependencies along the tidal phase, resulting in early warning accuracy that fails to meet practical application requirements. Therefore, how to achieve spatiotemporal collaborative monitoring and early warning of coastline changes by adaptively aligning tidal phases of multimodal monitoring data and accurately decoupling the physical interpretability of ocean dynamic processes has become a challenge for the industry. Summary of the Invention

[0004] This application provides a method and system for dynamic monitoring and early warning of coastlines, which can realize spatiotemporal collaborative monitoring and early warning of coastline changes by adaptively aligning tidal phases of multimodal monitoring data and accurately decoupling the physical interpretability of ocean dynamic processes.

[0005] Firstly, this application provides a method for dynamic monitoring and early warning of coastlines, including: Collect multimodal raw data of the coastline, determine the tidal phase corresponding to the acquisition time of each modality in the multimodal raw data, and then resample the multimodal raw data to a unified tidal phase time axis based on all tidal phases to obtain a tidal phase aligned multimodal dataset. The multimodal dataset is spatially referenced by UAV collaborative positioning without ground control points to obtain a multimodal spatiotemporal dataset jointly aligned with tidal phase and spatial coordinates. Based on a pre-built overcomplete dictionary of ocean dynamic processes, the multimodal spatiotemporal dataset is decomposed into the driving component of tides, the erosion component of storms, the action component of waves, and the disturbance component of human activities. Then, the deep features of the driving component, erosion component, action component, and disturbance component are extracted respectively to obtain the multimodal spatiotemporal features of the decoupled multimodal dataset. Spatiotemporal heterogeneous modeling is performed on the multimodal spatiotemporal features to generate an evolution feature tensor of a unified spatiotemporal benchmark for the coastline, and then coastline change monitoring and early warning are performed based on the evolution feature tensor.

[0006] In some embodiments, determining the tidal phase corresponding to the data acquisition time of each modality in the multimodal raw data specifically includes: For each modal data in the multimodal raw data, tidal harmonic analysis is performed on the acquisition time period of the modal data to obtain the tidal phase angle of each tidal component at each acquisition time. By performing multi-tidal synthesis on all tidal phase angles at the same acquisition time, the comprehensive tidal phase values ​​of the modal data at each acquisition time are obtained; The tidal phase corresponding to the acquisition time of the modal data is determined based on the comprehensive tidal phase value at all acquisition times, thereby obtaining the tidal phase corresponding to the acquisition time of each modal data in the multimodal raw data.

[0007] In some embodiments, resampling the original multimodal data onto a unified tidal phase time axis based on all tidal phases to obtain a tidal phase-aligned multimodal dataset specifically includes: Construct a regular sequence of phase nodes indexed by the tidal phase angle based on all tidal phases; Using the regular phase node sequence as the target time reference, interpolation is performed on each modal data in the multimodal raw data to obtain synchronous sampling data of each modal data on the regular phase node sequence; All synchronously sampled data are integrated to obtain a tidal phase-aligned multimodal dataset.

[0008] In some embodiments, decomposing the multimodal spatiotemporal dataset into tidal driving components, storm erosion components, wave action components, and human activity disturbance components based on a pre-built overcomplete dictionary of ocean dynamic processes specifically includes: A pre-constructed, complete dictionary of ocean dynamic processes; The multimodal spatiotemporal dataset is sparsely decomposed on the overcomplete dictionary to obtain the sparse coefficient matrix corresponding to each atom in the overcomplete dictionary; Based on the sparse coefficient matrix, the driving component of tides, the erosion component of storms, the action component of waves, and the disturbance component of human activities are separated from the multimodal spatiotemporal dataset.

[0009] In some embodiments, extracting deep features from the driving component, erosion component, action component, and perturbation component to obtain the multimodal spatiotemporal features of the decoupled multimodal dataset specifically includes: The dynamic feature map of tidal forces in the driving components is extracted by time-series modeling; The morphological feature maps of storm erosion in the erosion components were extracted through spatial analysis; Energy feature maps of waves in the active components are extracted by time-frequency transformation; The perturbation feature map of human activity in the perturbation component is extracted by pattern recognition; The dynamic feature map, the morphological feature map, the energy feature map, and the perturbation feature map are adaptively fused to obtain the multimodal spatiotemporal features of the decoupled multimodal dataset.

[0010] In some embodiments, performing spatiotemporal heterogeneous modeling on the multimodal spatiotemporal features to generate an evolution feature tensor for a unified spatiotemporal reference of the coastline specifically includes: An initial spatiotemporal topology graph is established using the coastal zone spatial grid as nodes and the aforementioned multimodal spatiotemporal features as node attributes; The initial spatiotemporal topology graph is subjected to graph convolutional neural network learning to extract spatially dependent co-evolutionary features; The evolution characteristics of each node in the initial spatiotemporal topology graph along the tidal phase dimension are captured by timing modeling using gated cyclic units. The co-evolutionary features are fused with all trend evolutionary features in a multidimensional manner to generate an evolutionary feature tensor of a unified spatiotemporal benchmark for the coastline.

[0011] In some embodiments, coastline change monitoring and early warning based on the evolution feature tensor specifically includes: A coastline change prediction model is constructed using a pre-trained long short-term memory network. The evolution feature tensor is input into the change prediction model for prediction, thereby obtaining the early warning level of the coastline change prediction index. The warning level is output for coastline change monitoring and early warning.

[0012] Secondly, this application provides a coastline dynamic monitoring and early warning system, comprising: The acquisition module is used to acquire multimodal raw data of the coastline, determine the tidal phase corresponding to the acquisition time of each modality in the multimodal raw data, and then resample the multimodal raw data to a unified tidal phase time axis based on all tidal phases to obtain a tidal phase aligned multimodal dataset. The processing module is used to perform spatial benchmark unification on the multimodal dataset through UAV collaborative positioning without ground control points, so as to obtain a multimodal spatiotemporal dataset with tidal phase-spatial coordinate joint alignment. The processing module is used to decompose the multimodal spatiotemporal dataset into tidal driving components, storm erosion components, wave action components, and human activity disturbance components based on a pre-built overcomplete dictionary of ocean dynamic processes. Then, it extracts the deep features of the driving components, erosion components, action components, and disturbance components respectively to obtain the multimodal spatiotemporal features of the decoupled multimodal dataset. The execution module is used to perform spatiotemporal heterogeneous modeling on the multimodal spatiotemporal features, generate an evolution feature tensor of a unified spatiotemporal benchmark for the coastline, and then perform coastline change monitoring and early warning based on the evolution feature tensor.

[0013] Thirdly, this application provides a computer device, the computer device including a memory and a processor, the memory storing code, and the processor being configured to acquire the code and execute the above-described method for dynamic monitoring and early warning of coastlines.

[0014] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method for dynamic monitoring and early warning of coastlines.

