An ice avalanche hidden danger intelligent identification method based on multi-source remote sensing and ground cooperation

Abstract

This invention discloses an intelligent identification method for ice avalanche hazards based on multi-source remote sensing and ground-based collaboration. The method includes: fusing optical remote sensing data, microwave remote sensing data, and lidar data as space-based observation data; combining ground meteorological station observation data and UAV-based measured data to form multi-source collaborative monitoring data; using an adaptive spatiotemporal fusion algorithm to reconstruct data gaps caused by cloudy snow cover in the multi-source collaborative monitoring data, generating high spatiotemporal resolution continuous snow cover dynamic monitoring data; constructing an ice avalanche hazard identification feature vector based on the snowline dynamic change index; standardizing the ice avalanche hazard identification feature vector and inputting it into a pre-trained ice avalanche hazard classification model to obtain intelligent identification results for ice avalanche hazard areas, forming a closed loop from data fusion and feature mining to intelligent early warning, significantly improving the accuracy and timeliness of ice avalanche hazard identification.

Description

Technical Field

[0001] This invention relates to the field of multi-source remote sensing technology, specifically to an intelligent identification method for ice avalanche hazards based on multi-source remote sensing and ground coordination. Background Technology

[0002] Ice avalanches are sudden and severe geological disasters in high-altitude glacial areas, possessing immense destructive power and posing a serious threat to downstream infrastructure, residents' safety, and the ecological environment. Therefore, early identification and monitoring of ice avalanche hazards has become an urgent need for disaster prevention and mitigation in high-altitude regions. Currently, methods for identifying ice avalanche hazards mainly rely on remote sensing technology and ground observation, and are gradually developing towards multi-source data fusion and intelligent analysis. Existing technologies mainly suffer from the following problems:

[0003] (1) Existing methods rely heavily on optical remote sensing data. However, in high-altitude and cold mountainous areas, the perennial cloud and snow cover results in a large number of spatiotemporal gaps in optical data. It is difficult to obtain continuous monitoring data during critical periods. There is a lack of a reliable method that can effectively integrate multi-source heterogeneous data, adaptively compensate for data gaps, and generate high spatiotemporal resolution continuous monitoring sequences, resulting in discontinuities and uncertainties in monitoring results.

[0004] (2) The large-scale preliminary screening results of satellite remote sensing lack effective verification and calibration of near-ground high-precision observation data such as UAVs and ground sensors. The key ground parameters are not deeply coupled with remote sensing macro information. This data island mode makes it difficult to substantially improve the accuracy and reliability of hidden danger identification, and is prone to misjudgment or omission. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent identification method for ice avalanche hazards based on multi-source remote sensing and ground coordination, so as to solve the technical problems in the prior art of insufficient spatiotemporal continuity of data coverage, susceptibility to environmental interference, and difficulty in substantially improving the accuracy and reliability of hazard identification.

[0006] To solve the above-mentioned technical problems, the present invention specifically provides the following technical solution:

[0007] This invention provides an intelligent identification method for ice avalanche hazards based on multi-source remote sensing and ground coordination, comprising the following steps:

[0008] By integrating optical remote sensing data, microwave remote sensing data, and lidar data as space-based observation data, and combining ground meteorological station observation data with UAV field measurement data, multi-source collaborative monitoring data is formed.

[0009] An adaptive spatiotemporal fusion algorithm is used to reconstruct the data missing areas caused by multi-cloud snow cover using the multi-source collaborative monitoring data, generating continuous snow cover dynamic monitoring data with high spatiotemporal resolution;

[0010] The continuous snow cover dynamic monitoring data is processed to automatically extract and output the snow cover start time, snow cover end time, snow water equivalent change trend map and snow line dynamic change index;

[0011] Based on the aforementioned snowline dynamic change index, and combined with glacier spatial geometric parameters, glacier surface characteristic parameters, glacier dynamic state parameters, topographically derived factors, meteorological driving factors, meltwater infiltration index, seismic activity parameters, and spatial density of historical ice avalanche events, an ice avalanche hazard identification feature vector is constructed.

[0012] After standardizing the feature vector for identifying ice avalanche hazards, it is input into a pre-trained ice avalanche hazard classification model to obtain intelligent identification results for ice avalanche hazard areas.

[0013] As a preferred embodiment of the present invention, optical remote sensing data, microwave remote sensing data, and lidar data are integrated as space-based observation data, combined with ground meteorological station observation data and UAV measured data to form multi-source collaborative monitoring data, including:

[0014] Space-based observation data is composed of optical remote sensing satellite imagery, synthetic aperture radar satellite imagery, and spaceborne lidar data.

[0015] Airborne observation data are constructed from remote sensing data acquired by UAVs equipped with optical cameras, radar, or lidar.

[0016] Ground-based observation data are constructed using temperature, precipitation, and radiation data collected from ground meteorological stations, as well as measured snow depth and snow density data.

[0017] Based on a unified spatiotemporal reference, three types of observation data—space-based observation data, airborne observation data, and ground-based observation data—are registered and fused to form a multi-source collaborative monitoring dataset that covers the target area and has different spatiotemporal resolutions and physical characteristics.

[0018] As a preferred embodiment of the present invention, observation data are registered and fused based on a unified spatiotemporal reference to form a multi-source collaborative monitoring dataset covering the target area and having different spatiotemporal resolutions and physical characteristics, including:

[0019] The space-based observation data and air-based observation data are uniformly resampled to a specified spatial resolution, and the spatial coordinates of the ground-based observation points are precisely correlated with the corresponding pixels;

[0020] Using Coordinated Universal Time as a unified time reference, the time labels of each observation data are standardized, and for data with different observation frequencies, time interpolation methods are used to generate datasets with the same time series intervals.

[0021] By employing a statistical regression model, the observations obtained from different sensors are uniformly converted into key physical parameters of snow cover, establishing a physical consistency relationship among multi-source parameters and forming a multi-dimensional collaborative monitoring dataset.

[0022] As a preferred embodiment of the present invention, an adaptive spatiotemporal fusion algorithm is used to reconstruct the data missing areas caused by cloud cover and snow cover from the multi-source collaborative monitoring data, including:

[0023] Cloud and snow mask identification is performed on the temporal optical remote sensing images of the target area to mark areas with missing data;

[0024] Based on the spatiotemporal resolution, physical characteristics, and data quality of each data source, dynamic weights are assigned to each type of observation data, where:

[0025] The areas in optical images without cloud or snow coverage are assigned the highest weight.

[0026] For microwave SAR imagery, based on its advantage in penetrating clouds and snow, higher weights are assigned to areas lacking optical data.

[0027] High-precision correction weights are assigned to the areas covered by the observation phase for the measured data of lidar and UAVs.

[0028] For regions lacking optical data, a fusion model based on spatiotemporal spectral similarity is constructed.

[0029] As a preferred embodiment of the present invention, the fusion model based on spatiotemporal spectral similarity generates high spatiotemporal resolution continuous snow cover dynamic monitoring data, including:

[0030] In the spatial dimension, high-resolution UAV data or historical cloudless images from the same period are used to fill in detailed information through spatial interpolation and texture transfer methods.

[0031] In the time dimension, by leveraging the high temporal continuity of microwave SAR time series data, a time series variation model of snow cover parameters is established, and the snow cover status of missing periods is recovered through deep learning time series prediction methods.

[0032] By fusing high-precision terrain and snow depth information obtained from lidar, the interpolation and prediction results are corrected using three-dimensional spatial constraints.

[0033] The snowfall and snowmelt process parameters measured by ground meteorological stations are used as boundary conditions, and the fusion results are optimized through energy balance equations.

[0034] An iterative optimization algorithm is used to adaptively adjust the fusion parameters within the effective observation area, with the goal of minimizing the error between the reconstructed data and the actual observation data, to obtain dynamic monitoring data of snow cover.

[0035] As a preferred embodiment of the present invention, the continuous snow cover dynamic monitoring data is processed to automatically extract and output the snow cover start time, snow cover end time, snow water equivalent change trend graph, and snow line dynamic change index, including:

[0036] The dynamic monitoring data of snow cover is processed, and a sliding window smoothing algorithm is used to obtain the time series curve of snow cover range;

[0037] Set a threshold for the proportion of snow cover area. When the proportion of snow cover area exceeds the threshold T for N consecutive days, the first day of the period will be determined as the start date of snow accumulation.

[0038] When the snow cover area ratio is lower than the threshold T for M consecutive days, the first day of that period is determined as the end date of snow cover.

[0039] Based on the time series of the snow cover dynamic monitoring data, calculate the cumulative snow water equivalent curve for each pixel throughout the entire snow cover period;

[0040] The key inflection points of the snow water equivalent accumulation, stabilization and ablation stages were extracted by piecewise linear fitting method, and a spatiotemporal variation trend map of snow water equivalent at the regional scale was generated.

[0041] As a preferred embodiment of the present invention, based on the aforementioned snowline dynamic change index, and combined with glacier spatial geometric parameters, glacier surface characteristic parameters, glacier dynamic state parameters, topographically derived factors, meteorological driving factors, meltwater infiltration index, seismic activity parameters, and spatial density of historical ice avalanche events, an ice avalanche hazard identification feature vector is constructed, including:

[0042] The characteristic parameters of glacier spatial geometry, glacier surface features, glacier dynamic state, topographic factors, meteorological driving factors, meltwater infiltration index, seismic activity parameters, and spatial density of historical glacial avalanches are unified to the same spatiotemporal scale.

[0043] Based on the aforementioned dynamic change index of the snow line, the target area is divided into a regular grid, and the point and line parameters are aggregated to the grid scale through spatial interpolation.

[0044] Using the snow season as the time unit, parameters from different observation frequencies are aggregated to the same time resolution through time series analysis.

[0045] The normalized multi-source feature parameters are arranged in a fixed order to construct a multi-dimensional feature vector corresponding to each spatial grid.

[0046] As a preferred embodiment of the present invention, the feature vector for identifying ice avalanche hazards is standardized, including:

[0047] The Z-score normalization method is used to calculate the mean and standard deviation of each feature dimension based on the training sample set, and the input feature vector for ice avalanche hazard identification is normalized as follows:

[0048]

[0049] in, Represents the original feature values. This represents the mean of the features in the training set. Indicates standard deviation, The value is represented by the standard.

[0050] For feature dimensions with missing values, the missing values ​​are filled in using the median after normalizing the features in the training set.

[0051] As a preferred embodiment of the present invention, the standardized ice avalanche hazard identification feature vector is input into a pre-trained ice avalanche hazard classification model, including:

[0052] The standardized ice avalanche hazard identification feature vector is input into a pre-trained ice avalanche hazard classification model, which is a support vector machine (SVM) model with radial basis function (RBF) as the kernel function. Its decision function expression is as follows:

[0053]

[0054] in, This represents the standardized feature vector of the input. Represents support vectors, This represents the category label corresponding to the support vector. Represents the Lagrange multipliers. Indicates the bias term. This represents the RBF kernel function.

[0055] As a preferred embodiment of the present invention, the intelligent identification result of the ice avalanche hazard area is obtained based on the discrimination result of the decision function output value, including:

[0056] like If so, the corresponding glacier unit or spatial grid is determined to be a potential ice avalanche zone;

[0057] like If so, it is determined to be a stable region;

[0058] After processing all input grids, a spatial distribution map of ice avalanche hazards at the regional scale is generated, along with a confidence score and a list of key contributing features for each hazard area.