[0015] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects: The method and system for dynamic monitoring and early warning of coastlines provided in this application first collect multimodal raw data of the coastline, determine the tidal phase corresponding to the data collection time of each modality in the multimodal raw data, and then resample the multimodal raw data to a unified tidal phase time axis based on all tidal phases to obtain a tidal phase aligned multimodal dataset. The multimodal dataset is then spatially aligned using UAV collaborative positioning without ground control points to obtain a tidal phase-spatial coordinate jointly aligned multimodal spatiotemporal dataset. Based on a pre-constructed overcomplete dictionary of ocean dynamic processes, the multimodal spatiotemporal dataset is decomposed into tidal driving components, storm erosion components, wave action components, and human activity disturbance components. Deep features of the driving, erosion, action, and disturbance components are then extracted to obtain the multimodal spatiotemporal features of the decoupled multimodal dataset. Spatiotemporal heterogeneous modeling is performed on the multimodal spatiotemporal features to generate an evolution feature tensor of a unified spatiotemporal reference for the coastline, and then coastline change monitoring and early warning are performed based on the evolution feature tensor.

[0016] Therefore, this application uses the aforementioned evolutionary feature tensor for coastline change monitoring and early warning. First, determining the multimodal dataset yields a standardized data matrix formed by integrating the original multimodal data according to the phase node order. The determination of the multimodal dataset involves introducing the tidal phase angle as a unified time index benchmark, transforming heterogeneous data such as satellite remote sensing, UAV waypoint clouds, shore-based video streams, buoy hydrological time series, and seabed fiber optic vibration from the original absolute time axis to the tidal evolution phase axis, fundamentally eliminating the data temporal misalignment problem caused by tidal phase differences at the time of acquisition. Furthermore, an adaptive interpolation algorithm is used to resample the tidal phase nodes of various modal data at different sampling frequencies, ensuring that subsequent... The decomposition of ocean dynamic processes enables physical source separation based on synchronous observations of the same tidal evolution stage, ensuring that the decoupled tidal-driven components, storm erosion components, wave-driven components, and human-induced disturbance components are physically interpretable in the temporal dimension. This provides high-quality, time-aligned input for subsequent unification of spatial benchmarks without control points, deep feature extraction, and spatiotemporal co-evolution modeling by constructing a unified data base indexed by tidal phase. This solves the technical challenges of feature extraction distortion and limited change prediction accuracy caused by tidal phase mismatch in multi-source data in existing technologies. Furthermore, determining the multimodal spatiotemporal features allows for a complete characterization of the coupling of multiple physical processes along the coastline. The unified high-dimensional representation of evolutionary patterns and the determination of multimodal spatiotemporal features can separate aliased tidal, storm, wave, and human activity signals into independent components with clear physical meanings and uncover their deep-seated evolutionary patterns. This effectively solves the problem of lack of physical interpretability in feature extraction caused by multi-source signal aliasing in existing technologies. Simultaneously, this feature serves as the input basis for subsequent spatiotemporal heterogeneous modeling, enabling deep fusion of cross-modal information while preserving the independent physical properties of each dynamic process. This provides high-quality feature representations for coastline change prediction that include both spatial neighborhood synergistic effects and tidal temporal dependencies, significantly improving the ability of early warning models to characterize complex marine dynamic environments. Prediction accuracy; Finally, determining the evolution feature tensor can yield a three-dimensional feature representation reflecting the nonlinear correlation, collaborative change patterns, and temporal evolution patterns along the tidal phase dimension between different spatial locations of the coastline. Determining the evolution feature tensor enables the change prediction model to simultaneously perceive spatial neighborhood aggregation information and historical temporal evolution information, thereby overcoming the technical limitations of traditional methods that separate spatiotemporal modeling, significantly improving the accuracy of coastline change prediction, the timeliness of disaster early warning, and the reliability of decision support. In summary, based on the above scheme, spatiotemporal collaborative monitoring and early warning of coastline changes can be achieved through adaptive alignment of multimodal monitoring data with tidal phase and precise decoupling of the physical interpretability of ocean dynamic processes. Attached Figure Description

[0017] Figure 1 This is an exemplary flowchart of a dynamic monitoring and early warning method for coastlines according to some embodiments of this application; Figure 2 This is a flowchart illustrating the operation of determining multimodal spatiotemporal features according to some embodiments of this application; Figure 3 This is an exemplary flowchart illustrating the determination of evolution feature tensors according to some embodiments of this application; Figure 4 This is a schematic diagram of the structure of a dynamic monitoring and early warning system for a coastline according to some embodiments of this application; Figure 5 This is an internal structural diagram of a computer device for implementing a dynamic monitoring and early warning method for coastlines, according to some embodiments of this application. Detailed Implementation

[0018] To better understand the technical solution of this application, the technical solution of this application will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] refer to Figure 1 The figure is an exemplary flowchart of a dynamic monitoring and early warning method for coastlines according to some embodiments of this application. The dynamic monitoring and early warning method for coastlines mainly includes the following steps: In step 101, multimodal raw data of the coastline is collected, the tidal phase corresponding to the acquisition time of each modality in the multimodal raw data is determined, and then the multimodal raw data is resampled to a unified tidal phase time axis based on all tidal phases to obtain a tidal phase aligned multimodal dataset.

[0020] It should be noted that, in this application, the multimodal raw data is a collection of unprocessed environmental elements covering different physical properties of the coastal zone, including satellite remote sensing images, UAV aerial point clouds, shore-based video streams, hydrological time-series data, and seabed distributed fiber optic sensing vibration signals. This can provide basic data support for subsequent tidal phase alignment and decoupling of ocean dynamic processes. In specific implementation, the collection of multimodal raw data of the coastline can be achieved in the following way: multimodal raw data of the coastline can be obtained through a multi-source sensing platform. This multimodal raw data covers satellite remote sensing images, UAV aerial point clouds, shore-based video streams, buoy hydrological time-series data, and seabed distributed fiber optic sensing vibration signals, so as to comprehensively cover the space-based, air-based, shore-based, sea-based, and seabed observation dimensions of the coastal zone.

[0021] In some embodiments, determining the tidal phase corresponding to the data acquisition time of each modality in the multimodal raw data can be achieved by the following steps: For each modal data in the multimodal raw data, tidal harmonic analysis is performed on the acquisition time period of the modal data to obtain the tidal phase angle of each tidal component at each acquisition time. By performing multi-tidal synthesis on all tidal phase angles at the same acquisition time, the comprehensive tidal phase values ​​of the modal data at each acquisition time are obtained; The tidal phase corresponding to the acquisition time of the modal data is determined based on the comprehensive tidal phase value at all acquisition times, thereby obtaining the tidal phase corresponding to the acquisition time of each modal data in the multimodal raw data.

[0022] It should be noted that, in this application, modal data refers to an independent data unit from a single sensing source that characterizes a specified physical property of the coastal zone, such as satellite remote sensing imagery, UAV aerial point cloud, shore-based video stream, hydrological time series data, or seabed distributed fiber optic sensing vibration signal, which can be used as the basic processing object for subsequent tidal phase alignment, marine dynamic process decomposition, and deep feature extraction.