[0059] Compared with the prior art, the present invention has the following advantages:

[0060] This invention overcomes the data gap problem in high-altitude, cold mountainous areas with cloudy and snowy environments by constructing an integrated air-space-ground collaborative monitoring system and an adaptive spatiotemporal fusion algorithm. It achieves continuous dynamic monitoring of snow cover with high spatiotemporal resolution. The system automatically extracts key indicators such as snow phenology, snow water equivalent, and snowline dynamics, and constructs a feature vector with complete mechanisms by combining multiple disaster-causing factors. This feature vector is then input into an optimized support vector machine model for intelligent discrimination. Finally, it outputs a decision function that integrates the spatial distribution map of potential hazards, risk confidence, and key causes, forming a closed loop from data fusion and feature mining to intelligent early warning. This significantly improves the accuracy and timeliness of ice avalanche hazard identification. Attached Figure Description

[0061] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0062] Figure 1 The flowchart illustrates the intelligent identification method for ice avalanche hazards based on multi-source remote sensing and ground coordination, as provided in this embodiment of the invention. Detailed Implementation

[0063] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0064] like Figure 1 As shown, this invention provides an intelligent identification method for ice avalanche hazards based on multi-source remote sensing and ground coordination, comprising the following steps:

[0065] By integrating optical remote sensing data, microwave remote sensing data, and lidar data as space-based observation data, and combining ground meteorological station observation data with UAV field measurement data, multi-source collaborative monitoring data is formed.

[0066] An adaptive spatiotemporal fusion algorithm is used to reconstruct the data missing areas caused by multi-cloud snow cover using the multi-source collaborative monitoring data, generating continuous snow cover dynamic monitoring data with high spatiotemporal resolution;

[0067] In this embodiment, an adaptive spatiotemporal fusion algorithm is adopted, which utilizes the penetrability of microwave data and the high precision of lidar to dynamically compensate for the lack of optical data, generating a continuous dynamic sequence of snow cover with high spatiotemporal resolution. This fundamentally overcomes the limitations of traditional optical remote sensing in high-altitude and cold mountainous areas, ensuring the integrity of key phenological data.

[0068] In this embodiment, the adaptive spatiotemporal fusion algorithm employs a pre-trained model based on a three-dimensional convolutional neural network. This neural network is specifically designed to capture long-term temporal features such as the start / end time of snow accumulation and the trend of snow water equivalent changes. These long-term temporal features are used as the basic data for the adaptive spatiotemporal fusion algorithm, generating high spatiotemporal resolution continuous data as the main input stream, thus ensuring the integrity of the time series. This enables the model to distinguish between surface melting caused by short-term temperature fluctuations and deep deformation caused by long-term cumulative effects, thereby identifying potential precursor signals.

[0069] Convolutional layers are used to extract the spatial patterns of dynamic changes in the snow line, such as the retreat rate of the snow line and the irregular retreat. These spatial patterns are often directly related to the terrain undulations and sub-ice topography below. The terrain undulations and sub-ice topography below are acquired by lidar. By fusing these two, stress concentration areas controlled by the terrain can be identified.

[0070] When the pre-trained model encounters key areas of uncertainty during training, it will backtrack and call the original UAV measured data for local verification, or refer to the observation data of ground weather stations to correct the feature map of the intermediate layer of the model. This ensures that the final intelligent recognition result is a consistent output of the multi-source evidence chain of sky-air-ground in the internal logic of the model, rather than the interpretation result of a single remote sensing data.

[0071] The continuous snow cover dynamic monitoring data is processed to automatically extract and output the snow cover start time, snow cover end time, snow water equivalent change trend map and snow line dynamic change index;

[0072] Based on the aforementioned snowline dynamic change index, and combined with glacier spatial geometric parameters, glacier surface characteristic parameters, glacier dynamic state parameters, topographically derived factors, meteorological driving factors, meltwater infiltration index, seismic activity parameters, and spatial density of historical ice avalanche events, an ice avalanche hazard identification feature vector is constructed.

[0073] In this embodiment, a feature vector is constructed, which breaks through the limitations of relying only on a few remote sensing-derived parameters such as deformation and cracks. It systematically integrates snowline dynamics, multi-source disaster-causing factors, and historical disaster background. This integration enables the identification model to not only "see" the surface changes of glaciers, but also "understand" the internal driving mechanism of their instability and external environmental stress, which greatly improves the scientific nature and foresight of the identification.

[0074] After standardizing the feature vector for identifying ice avalanche hazards, it is input into a pre-trained ice avalanche hazard classification model to obtain intelligent identification results for ice avalanche hazard areas.

[0075] In this embodiment, by deeply coordinating the large-scale coverage of satellites, the localized and refined monitoring by UAVs, and the continuous measured data from ground stations, the monitoring range is not only expanded, but the remote sensing inversion results are also verified and calibrated using ground data. This significantly improves the credibility and accuracy of the entire monitoring system, forming a complete chain of evidence from macro to micro and from indirect to direct.

[0076] By integrating optical remote sensing data, microwave remote sensing data, and lidar data as space-based observation data, and combining them with ground-based meteorological station observation data and UAV field measurement data, multi-source collaborative monitoring data is formed, including:

[0077] Space-based observation data is composed of optical remote sensing satellite imagery, synthetic aperture radar satellite imagery, and spaceborne lidar data.

[0078] In this embodiment, optical remote sensing data provides high spatial resolution texture, color, and morphological information, which is the basis for identifying surface features such as fissures, meltwater channels, and glacial lakes. Microwave remote sensing data has all-weather and all-day working capabilities, can penetrate clouds and a certain degree of snow layer, and can monitor glacier surface deformation and obtain snow moisture / snow water equivalent. LiDAR data provides three-dimensional terrain and surface elevation information with centimeter-level to meter-level accuracy, which is the most direct means of quantifying ice cliff height, ice surface change rate, and spatial distribution of snow depth.