[0023] In practice, tidal harmonic analysis is performed on the acquisition time period of the modal data to obtain the tidal phase angle of each tidal constituent at each acquisition time. This can be achieved in the following way: First, the water level time series data of the tide gauge station or the tidal level sequence data extracted from the global tidal model can be obtained within the acquisition time period corresponding to the modal data. Then, the harmonic constant of each tidal constituent within the acquisition time period is calculated using a tidal harmonic analysis algorithm (such as the Foreman tidal harmonic analysis package). This constant includes the average amplitude of each tidal constituent and the phase lag between the actual high tide time and the astronomical equilibrium high tide time, which characterizes the seabed friction and seawater inertia. The lag angle is then calculated. Next, using the phase angle calculation formula in the tidal height expression, the phase angle is calculated for each acquisition moment (e.g., per second) within the acquisition period based on all calculated harmonic constants and known astronomical parameters (e.g., tidal angular rate and astronomical initial phase angle). This yields the tidal phase angle for each tidal constituent at each acquisition moment. The tidal phase angle refers to the instantaneous phase of each major tidal constituent at a specified moment, which can finely characterize the evolution of tidal components in different astronomical cycles, thus providing a physically interpretable basic phase component for subsequent multi-tidal synthesis and improving the accuracy of tidal decomposition and reconstruction.

[0024] In specific implementation, the comprehensive tidal phase value of modal data at each acquisition time can be obtained by multi-tidal synthesis of all tidal phase angles at the same acquisition time in the following way: First, based on the principle of tidal synthesis (e.g., using the tidal height reconstruction method of multi-tidal superposition), the tidal phase angle of each tidal component at the same acquisition time and its corresponding average amplitude can be weighted, and the weighted tidal components can be linearly summed by cosine function superposition to generate a synthetic tidal waveform that fully represents the actual tidal evolution state within the acquisition time period; then, the instantaneous phase corresponding to each acquisition time is extracted from the synthetic tidal waveform as the comprehensive tidal phase value of the modal data at that acquisition time, thereby obtaining the comprehensive tidal phase value of the modal data at each acquisition time; wherein, the comprehensive tidal phase value is a quantitative index representing the actual tidal evolution state at a specified acquisition time, which can fuse the phase information of multiple tidal components into a comprehensive tidal characteristic parameter, providing a unified time reference for tidal phase alignment.

[0025] In practice, determining the tidal phase corresponding to the acquisition time of modal data based on the comprehensive tidal phase value at all acquisition times can be achieved in the following way: First, a mapping table between acquisition times and comprehensive tidal phase values ​​can be constructed to obtain the projection index of the modal data time series onto the tidal phase space; then, using the comprehensive tidal phase value as the time anchor reference for each acquisition time, a phase space transformation is performed on the original time axis to obtain the tidal phase corresponding to the modal data at each acquisition time.

[0026] It should be noted that, in this application, tidal phase refers to the tidal state identifier at each acquisition time in the multimodal raw data. It can convert the absolute timestamps of different modal data into a unified tidal evolution stage marker, thereby solving the problem of temporal misalignment of multimodal data caused by tidal changes.

[0027] In some embodiments, resampling the original multimodal data onto a unified tidal phase time axis based on all tidal phases to obtain a tidal phase-aligned multimodal dataset can be achieved through the following steps: Construct a regular sequence of phase nodes indexed by the tidal phase angle based on all tidal phases; Using the regular phase node sequence as the target time reference, interpolation is performed on each modal data in the multimodal raw data to obtain synchronous sampling data of each modal data on the regular phase node sequence; All synchronously sampled data are integrated to obtain a tidal phase-aligned multimodal dataset.

[0028] In specific implementation, constructing a regular phase node sequence indexed by the tidal phase angle based on all tidal phases can be achieved in the following way: First, determine the complete tidal cycle range from 0 degrees to 360 degrees based on the values ​​of all tidal phases, and use an equal-interval sampling strategy (e.g., every 1 degree or every 5 degrees as an interval) to uniformly divide the complete tidal cycle into multiple consecutive phase nodes; then, arrange each divided phase node in order according to its corresponding angle value to form a regular phase node sequence with clear physical meaning; wherein, the regular phase node sequence is a standardized time reference sequence formed by uniformly dividing the complete tidal cycle with tidal phase values ​​as indexes, which can be used to unify multimodal data from different acquisition times to the same tidal evolution stage, realizing the spatiotemporal reference conversion from the original absolute time axis to the tidal evolution phase axis.

[0029] In specific implementation, interpolating each modal data in the multimodal raw data using the regular phase node sequence as the target time reference to obtain the synchronous sampling data of each modal data on the regular phase node sequence can be achieved in the following way: First, an appropriate interpolation basis function can be selected for the numerical characteristics of each type of modal data. For example, piecewise cubic Hermit interpolation can be used for continuous hydrological time series data to ensure the smoothness of the interpolation curve, and nearest neighbor interpolation can be used for discrete event data to maintain the original attributes of the event; then, for each modal data, the position of the node to be interpolated on the original time axis is identified according to the position of the node in the regular phase node sequence. Two adjacent measured data points are used as input parameters, with the tidal phase difference and numerical change gradient corresponding to the acquisition time of these two adjacent measured data points. The numerical estimate of the node to be interpolated is calculated. Then, all nodes in the regular phase node sequence are traversed, and the above interpolation calculation process is repeated to generate a synchronous sampling data sequence that corresponds one-to-one with the target phase node. Through the above steps, the synchronous sampling data of each modality data on the regular phase node sequence can be obtained. The synchronous sampling data is a set of observations with a consistent time index generated after mapping the original modality data to the regular phase node sequence, which can be used to eliminate the problem of data time-domain misalignment caused by tidal changes.

[0030] In specific implementation, the integration of all synchronously sampled data to obtain a tidal phase-aligned multimodal dataset can be achieved in the following way: First, the self-consistency of the synchronously sampled data for each mode can be checked, for example, by using the chi-square test to verify the reliability of the interpolation results by calculating the statistical characteristic differences of the data before and after interpolation; then, the sampled values ​​of different modes at the same phase node are cross-validated, for example, by comparing the spatial position of the waterline retrieved from satellite remote sensing images with the waterline observed from shore-based video, identifying and eliminating alignment samples that may have anomalies; finally, all the validated modal data are horizontally stitched together according to the phase node order to construct a two-dimensional data matrix with the tidal phase value as the row index and different modes as the column index, and this data matrix is ​​used as the tidal phase-aligned multimodal dataset; wherein, the multimodal dataset is a standardized data matrix formed by integrating the original multimodal data according to the phase node order, which can be used to provide a unified input basis for tidal phase alignment for subsequent ocean dynamic process decomposition.