[0079] Airborne observation data are constructed from remote sensing data acquired by UAVs equipped with optical cameras, radar, or lidar.

[0080] Ground-based observation data are constructed using temperature, precipitation, and radiation data collected from ground meteorological stations, as well as measured snow depth and snow density data.

[0081] Based on a unified spatiotemporal reference, three types of observation data—space-based observation data, airborne observation data, and ground-based observation data—are registered and fused to form a multi-source collaborative monitoring dataset that covers the target area and has different spatiotemporal resolutions and physical characteristics.

[0082] In this embodiment, three types of observation data—space-based observation data, air-based observation data, and ground-based observation data—are integrated to ensure that data from different sources are geometrically precisely aligned, avoiding information misalignment. Time synchronization enables observations of different frequencies to be integrated onto a unified time axis for analyzing dynamic processes. The resulting multi-source collaborative monitoring dataset has a much higher accuracy and reliability than any product from a single data source.

[0083] Based on a unified spatiotemporal reference, observation data are registered and fused to form a multi-source collaborative monitoring dataset covering the target area and possessing different spatiotemporal resolutions and physical characteristics, including:

[0084] The space-based observation data and air-based observation data are uniformly resampled to a specified spatial resolution, and the spatial coordinates of the ground-based observation points are precisely correlated with the corresponding pixels;

[0085] In this embodiment, by resampling to a specified resolution and accurately associating with coordinates, it is ensured that satellite pixels, UAV pixels and ground measurement points are strictly corresponding in spatial position, fundamentally avoiding overlay analysis errors caused by inconsistent data spatial benchmarks, and ensuring reliable fusion and comparison of multi-source information in spatial dimensions.

[0086] Using Coordinated Universal Time as a unified time reference, the time labels of each observation data are standardized, and for data with different observation frequencies, time interpolation methods are used to generate datasets with the same time series intervals.

[0087] By employing a statistical regression model, the observations obtained from different sensors are uniformly converted into key physical parameters of snow cover, establishing a physical consistency relationship among multi-source parameters and forming a multi-dimensional collaborative monitoring dataset.

[0088] In this embodiment, Coordinated Universal Time (UTC) is used for time synchronization, and time interpolation is combined to "align" all data to the same time series grid. This allows high-frequency UAV data, daily weather station data, and low-frequency satellite data to be comprehensively analyzed within a unified time frame, thereby seamlessly reconstructing the continuous dynamic process of snow accumulation and melting. This overcomes the limitation of traditional methods, which can only perform snapshot-style comparisons due to data time asynchrony.

[0089] An adaptive spatiotemporal fusion algorithm is used to reconstruct data missing areas caused by cloud cover and snow cover in the multi-source collaborative monitoring data, including:

[0090] Cloud and snow mask identification is performed on the temporal optical remote sensing images of the target area to mark areas with missing data;

[0091] Based on the spatiotemporal resolution, physical characteristics, and data quality of each data source, dynamic weights are assigned to each type of observation data, where:

[0092] The areas in optical images without cloud or snow coverage are assigned the highest weight.

[0093] For microwave SAR imagery, based on its advantage in penetrating clouds and snow, higher weights are assigned to areas lacking optical data.

[0094] High-precision correction weights are assigned to the areas covered by the observation phase for the measured data of lidar and UAVs.

[0095] For regions lacking optical data, a fusion model based on spatiotemporal spectral similarity is constructed.

[0096] In this embodiment, the spatiotemporal spectral similarity-based fusion model consists of three parts: an input layer, an adaptive convolutional neural network module, and an output layer.

[0097] The input layer data selects high-quality optical images with no cloud or snow cover that are closest to the target reconstruction time as the spectral reference. At the same time, it aligns with the lidar point cloud derived data of the same time phase. This invention selects UAV orthophotos as point cloud derived data to provide high-precision geometric structure and texture priors. It also uses terrain factors derived from the digital elevation model (DEM) and the latitude and longitude coordinates of pixels to obtain the time to constrain the spatiotemporal variation of ground features during the fusion process.

[0098] The adaptive convolutional neural network module mines the nonlinear mapping relationship from auxiliary data to the target spectrum through deep learning networks, and performs local weighted optimization using similar pixels, specifically:

[0099] For any missing pixel p in the optical image at time T1, in the high-quality optical image at time T0, combined with DEM data, search for non-neighboring similar pixels with similar spectral and topographic features to p.

[0100] The input layer of the adaptive convolutional neural network module combines the microwave SAR image at time T1, the optical image at time T0, and the DEM data into a multi-channel input tensor.

[0101] A convolutional neural network with an encoder-decoder structure is used. In the encoder part, texture structure features of SAR data and spectral features of optical data are extracted simultaneously through 3D convolution. In the attention mechanism part, a dynamic weight layer based on data quality is introduced. This layer automatically adjusts the contribution of SAR features and optical features during fusion according to the weights assigned by the data source. In cloud and snow-covered areas, the gradient weights of SAR features are automatically amplified by the network. The fusion layer fuses the depth features extracted by the encoder and performs upsampling and spectral reconstruction through transposed convolution.

[0102] Using UAV measured data as a hard constraint, after the model initially generates the fused image at time T1, the residual between the fusion result and the measured high-precision data is calculated using UAV data from the covered area. Spatial interpolation correction is performed on the fusion residual at full resolution to finally correct the fusion result, ensuring that the absolute accuracy of the points with measured data meets the standard and improving the geometric fineness of the overall result.

[0103] The output layer, based on the reconstructed spectral data and combined with the microwave inversion model, outputs pixel-by-pixel snow cover, snow depth, and snow water equivalent products. Especially in areas where optical data is completely missing, it mainly relies on the pseudo-optical inversion results of microwave data after deep network transformation, ensuring the continuity of the time series.

[0104] In this embodiment, by automatically identifying cloud and snow masks, the effective information gaps in time-series optical data can be located accurately and efficiently. This allows algorithm resources to be concentrated on the areas that need to be filled most, avoiding indiscriminate global calculations and significantly improving processing efficiency.

[0105] In this embodiment, weights are dynamically allocated based on the advantages of the data sources to achieve the optimal combination. The dynamic weight allocation strategy based on physical characteristics and data quality is the core manifestation of the algorithm's adaptive capability, which maximizes the combination of the advantages of multi-source data and ensures the overall optimality of the reconstruction results under different regions and conditions.

[0106] The fusion model based on spatiotemporal spectral similarity generates high spatiotemporal resolution continuous snow cover dynamic monitoring data, including:

[0107] In the spatial dimension, high-resolution UAV data or historical cloudless images from the same period are used to fill in detailed information through spatial interpolation and texture transfer methods.

[0108] In this embodiment, high-resolution UAV data or historical cloudless imagery is used to fill in pixel values ​​and reconstruct the real spatial structure and texture features of ground features through spatial interpolation and texture transfer. This avoids the image blurring or homogenization problems caused by simple interpolation, making the generated snow cover range map and snow depth distribution map closer to the real scene in terms of visual and statistical characteristics. This is especially important for identifying micro-topography such as ice surface cracks and meltwater gullies.

[0109] In the time dimension, by leveraging the high temporal continuity of microwave SAR time series data, a time series variation model of snow cover parameters is established, and the snow cover status of missing periods is recovered through deep learning time series prediction methods.

[0110] In this embodiment, the deep learning-based time series prediction method is specifically as follows:

[0111] The input data for time series prediction in deep learning is: taking the target pixel as the center, extracting the microwave SAR multi-polarization backscattering coefficients of the past N time phases, and extracting the snow parameters retrieved from sparse optical remote sensing images without cloud and snow cover and the accurate snow depth point cloud data obtained by lidar within the same time series window.

[0112] The time series forecasting process is as follows:

[0113] First, the input microwave SAR time series feature sequence is normalized. At the same time, a missing mask matrix is ​​constructed to mark which time phases have optical / liquid radar ground truth and which time phases are completely dependent on SAR.

[0114] The encoded temporal feature sequences are input into the forward long short-term memory network LSTM and the backward long short-term memory network LSTM, respectively.

[0115] The forward LSTM learns the cumulative effect of snow accumulation from the past to the present, capturing the "impact of historical state on current state", such as how previous precipitation affects the current snow layer thickness;

[0116] Reverse LSTM learns the backward dependencies from the future to the past, capturing "the correction of previous states by subsequent observations", for example, inferring the snow melt rate from the date of snow melt;

[0117] A temporal attention layer is introduced above the hidden state of the bidirectional LSTM output. This layer automatically learns to assign different weights to different time steps.

[0118] Using a sequence-to-sequence structure, the encoder compresses the input sequence into a context vector, and the decoder predicts the snow parameter sequence for the target time period, i.e. the time period when optical data is missing, based on this vector. During the decoding process, whenever a time point with lidar or ground measurement data is encountered, the model will force a state reset, correct the predicted value to the measured value, and use this as a new starting point to continue predicting.

[0119] The output data of time series prediction by deep learning is as follows: for each spatial cell, the output is the binary values ​​of snow depth, snow water equivalent or snow cover in a continuous time series, and the confidence interval of each predicted value is also output. The confidence interval is wider for the time period further away from the effective observation, which objectively reflects the uncertainty of the prediction and provides risk-controllable input data for subsequent identification of ice avalanche hazards.

[0120] In this embodiment, a time series change model is established using microwave SAR time series data, and a deep learning time series prediction method is used. This makes the reconstruction not only dependent on spatially neighboring pixels, but also follows the physical evolution law and historical memory of snow accumulation and melting at that location. Especially for long-lasting cloud cover periods, the snow status of the missing period can be more reliably inferred based on effective observations before and after the period, ensuring the rationality of the entire time series in terms of physical process.

[0121] By fusing high-precision terrain and snow depth information obtained from lidar, the interpolation and prediction results are corrected using three-dimensional spatial constraints.

[0122] In this embodiment, high-precision terrain and snow depth information from lidar are fused for three-dimensional spatial correction to ensure that the reconstructed snow parameters strictly follow the physical laws of the actual terrain slope and aspect. This provides the most robust spatial geometric constraints from direct observation for the reconstruction results, effectively correcting the physical inconsistencies that may be generated by two-dimensional plane interpolation.

[0123] The snowfall and snowmelt process parameters measured by ground meteorological stations are used as boundary conditions, and the fusion results are optimized through energy balance equations.

[0124] An iterative optimization algorithm is used to adaptively adjust the fusion parameters within the effective observation area, with the goal of minimizing the error between the reconstructed data and the actual observation data, to obtain dynamic monitoring data of snow cover.

[0125] In this embodiment, the iterative optimization algorithm is specifically as follows:

[0126] The core parameters controlling the contribution of each data source in the adaptive spatiotemporal fusion model are used as the data input for iterative optimization. An iterative optimization algorithm based on a hybrid strategy of Bayesian optimization and gradient descent is adopted to find a set of optimal fusion parameters that minimize the error between the reconstructed data and the real observation data, while satisfying the physical process constraints.