[0031] In step 102, the multimodal dataset is spatially referenced and unified by UAV collaborative positioning without ground control points to obtain a multimodal spatiotemporal dataset with tidal phase-spatial coordinate joint alignment.

[0032] In some embodiments, the spatial reference unification of the multimodal dataset through UAV cooperative positioning without ground control points, to obtain a multimodal spatiotemporal dataset jointly aligned with tidal phase and spatial coordinates, can be achieved by the following steps: By using collaborative localization of a drone swarm without ground control points, geometric correction and coordinate calculation are performed on the spatial data in the multimodal dataset to obtain the spatial coordinate data of the absolute pose of the drone swarm. Based on the spatiotemporal correlation between the spatial coordinate data and the modal data in the multimodal dataset, a multimodal spatiotemporal dataset with tidal phase-spatial coordinate joint alignment is obtained.

[0033] In specific implementation, the spatial coordinate data of the multimodal dataset is geometrically corrected and coordinately solved through UAV swarm collaborative positioning without ground control points. This results in the spatial coordinate data of the UAV swarm's absolute pose, which can be achieved in the following way: First, a collaborative positioning technique based on visual markers (e.g., a collaborative visual deformation monitoring method for UAVs without control points) can be used to solve the stereo image pair constructed by the hovering observation UAV and the reference UAV, and the precise pixel position of the target markers set on the UAV body is obtained in the image. Second, a high-precision ranging value between each UAV is obtained using a UAV swarm networking data link system, and a relative positioning algorithm based on multidimensional scaling technology (e.g., the relative positioning technology based on UAV swarm self-organizing network data link ranging proposed by the Beijing Telemetry Technology Research Institute) is used to perform dual-centering processing and eigenvalue decomposition on the high-precision ranging value to solve the relative spatial position relationship of all UAVs in the swarm. Then, a receiver equipped with real-time dynamic positioning technology is used... Using a reference UAV as a spatial reference anchor node, the relative position relationship is fused with the absolute position information of the anchor node through a fusion algorithm based on extended Kalman filtering (e.g., the 3D distributed UAV cooperative localization method based on high-precision ranging information proposed by the China Academy of Launch Vehicle Technology). A covariance intersection method is introduced to handle the unknown correlation between the state estimates of different UAVs. Simultaneously, an adaptive adjustment method based on chi-square detection is used to filter out inconsistencies and false information in ultra-wideband dynamic ranging. Finally, the high-precision absolute pose obtained from the fusion solution is used as the spatial reference coordinates to inversely calculate the spatial position of the shore-based fixed camera and correct the geometric distortion of the UAV aerial point cloud in the multimodal dataset, obtaining the spatial coordinate data of the absolute pose of the UAV cluster. The spatial coordinate data refers to the high-precision absolute pose information serving as a unified spatial reference frame for the coastal monitoring area, which can be used to correct the geometric distortion of multi-source spatial data and achieve accurate alignment of heterogeneous spatial information in the same coordinate system.

[0034] In specific implementation, the multimodal spatiotemporal dataset with tidal phase-spatial coordinate joint alignment is obtained by spatiotemporally associating the spatial coordinate data with the modal data in the multimodal dataset. This can be achieved in the following way: First, the spatial coordinate data can be used as a unified spatial reference framework to establish a high-precision spatial reference system for the coastal monitoring area. The inherent spatial attributes of each modal data in the multimodal dataset are extracted, including the four corner geographic coordinates of satellite remote sensing images, the three-dimensional coordinates of each point in the UAV aerial point cloud, the installation position of the camera corresponding to the shore-based video stream, and the anchoring latitude and longitude of the buoy hydrological time series data. These heterogeneous spatial information are then uniformly converted to the same spatial coordinate system based on the absolute pose of the UAV cluster. Then, using the tidal phase value as a unified time index reference, the spatially unified modal data are arranged according to their acquisition... The tidal phase corresponding to each moment is matched in both spatiotemporal context. For continuous data streams (e.g., shore-based video streams and buoy hydrological time-series data), a spatiotemporal interpolation algorithm is used to ensure that they correspond precisely with discrete data (e.g., satellite remote sensing images and UAV aerial point clouds) at discrete tidal phase angle nodes. Finally, all modal data that have undergone spatiotemporal joint alignment are integrated and encapsulated according to a unified data structure to generate a spatiotemporal dataset with tidal phase angle as the time index and UAV spatial coordinate system as the spatial reference. This dataset serves as a multimodal spatiotemporal dataset with tidal phase-spatial coordinate joint alignment. The multimodal spatiotemporal dataset refers to multi-source monitoring data that has been spatiotemporally and spatially integrated through tidal phase alignment and geometric correction. It can be used to eliminate spatiotemporal heterogeneity differences between multimodal data and provide a unified input for subsequent decoupling of ocean dynamic processes and feature extraction.

[0035] In step 103, the multimodal spatiotemporal dataset is decomposed into tidal driving components, storm erosion components, wave action components, and human activity disturbance components based on a pre-constructed overcomplete dictionary of ocean dynamic processes. Then, deep features of the driving components, erosion components, action components, and disturbance components are extracted respectively to obtain the multimodal spatiotemporal features of the decoupled multimodal dataset.

[0036] In some embodiments, decomposing the multimodal spatiotemporal dataset into tidal driving components, storm erosion components, wave action components, and human activity disturbance components based on a pre-built overcomplete dictionary of ocean dynamic processes can be achieved through the following steps: A pre-constructed, complete dictionary of ocean dynamic processes; The multimodal spatiotemporal dataset is sparsely decomposed on the overcomplete dictionary to obtain the sparse coefficient matrix corresponding to each atom in the overcomplete dictionary; Based on the sparse coefficient matrix, the driving component of tides, the erosion component of storms, the action component of waves, and the disturbance component of human activities are separated from the multimodal spatiotemporal dataset.

[0037] In practical implementation, the pre-constructed complete dictionary of ocean dynamic processes can be achieved in the following way: First, multimodal data from historical ocean monitoring data can be collected, and physical process analysis based on the geohydrodynamic mechanism can be performed on the multimodal data. That is, based on the differences in dynamic attributes of six typical motion processes—large-scale circulation, mesoscale eddies, barotropic tides, and low-mode and high-mode internal waves—standard waveforms and spatial modes describing different types of ocean dynamic processes can be established. Then, the K-singular value decomposition method in the dictionary learning algorithm is used to iteratively train and atomically optimize the standard waveforms and spatial modes to generate a dictionary containing tidal fundamental waves, storm impact waveforms, wave spectrum templates, and human activity vibrations. An overcomplete dictionary is constructed using multiple representative atoms, such as dynamic features. Each atom corresponds to a pure waveform or spatial structure of a specified type of ocean dynamic process. The overcomplete dictionary is then expanded and verified by generating theoretical waveforms covering different sea states and topographic conditions through numerical simulation, ensuring that the dictionary atoms can fully represent various ocean dynamic processes induced by tides, storms, waves, and human activities. Finally, the verified overcomplete dictionary is permanently stored as an overcomplete dictionary of ocean dynamic processes. The overcomplete dictionary refers to a signal representation basis composed of atoms representing the pure waveforms of various ocean dynamic processes, which can sparsely decompose complex coastal multimodal observation data into components from different physical sources.

[0038] In specific implementation, the multimodal spatiotemporal dataset is sparsely decomposed on the overcomplete dictionary to obtain the sparse coefficient matrix corresponding to each atom in the overcomplete dictionary. This can be achieved in the following way: First, each modal data in the multimodal spatiotemporal dataset can be represented as a high-dimensional signal vector, and the high-dimensional signal vector is projected onto the feature space formed by the pre-constructed overcomplete dictionary. The orthogonal matching pursuit algorithm is used to select the atom most relevant to the observation residual of the current high-dimensional signal vector from the overcomplete dictionary in each iteration, and the corresponding sparse coefficients are updated by the least squares method. Then, sparse encoding is performed on each modal data, ensuring that the reconstruction error is less than a preset threshold, so that the atoms participating in the representation are sparsely encoded. The minimum number of atoms is required. Then, during the iteration of the orthogonal matching pursuit algorithm, the Gram matrix between the atoms in the dictionary is solved quickly by the Cholesky decomposition to accelerate the selection process of the optimal atom in each iteration. Finally, after the iteration of all modal data converges, a sparse coefficient matrix is ​​output. The rows of the sparse coefficient matrix correspond to each atom in the overcomplete dictionary, and the columns correspond to each modal data in the multimodal spatiotemporal dataset. Each non-zero element in the matrix represents the contribution weight of the corresponding atom in reconstructing the modal data. The sparse coefficient matrix is ​​a numerical matrix that reflects the contribution weight of each atom in the overcomplete dictionary and can quantitatively describe the degree of participation of different ocean dynamic processes in the original signal.

[0039] In specific implementation, the separation of the tidal driving component, storm erosion component, wave impact component, and human activity disturbance component from the multimodal spatiotemporal dataset based on the sparse coefficient matrix can be achieved in the following way: First, based on the distribution of non-zero elements in the sparse coefficient matrix, the atoms used by each modal data in the overcomplete dictionary and their corresponding contribution weights can be identified, and these atoms can be classified into four categories according to their physical labels: tide, storm, wave, or human activity; Second, for each category, all atoms classified into that category are weighted linearly according to their sparse coefficients by matrix multiplication. The pure signal waveforms corresponding to each class are reconstructed by multiplying the subset of atoms belonging to that class in the complete dictionary by the coefficient vector of the corresponding row in the sparse coefficient matrix. Then, the orthogonality of the four classes of pure signal waveforms obtained by reconstruction is checked to ensure that the separated components satisfy the physical dynamic independence. For example, the tidal driving component should have significant periodic spectral peaks, while the storm erosion component should exhibit non-periodic energy. Finally, the pure signal waveforms corresponding to tides, storms, waves, or human activities are respectively used as the driving component of tides, the erosion component of storms, the impact component of waves, and the disturbance component of human activities.

[0040] It should be noted that, in this application, the driving component refers to the periodic tidal fluctuation signal dominated by celestial tidal forces, used to reveal the regular changes in the coastline controlled by tides; the erosion component refers to the sudden high-energy impact signal triggered by storm events, used to characterize the impact intensity and erosion range of extreme weather on the coastline; the action component refers to the high-frequency fluctuation signal continuously acted by wind-generated waves, used to characterize the dynamic background of sediment transport and shoreline shaping under normal wave conditions; the disturbance component refers to the non-natural vibration or structural change signal generated by human activities, used to identify the interference patterns of human factors such as port shipping and coastal engineering on the coastline; the common function of these components is to decouple the chaotic and mixed original observation data into independent components with clear physical meaning, so that subsequent deep feature extraction can model different dynamic processes separately, thereby significantly improving the accuracy of coastline change prediction and the physical interpretability of disaster early warning.

[0041] In some embodiments, reference Figure 2 The figure is a flowchart illustrating the operation of determining multimodal spatiotemporal features according to some embodiments of this application. In this application, the deep features of the driving component, erosion component, action component, and disturbance component are extracted respectively to obtain the multimodal spatiotemporal features of the decoupled multimodal dataset, which can be achieved by the following steps: The dynamic feature map of tidal forces in the driving components is extracted by time-series modeling; The morphological feature maps of storm erosion in the erosion components were extracted through spatial analysis; Energy feature maps of waves in the active components are extracted by time-frequency transformation; The perturbation feature map of human activity in the perturbation component is extracted by pattern recognition; The dynamic feature map, the morphological feature map, the energy feature map, and the perturbation feature map are adaptively fused to obtain the multimodal spatiotemporal features of the decoupled multimodal dataset.

[0042] In specific implementation, the extraction of the tidal dynamic feature map from the driving component through time-series modeling can be achieved in the following way: the tidal time-series data in the driving component can be decomposed using an improved symplectic geometric mode decomposition method to eliminate endpoint effects and spurious components in the decomposition results, and improve the stability and correlation of each component. Then, the highly correlated intrinsic mode components obtained after decomposition are input into a convolutional neural network combined with time pattern attention and a bidirectional long short-term memory network (e.g., a time pattern attention-convolutional neural network-bidirectional long short-term memory network architecture) for deep feature extraction. That is, the hidden state of relevant historical time steps is adaptively selected through the time pattern attention mechanism to enhance the ability to capture key time patterns. Local fluctuation features are extracted from the reconstructed map using the convolutional neural network, and the bidirectional dependence of the tidal sequence in the time dimension is captured through the bidirectional long short-term memory network to obtain the deep time-series features. Finally, the extracted deep time-series features are aggregated into a tidal dynamic feature map; wherein, the dynamic feature map is a set of deep features characterizing the periodic evolution law of the tide.

[0043] In specific implementation, the extraction of storm erosion morphological feature maps from the erosion component through spatial analysis can be achieved in the following way: First, the optical remote sensing images in the erosion component can be automatically extracted at the sub-pixel level using open-source coastline analysis and extraction tools (such as the coastline analysis and extraction tools developed under the European Coastal Flood Warning System framework) to obtain the instantaneous coastline positions before and after the storm event; then, a spatiotemporal model is constructed based on the extracted instantaneous coastline positions. By comparing the changes in coastline positions before and after the storm, the distance and spatial distribution range of coastline retreat caused by the storm are quantified, enabling spatial heterogeneity analysis of storm erosion intensity. Combined with synchronously acquired wave data (such as data from extreme storm events with significant wave heights exceeding seven meters), the response differences of different coastal sections to storm erosion are analyzed to identify erosion hotspots and extract the spatial distribution patterns of erosion morphology; finally, this spatial distribution pattern is integrated into a storm erosion morphological feature map; wherein, the morphological feature map is a set of deep features characterizing the spatial distribution and intensity characteristics of storm erosion.