[0127] The processing flow is as follows:

[0128] Using the current combination of fusion parameters as input, drive the entire adaptive spatiotemporal fusion model to run once and generate complete dynamic monitoring data of snow cover;

[0129] Within the effective observation area, the reconstructed data generated by the model is compared pixel by pixel with the real observation data;

[0130] Calculate the gradient of the composite loss function with respect to each fusion parameter to determine which parameter adjustments can most effectively reduce the error;

[0131] By combining the global search capability of Bayesian optimization, multiple sets of candidate parameter combinations are sampled in the neighborhood of the current parameter to avoid getting trapped in local optima. Especially in complex terrain areas with high parameter space dimensionality and strong nonlinearity, this hybrid strategy can effectively explore the optimal solution region.

[0132] After iterative optimization, the output is the optimal fusion parameters that are automatically learned and determined based on the current monitoring area and the current seasonal characteristics. These parameters enable the model to achieve the best fusion effect in this complex environment, realizing the adaptive capability of the algorithm.

[0133] In this embodiment, an iterative optimization algorithm is adopted to minimize the error between the reconstructed data and the real observation data in the effective observation area. The internal parameters of the fusion model are automatically adjusted, which forms an adaptive learning closed loop of fusion-verification-correction. It can learn and adjust the optimal fusion strategy according to the characteristics of different regions and seasons, so as to maintain high accuracy in the complex and ever-changing glacial environment.

[0134] The continuous snow cover dynamic monitoring data is processed to automatically extract and output the snow cover start time, snow cover end time, snow water equivalent change trend map, and snow line dynamic change index, including:

[0135] The dynamic monitoring data of snow cover is processed, and a sliding window smoothing algorithm is used to obtain the time series curve of snow cover range;

[0136] Set a threshold for the proportion of snow cover area. When the proportion of snow cover area exceeds the threshold T for N consecutive days, the first day of the period will be determined as the start date of snow accumulation.

[0137] When the snow cover area ratio is lower than the threshold T for M consecutive days, the first day of that period is determined as the end date of snow cover.

[0138] In this embodiment, by setting a clear threshold T and consecutive days standards N and M, the identification rules for the start and end of snow cover are strictly defined. This completely replaces the subjectivity and inconsistency of traditional visual interpretation or personal experience judgment, ensuring the objectivity, fairness and spatiotemporal comparability of phenological parameter extraction results in different years in the same region or between different regions.

[0139] Based on the time series of the snow cover dynamic monitoring data, calculate the cumulative snow water equivalent curve for each pixel throughout the entire snow cover period;

[0140] The key inflection points of the snow water equivalent accumulation, stabilization and ablation stages were extracted by piecewise linear fitting method, and a spatiotemporal variation trend map of snow water equivalent at the regional scale was generated.

[0141] In this embodiment, a sliding window smoothing algorithm is used to preprocess the time series curve, which effectively filters out short-term drastic fluctuations caused by brief weather events. This allows the algorithm to robustly capture the essential and continuous changing trends of snow cover, avoid misjudgments caused by occasional fluctuations, and extract start / end dates that better reflect the true seasonal transition information.

[0142] Based on the aforementioned snowline dynamic change indicators, and combined with glacier spatial geometric parameters, glacier surface characteristic parameters, glacier dynamic state parameters, topographically derived factors, meteorological driving factors, meltwater infiltration index, seismic activity parameters, and spatial density of historical ice avalanche events, an ice avalanche hazard identification feature vector is constructed, including:

[0143] The characteristic parameters of glacier spatial geometry, glacier surface features, glacier dynamic state, topographic factors, meteorological driving factors, meltwater infiltration index, seismic activity parameters, and spatial density of historical glacial avalanches are unified to the same spatiotemporal scale.

[0144] Based on the aforementioned dynamic change index of the snow line, the target area is divided into a regular grid, and the point and line parameters are aggregated to the grid scale through spatial interpolation.

[0145] Using the snow season as the time unit, parameters from different observation frequencies are aggregated to the same time resolution through time series analysis.

[0146] The normalized multi-source feature parameters are arranged in a fixed order to construct a multi-dimensional feature vector corresponding to each spatial grid.

[0147] In this embodiment, a regular grid is used as the basic analysis unit, rather than the traditional entire glacier or sub-basin, which achieves higher spatial resolution for hazard characterization. By spatial interpolation, point and line data are reasonably diffused to the surface, reflecting the spatial heterogeneity of environmental factors more precisely. This avoids the problem of local high-risk signals being averaged out and submerged due to treating the entire glacier as a homogeneous body.

[0148] The feature vector for identifying ice avalanche hazards is standardized, including:

[0149] The Z-score normalization method is used to calculate the mean and standard deviation of each feature dimension based on the training sample set, and the input feature vector for ice avalanche hazard identification is normalized as follows:

[0150]

[0151] in, Represents the original feature values. This represents the mean of the features in the training set. Indicates standard deviation, The value is represented by the standard.

[0152] For feature dimensions with missing values, the missing values ​​are filled in using the median after normalizing the features in the training set.

[0153] In this embodiment, Z-score standardization completely eliminates the influence of units and absolute numerical values ​​by transforming each feature into a distribution with a mean of 0 and a standard deviation of 1. This ensures that all features start on a completely equal footing when the model learns, guaranteeing that the model can learn fairly based on the true correlation between each feature and the target, rather than being misled by numerical values.

[0154] The standardized feature vector for identifying ice avalanche hazards is input into a pre-trained ice avalanche hazard classification model, including:

[0155] The standardized ice avalanche hazard identification feature vector is input into a pre-trained ice avalanche hazard classification model, which is a support vector machine (SVM) model with radial basis function (RBF) as the kernel function. Its decision function expression is as follows:

[0156]

[0157] in, This represents the standardized feature vector of the input. Represents support vectors, This represents the category label corresponding to the support vector. Represents the Lagrange multipliers. Indicates the bias term. This represents the RBF kernel function.