[0044] In specific implementation, the extraction of wave energy feature maps from the active component through time-frequency transformation can be achieved in the following way: First, the wave time series signal in the active component can be converted from a one-dimensional time series to a two-dimensional time spectrum map using short-time Fourier transform, so as to simultaneously retain the localization information of the signal in the time domain and frequency domain. Then, an improved deep learning semantic segmentation model (e.g., based on an improved Vgg16-Unet network model) is used to automatically detect and segment the energy accumulation regions in the two-dimensional time spectrum map, accurately extracting the energy distribution regions representing different wave modes from the complex background. Then, frequency features are extracted from all segmented energy distribution regions, that is, linear frequency modulated wavelet transform based on a dispersion model is used to adaptively match the instantaneous frequency and group delay of the wave signal, thereby separating the wave energy under different propagation modes and obtaining the frequency features of different wave modes. Finally, all frequency features are fused into a two-dimensional data matrix as the wave energy spectrum feature map. The energy feature map is a set of deep features characterizing the energy distribution law of waves in different frequency bands.

[0045] In specific implementation, the extraction of disturbance feature maps of human activities in the disturbance component through pattern recognition can be achieved in the following way: First, an object-level change detection network (e.g., an object-level binary classification change detection network based on YOLO v5) can be used to extract the image features of the preceding and following time phases of the optical remote sensing images related to human activities in the disturbance component, and feature fusion is performed to output the human activity change area in the form of a target box; Second, a lightweight scene classification network (e.g., MobileNet v2 network) is used to accurately classify the preceding and following time phase images corresponding to the human activity change area to identify the specific human activity type, such as industrial construction, agricultural development, or aquaculture; Then, a pattern recognition algorithm combined with vibration signals collected by a distributed fiber optic vibration sensing system is used to identify the vibration patterns corresponding to the specified human activities (such as ship navigation and engineering construction), thereby realizing multimodal collaborative identification of nearshore human activities, and integrating the identified human activity type, spatial distribution range, and activity intensity information into a disturbance feature map; wherein, the disturbance feature map is a deep feature set characterizing the type, intensity, and spatial distribution of human activities, used to quantify the degree of impact of human activities on coastline evolution.

[0046] It should be noted that in this application, multimodal spatiotemporal features are high-dimensional unified features that fully characterize the coupled evolution of multiple physical processes along the coastline, and can serve as the basic input for subsequent spatiotemporal modeling and change early warning. Specifically, the adaptive fusion of the dynamic feature map, morphological feature map, energy feature map, and perturbation feature map to obtain the multimodal spatiotemporal features of the decoupled multimodal dataset can be achieved in the following way: the dynamic feature map, morphological feature map, energy feature map, and perturbation feature map can be mapped to a unified feature embedding space through independent encoder networks to eliminate the feature dimensionality and semantic differences between different modalities, and to induce... A dual-level attention mechanism adaptively weights four feature maps in the feature embedding space. This mechanism can automatically select key local features in each feature map at the intra-modal level and dynamically adjust the fusion weights of different physical components at the inter-modal level according to the current task requirements to highlight information sources with higher contribution. A multimodal deep reinforcement learning framework is used to optimize the feature fusion strategy end-to-end. Through the interaction feedback between the agent and the environment, the parameters and methods of feature fusion are adaptively adjusted to cope with the complex and ever-changing marine environment. Finally, the multimodal features optimized by attention weighting and reinforcement learning are vectorized and concatenated to generate multimodal spatiotemporal features of the decoupled multimodal dataset.

[0047] In step 104, the multimodal spatiotemporal features are spatiotemporally heterogeneously modeled to generate an evolution feature tensor of a unified spatiotemporal benchmark for the coastline, and then coastline change monitoring and early warning are performed based on the evolution feature tensor.

[0048] In some embodiments, reference Figure 3 The figure is an exemplary flowchart illustrating the determination of the evolution feature tensor according to some embodiments of this application. In this application, the evolution feature tensor of the unified spatiotemporal reference of the coastline by performing spatiotemporal heterogeneous modeling on the multimodal spatiotemporal features can be achieved by the following steps: In step 1041, an initial spatiotemporal topology graph is established using the coastal zone spatial grid as nodes and the multimodal spatiotemporal features as node attributes; In step 1042, the initial spatiotemporal topology graph is subjected to graph convolutional neural network learning to extract spatially dependent co-evolutionary features; In step 1043, the evolution characteristics of each node in the initial spatiotemporal topology graph along the tidal phase dimension are captured by time-series modeling using gated loop units; In step 1044, the co-evolutionary features are multidimensionally fused with all trend evolutionary features to generate an evolutionary feature tensor of a unified spatiotemporal reference for the coastline.

[0049] In specific implementation, the initial spatiotemporal topology graph can be established using the coastal spatial grid as nodes and the multimodal spatiotemporal features as node attributes in the following manner: The coastal region can be divided into regular grid cells according to a preset spatial resolution, and each grid cell can be used as an independent graph node in the topology graph. At the same time, the multimodal spatiotemporal features corresponding to each graph node at continuous tidal phase values ​​can be used as the original attribute vector of the node, thereby constructing a set of nodes with clear spatial location indices. Then, the correlation strength between nodes can be defined by calculating the spatial Euclidean distance between any two graph nodes. For example, the radial basis function can be used to weight and fuse the spatial Euclidean distance to generate the edge weights in the adjacency matrix, thereby establishing the initial spatiotemporal topology graph. The initial spatiotemporal topology graph refers to graph structure data that represents the intrinsic correlation structure between different spatial locations of the coastal zone. It can explicitly model the intrinsic correlation between different spatial locations of the coastal zone through graph structure, transforming irregular spatial correlation relationships of the coastal zone into structured graph data to support subsequent graph neural network learning, and providing a unified data organization foundation for the mining of spatial dependencies.

[0050] In specific implementation, the initial spatiotemporal topology graph is learned using a graph convolutional neural network (GNN). The extraction of spatially dependent co-evolutionary features can be achieved in the following way: the initial spatiotemporal topology graph and its node attributes are input into the GNN. Graph convolution operations in the graph Fourier domain are used to aggregate and update the neighborhood information of the central node. Specifically, after normalizing the adjacency matrix using the Laplacian matrix, a graph convolution kernel is defined in the spectral domain to weighted aggregate the feature information of adjacent nodes, thereby learning the nonlinear spatial dependencies and co-evolutionary patterns between different spatial locations in the coastal zone. Then, by stacking multiple layers of graph convolutional layers to gradually expand the receptive field, each graph node can aggregate the spatial information of its multi-order neighbors, thereby extracting the spatial propagation patterns and co-evolutionary features of the coastline change process under the interaction of land and sea. The co-evolutionary features refer to node feature representations that reflect the nonlinear correlations and co-change patterns between different spatial locations in the coastal zone. They can capture the spatial propagation patterns of the coastline change process under the interaction of land and sea, enabling subsequent time-series modeling to simultaneously perceive the influence of the neighborhood space.

[0051] In specific implementation, capturing the trend evolution features of each node in the initial spatiotemporal topology graph along the tidal phase dimension through time-series modeling of gated recurrent units can be achieved in the following way: the feature vector sequence obtained by multi-level spatial aggregation of each graph node after learning by a graph convolutional neural network can be organized into a time-series input into the gated recurrent unit network according to the tidal phase time order. Through the update gate and reset gate mechanism in the gated recurrent unit, the fusion ratio of historical state information and current input information is adaptively controlled, thereby effectively capturing the long short-term memory pattern and trend change law of each spatial node feature over time in the continuous tidal phase dimension. Finally, the feature vector encoded by each graph node in the time-series dimension is output as the trend evolution feature. The trend evolution feature is a feature representation reflecting the long short-term memory pattern of each spatial node feature along the tidal phase dimension over time, which can effectively characterize the periodic trends and abrupt events in the process of coastline evolution, and provide accurate time-series evolution information for change prediction.

[0052] It should be noted that, in this application, the evolution feature tensor is a three-dimensional feature representation reflecting the nonlinear correlation, cooperative change patterns, and temporal evolution patterns along the tidal phase dimension between different spatial locations of the coastal zone. It can unify spatiotemporal semantic information into a single data base, providing a standardized feature representation that can be directly input for subsequent coastline change prediction models. In specific implementation, the evolution feature tensor of the unified spatiotemporal benchmark of the coastline can be generated by multidimensionally fusing the cooperative evolution features with all trend evolution features. This can be achieved in the following way: the cooperative evolution features and all trend evolution features can be concatenated as vectors along the feature dimension, so that each graph node carries both spatial neighborhood aggregation information and historical temporal evolution information, forming a fused feature vector with spatiotemporal semantics. Then, according to the spatial grid arrangement order of all graph nodes in the initial spatiotemporal topology graph, these fused feature vectors are reorganized into a three-dimensional tensor form, where the first two dimensions of the tensor correspond to the row and column indices of the spatial grid, and the third dimension corresponds to the fused feature channel dimension, thereby generating the evolution feature tensor of the unified spatiotemporal benchmark of the coastline.

[0053] In some embodiments, coastline change monitoring and early warning based on the evolution feature tensor can be achieved through the following steps: A coastline change prediction model is constructed using a pre-trained long short-term memory network. The evolution feature tensor is input into the change prediction model for prediction, thereby obtaining the early warning level of the coastline change prediction index. The warning level is output for coastline change monitoring and early warning.

[0054] In specific implementation, the coastline change prediction model can be constructed using a pre-trained Long Short-Term Memory (LSTM) network in the following manner: First, a coastline change prediction model architecture can be built based on the LSTM network. This network effectively solves the gradient vanishing problem in traditional recurrent neural networks (RNNs) for long-sequence modeling by introducing a gating mechanism, enabling it to capture the long-term dependencies and nonlinear change characteristics in the coastline evolution process. Second, the constructed LSTM network is trained offline using a historical coastline evolution feature tensor dataset. This dataset contains evolution feature tensors obtained from previous monitoring and their corresponding actual coastline change observations. The network weight parameters are optimized using an error backpropagation algorithm, allowing the model to learn the complex mapping relationship between evolution features and coastline changes. Then, regularization techniques and early stopping mechanisms are introduced during training to prevent overfitting, and the optimal hyperparameter combination is selected using cross-validation. Finally, the trained LSTM network parameters are solidified to form a coastline change prediction model with generalization capabilities. The prediction model is a deep learning engine based on evolution feature tensors that captures long-term dependencies in the coastline to output quantitative indicators of future changes.

[0055] In specific implementation, the evolutionary feature tensor is input into the change prediction model for prediction, and the resulting early warning level of the coastline change prediction index can be achieved in the following way: First, the real-time generated evolutionary feature tensor can be sliced ​​and recombined according to a preset time window length to form a temporal feature sequence that meets the input requirements of the change prediction model. The constructed temporal feature sequence is then input into the trained change prediction model for forward computation. Through its memory unit and gating structure, the temporal dependencies in the feature sequence are automatically extracted, and change prediction indicators such as the coastline position shift or erosion and siltation rate in future periods are output. Then, the generated change prediction index is classified according to a preset multi-level early warning threshold, and the continuous prediction index values ​​are converted into discrete coastline change early warning levels to obtain the early warning level of the coastline change prediction index. The multi-level early warning threshold can be comprehensively determined based on historical disaster statistics and the vulnerability assessment results of coastal disaster-bearing bodies. The early warning level is a discrete risk level that quantifies the coastline change prediction results, which can transform complex prediction results into intuitive decision-making basis to achieve graded response.

[0056] In practice, the output of the warning level for coastline change monitoring and early warning can be achieved in the following way: the corresponding warning color label and risk description text can be matched according to the generated warning level to form standardized warning information, and the warning information can be pushed to relevant user terminals such as coastal zone management departments, coastal communities and marine enterprises in real time through a multi-channel distribution mechanism.

[0057] In another aspect, in some embodiments, this application provides a coastline dynamic monitoring and early warning system, with reference to... Figure 4 The figure is a schematic diagram of the structure of a dynamic monitoring and early warning system for coastlines according to some embodiments of this application. The dynamic monitoring and early warning system 400 for coastlines includes: a data acquisition module 401, a processing module 402, and an execution module 403, which are described below: The acquisition module 401 in this application is mainly used to acquire multimodal raw data of the coastline, determine the tidal phase corresponding to the acquisition time of each modal data in the multimodal raw data, and then resample the multimodal raw data to a unified tidal phase time axis based on all tidal phases to obtain a tidal phase aligned multimodal dataset. Processing module 402, in this application, is mainly used to perform spatial reference unification on the multimodal dataset through UAV cooperative positioning without ground control points, so as to obtain a multimodal spatiotemporal dataset with tidal phase-spatial coordinate joint alignment; It should be noted that the processing module 402 in this application is also used to decompose the multimodal spatiotemporal dataset into the driving component of tides, the erosion component of storms, the action component of waves and the disturbance component of human activities based on the pre-constructed overcomplete dictionary of ocean dynamic processes, and then extract the deep features of the driving component, the erosion component, the action component and the disturbance component respectively to obtain the multimodal spatiotemporal features of the decoupled multimodal dataset. The execution module 403 in this application is mainly used to perform spatiotemporal heterogeneous modeling on the multimodal spatiotemporal features, generate an evolution feature tensor of a unified spatiotemporal benchmark for the coastline, and then perform coastline change monitoring and early warning based on the evolution feature tensor.