[0158] In this embodiment, the processing flow of the ice avalanche hazard classification model is as follows:

[0159] The input layer is responsible for receiving the standardized feature vector for ice avalanche hazard identification. This vector is a sample point in a multi-dimensional feature space, and its dimension is determined by multiple sources of features such as the snowline dynamic index, glacier geometric parameters, and meteorological driving factors.

[0160] The model training layer is as follows: the feature vector for identifying ice avalanche hazards is standardized to eliminate the influence of different dimensions between features and prevent features with large magnitudes from dominating the similarity calculation of the kernel function;

[0161] A penalty parameter C is set to control the severity of the penalty for misclassified samples, and Bayesian optimization is used within a preset parameter space. The process involves iterating through the training set and performing cross-validation on each set of parameters. The average classification accuracy on the validation set is used as the evaluation metric, and the set with the best performance is selected as the final model parameters.

[0162] After training, the final model is evaluated using a test set that has never been used for training and validation. Indicators such as accuracy and recall are calculated to objectively measure the model's recognition ability in practical applications. Only models whose test set performance meets the preset threshold can be deployed for actual intelligent identification of ice avalanche hazards.

[0163] In this embodiment, a support vector machine with radial basis function kernel is used. Based on the principle of structural risk minimization, by finding the maximum margin hyperplane, overfitting can be effectively avoided, and the generalization ability is strong. Moreover, the RBF kernel function can map linearly inseparable features to a high-dimensional space, and has excellent nonlinear processing ability, making it very suitable for characterizing the complex process of multi-factor nonlinear coupling such as glacier instability.

[0164] Based on the decision function output value, the intelligent identification result of the ice avalanche hazard area is obtained, including:

[0165] like If so, the corresponding glacier unit or spatial grid is determined to be a potential ice avalanche zone;

[0166] like If so, it is determined to be a stable region;

[0167] After processing all input grids, a spatial distribution map of ice avalanche hazards at the regional scale is generated, along with a confidence score and a list of key contributing features for each hazard area.

[0168] In this embodiment, the decision function provides a clear discrimination boundary based on the symbol function, avoiding the ambiguity in interpretation that may be caused by vague or probabilistic outputs, and providing a clear and decisive "yes / no" judgment for business-oriented early warning, which facilitates rapid response from management departments.

[0169] This invention overcomes the data gap problem in high-altitude, cold mountainous areas with cloudy and snowy environments by constructing an integrated air-space-ground collaborative monitoring system and an adaptive spatiotemporal fusion algorithm. It achieves continuous dynamic monitoring of snow cover with high spatiotemporal resolution. The system automatically extracts key indicators such as snow phenology, snow water equivalent, and snowline dynamics, and constructs a feature vector with complete mechanisms by combining multiple disaster-causing factors. This feature vector is then input into an optimized support vector machine model for intelligent discrimination. Finally, it outputs a decision function that integrates the spatial distribution map of potential hazards, risk confidence, and key causes, forming a closed loop from data fusion and feature mining to intelligent early warning. This significantly improves the accuracy and timeliness of ice avalanche hazard identification.

[0170] The above embodiments are merely exemplary embodiments of this application and are not intended to limit this application. The scope of protection of this application is defined by the claims. Those skilled in the art can make various modifications or equivalent substitutions to this application within its substance and scope of protection, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of this application.

Claims

1. A method for intelligent identification of ice avalanche hazards based on multi-source remote sensing and ground coordination, characterized in that, Includes the following steps: By integrating optical remote sensing data, microwave remote sensing data, and lidar data as space-based observation data, and combining ground meteorological station observation data with UAV field measurement data, multi-source collaborative monitoring data is formed. An adaptive spatiotemporal fusion algorithm is used to reconstruct the data missing areas caused by multi-cloud snow cover using the multi-source collaborative monitoring data, generating continuous snow cover dynamic monitoring data with high spatiotemporal resolution; The continuous snow cover dynamic monitoring data is processed to automatically extract and output the snow cover start time, snow cover end time, snow water equivalent change trend map and snow line dynamic change index; Based on the aforementioned snowline dynamic change index, and combined with glacier spatial geometric parameters, glacier surface characteristic parameters, glacier dynamic state parameters, topographically derived factors, meteorological driving factors, meltwater infiltration index, seismic activity parameters, and spatial density of historical ice avalanche events, an ice avalanche hazard identification feature vector is constructed. After standardizing the feature vector for identifying ice avalanche hazards, it is input into a pre-trained ice avalanche hazard classification model to obtain intelligent identification results for ice avalanche hazard areas.

2. The intelligent identification method for ice avalanche hazards based on multi-source remote sensing and ground coordination according to claim 1, characterized in that, By integrating optical remote sensing data, microwave remote sensing data, and lidar data as space-based observation data, and combining them with ground-based meteorological station observation data and UAV field measurement data, multi-source collaborative monitoring data is formed, including: Space-based observation data is composed of optical remote sensing satellite imagery, synthetic aperture radar satellite imagery, and spaceborne lidar data. Airborne observation data are constructed from remote sensing data acquired by UAVs equipped with optical cameras, radar, or lidar. Ground-based observation data are constructed using temperature, precipitation, and radiation data collected from ground meteorological stations, as well as measured snow depth and snow density data. Based on a unified spatiotemporal reference, three types of observation data—space-based observation data, airborne observation data, and ground-based observation data—are registered and fused to form a multi-source collaborative monitoring dataset that covers the target area and has different spatiotemporal resolutions and physical characteristics.