[0058] Each module in the aforementioned dynamic monitoring and early warning system for the coastline can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of the computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0059] In another embodiment, this application provides a computer device, which may be a server, and its internal structure diagram may be as follows. Figure 5As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores data related to dynamic coastline monitoring and early warning methods. The network interface communicates with external terminals via a network. When executed by the processor, the computer program implements a dynamic coastline monitoring and early warning method.

[0060] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0061] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described embodiment of the dynamic monitoring and early warning method for coastlines.

[0062] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps described in the above embodiment of the dynamic monitoring and early warning method for coastlines.

[0063] In one embodiment, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps described in the above-described embodiment of the coastline dynamic monitoring and early warning method.

[0064] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0065] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0066] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for dynamic monitoring and early warning of coastlines, characterized in that, Includes the following steps: Collect multimodal raw data of the coastline, determine the tidal phase corresponding to the acquisition time of each modality in the multimodal raw data, and then resample the multimodal raw data to a unified tidal phase time axis based on all tidal phases to obtain a tidal phase aligned multimodal dataset. The multimodal dataset is spatially referenced by UAV collaborative positioning without ground control points to obtain a multimodal spatiotemporal dataset jointly aligned with tidal phase and spatial coordinates. Based on a pre-built overcomplete dictionary of ocean dynamic processes, the multimodal spatiotemporal dataset is decomposed into the driving component of tides, the erosion component of storms, the action component of waves, and the disturbance component of human activities. Then, the deep features of the driving component, erosion component, action component, and disturbance component are extracted respectively to obtain the multimodal spatiotemporal features of the decoupled multimodal dataset. Spatiotemporal heterogeneous modeling is performed on the multimodal spatiotemporal features to generate an evolution feature tensor of a unified spatiotemporal benchmark for the coastline, and then coastline change monitoring and early warning are performed based on the evolution feature tensor.

2. The method as described in claim 1, characterized in that, Determining the tidal phase corresponding to the data acquisition time of each modality in the multimodal raw data specifically includes: For each modal data in the multimodal raw data, tidal harmonic analysis is performed on the acquisition time period of the modal data to obtain the tidal phase angle of each tidal component at each acquisition time. By performing multi-tidal synthesis on all tidal phase angles at the same acquisition time, the comprehensive tidal phase values ​​of the modal data at each acquisition time are obtained; The tidal phase corresponding to the acquisition time of the modal data is determined based on the comprehensive tidal phase value at all acquisition times, thereby obtaining the tidal phase corresponding to the acquisition time of each modal data in the multimodal raw data.

3. The method as described in claim 1, characterized in that, Based on all tidal phases, the original multimodal data is resampled onto a unified tidal phase time axis to obtain a tidal phase-aligned multimodal dataset, specifically including: Construct a regular sequence of phase nodes indexed by the tidal phase angle based on all tidal phases; Using the regular phase node sequence as the target time reference, interpolation is performed on each modal data in the multimodal raw data to obtain synchronous sampling data of each modal data on the regular phase node sequence; All synchronously sampled data are integrated to obtain a tidal phase-aligned multimodal dataset.

4. The method as described in claim 1, characterized in that, Based on a pre-built overcomplete dictionary of ocean dynamic processes, the multimodal spatiotemporal dataset is decomposed into tidal driving components, storm erosion components, wave action components, and human activity disturbance components, specifically including: A pre-constructed, complete dictionary of ocean dynamic processes; The multimodal spatiotemporal dataset is sparsely decomposed on the overcomplete dictionary to obtain the sparse coefficient matrix corresponding to each atom in the overcomplete dictionary; Based on the sparse coefficient matrix, the driving component of tides, the erosion component of storms, the action component of waves, and the disturbance component of human activities are separated from the multimodal spatiotemporal dataset.

5. The method as described in claim 1, characterized in that, The deep features of the driving component, erosion component, action component, and perturbation component are extracted respectively to obtain the multimodal spatiotemporal features of the decoupled multimodal dataset, specifically including: The dynamic feature map of tidal forces in the driving components is extracted by time-series modeling; The morphological feature maps of storm erosion in the erosion components were extracted through spatial analysis; Energy feature maps of waves in the active components are extracted by time-frequency transformation; The perturbation feature map of human activity in the perturbation component is extracted by pattern recognition; The dynamic feature map, the morphological feature map, the energy feature map, and the perturbation feature map are adaptively fused to obtain the multimodal spatiotemporal features of the decoupled multimodal dataset.

6. The method as described in claim 1, characterized in that, Spatiotemporal heterogeneous modeling of the multimodal spatiotemporal features to generate the evolution feature tensor of the unified spatiotemporal reference for the coastline specifically includes: An initial spatiotemporal topology graph is established using the coastal zone spatial grid as nodes and the aforementioned multimodal spatiotemporal features as node attributes; The initial spatiotemporal topology graph is subjected to graph convolutional neural network learning to extract spatially dependent co-evolutionary features; The evolution characteristics of each node in the initial spatiotemporal topology graph along the tidal phase dimension are captured by timing modeling using gated cyclic units. The co-evolutionary features are fused with all trend evolutionary features in a multidimensional manner to generate an evolutionary feature tensor of a unified spatiotemporal benchmark for the coastline.

7. The method as described in claim 1, characterized in that, Coastline change monitoring and early warning based on the aforementioned evolutionary feature tensor specifically includes: A coastline change prediction model is constructed using a pre-trained long short-term memory network. The evolution feature tensor is input into the change prediction model for prediction, thereby obtaining the early warning level of the coastline change prediction index. The warning level is output for coastline change monitoring and early warning.

8. A dynamic monitoring and early warning system for coastlines, characterized in that, include: The acquisition module is used to acquire multimodal raw data of the coastline, determine the tidal phase corresponding to the acquisition time of each modality in the multimodal raw data, and then resample the multimodal raw data to a unified tidal phase time axis based on all tidal phases to obtain a tidal phase aligned multimodal dataset. The processing module is used to perform spatial benchmark unification on the multimodal dataset through UAV collaborative positioning without ground control points, so as to obtain a multimodal spatiotemporal dataset with tidal phase-spatial coordinate joint alignment. The processing module is used to decompose the multimodal spatiotemporal dataset into tidal driving components, storm erosion components, wave action components, and human activity disturbance components based on a pre-built overcomplete dictionary of ocean dynamic processes. Then, it extracts the deep features of the driving components, erosion components, action components, and disturbance components respectively to obtain the multimodal spatiotemporal features of the decoupled multimodal dataset. The execution module is used to perform spatiotemporal heterogeneous modeling on the multimodal spatiotemporal features, generate an evolution feature tensor of a unified spatiotemporal benchmark for the coastline, and then perform coastline change monitoring and early warning based on the evolution feature tensor.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the coastline dynamic monitoring and early warning method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the dynamic monitoring and early warning method for the coastline as described in any one of claims 1 to 7.

Images