3. The intelligent identification method for ice avalanche hazards based on multi-source remote sensing and ground coordination according to claim 2, characterized in that, Based on a unified spatiotemporal reference, observation data are registered and fused to form a multi-source collaborative monitoring dataset covering the target area and possessing different spatiotemporal resolutions and physical characteristics, including: The space-based observation data and air-based observation data are uniformly resampled to a specified spatial resolution, and the spatial coordinates of the ground-based observation points are precisely correlated with the corresponding pixels; Using Coordinated Universal Time as a unified time reference, the time labels of each observation data are standardized, and for data with different observation frequencies, time interpolation methods are used to generate datasets with the same time series intervals. By employing a statistical regression model, the observations obtained from different sensors are uniformly converted into key physical parameters of snow cover, establishing a physical consistency relationship among multi-source parameters and forming a multi-dimensional collaborative monitoring dataset.

4. The intelligent identification method for ice avalanche hazards based on multi-source remote sensing and ground coordination according to claim 3, characterized in that, An adaptive spatiotemporal fusion algorithm is used to reconstruct data missing areas caused by cloud cover and snow cover in the multi-source collaborative monitoring data, including: Cloud and snow mask identification is performed on the temporal optical remote sensing images of the target area to mark areas with missing data; Based on the spatiotemporal resolution, physical characteristics, and data quality of each data source, dynamic weights are assigned to each type of observation data, where: The areas in optical images without cloud or snow coverage are assigned the highest weight. For microwave SAR imagery, based on its advantage in penetrating clouds and snow, higher weights are assigned to areas lacking optical data. High-precision correction weights are assigned to the areas covered by the observation phase for the measured data of lidar and UAVs. For regions lacking optical data, a fusion model based on spatiotemporal spectral similarity is constructed.

5. The intelligent identification method for ice avalanche hazards based on multi-source remote sensing and ground coordination according to claim 4, characterized in that, The fusion model based on spatiotemporal spectral similarity generates high spatiotemporal resolution continuous snow cover dynamic monitoring data, including: In the spatial dimension, high-resolution UAV data or historical cloudless images from the same period are used to fill in detailed information through spatial interpolation and texture transfer methods. In the time dimension, by leveraging the high temporal continuity of microwave SAR time series data, a time series variation model of snow cover parameters is established, and the snow cover status of missing periods is recovered through deep learning time series prediction methods. By fusing high-precision terrain and snow depth information obtained from lidar, the interpolation and prediction results are corrected using three-dimensional spatial constraints. The snowfall and snowmelt process parameters measured by ground meteorological stations are used as boundary conditions, and the fusion results are optimized through energy balance equations. An iterative optimization algorithm is used to adaptively adjust the fusion parameters within the effective observation area, with the goal of minimizing the error between the reconstructed data and the actual observation data, to obtain dynamic monitoring data of snow cover.

6. The intelligent identification method for ice avalanche hazards based on multi-source remote sensing and ground coordination according to claim 5, characterized in that, The continuous snow cover dynamic monitoring data is processed to automatically extract and output the snow cover start time, snow cover end time, snow water equivalent change trend map, and snow line dynamic change index, including: The dynamic monitoring data of snow cover is processed, and a sliding window smoothing algorithm is used to obtain the time series curve of snow cover range; Set a threshold for the proportion of snow cover area. When the proportion of snow cover area exceeds the threshold T for N consecutive days, the first day of the period will be determined as the start date of snow accumulation. When the snow cover area ratio is lower than the threshold T for M consecutive days, the first day of that period is determined as the end date of snow cover. Based on the time series of the snow cover dynamic monitoring data, calculate the cumulative snow water equivalent curve for each pixel throughout the entire snow cover period; The key inflection points of the snow water equivalent accumulation, stabilization and ablation stages were extracted by piecewise linear fitting method, and a spatiotemporal variation trend map of snow water equivalent at the regional scale was generated.

7. The intelligent identification method for ice avalanche hazards based on multi-source remote sensing and ground coordination according to claim 6, characterized in that, Based on the aforementioned snowline dynamic change indicators, and combined with glacier spatial geometric parameters, glacier surface characteristic parameters, glacier dynamic state parameters, topographically derived factors, meteorological driving factors, meltwater infiltration index, seismic activity parameters, and spatial density of historical ice avalanche events, an ice avalanche hazard identification feature vector is constructed, including: The characteristic parameters of glacier spatial geometry, glacier surface features, glacier dynamic state, topographic factors, meteorological driving factors, meltwater infiltration index, seismic activity parameters, and spatial density of historical glacial avalanches are unified to the same spatiotemporal scale. Based on the aforementioned dynamic change index of the snow line, the target area is divided into a regular grid, and the point and line parameters are aggregated to the grid scale through spatial interpolation. Using the snow season as the time unit, parameters from different observation frequencies are aggregated to the same time resolution through time series analysis. The normalized multi-source feature parameters are arranged in a fixed order to construct a multi-dimensional feature vector corresponding to each spatial grid.

8. The intelligent identification method for ice avalanche hazards based on multi-source remote sensing and ground coordination according to claim 7, characterized in that, The feature vector for identifying ice avalanche hazards is standardized, including: The Z-score normalization method is used to calculate the mean and standard deviation of each feature dimension based on the training sample set, and the input feature vector for ice avalanche hazard identification is normalized as follows: ； in, Represents the original feature values. This represents the mean of the features in the training set. Indicates standard deviation, The value is represented by the standard. For feature dimensions with missing values, the missing values ​​are filled in using the median after normalizing the features in the training set.

9. The intelligent identification method for ice avalanche hazards based on multi-source remote sensing and ground coordination according to claim 8, characterized in that, The standardized feature vector for identifying ice avalanche hazards is input into a pre-trained ice avalanche hazard classification model, including: The standardized ice avalanche hazard identification feature vector is input into a pre-trained ice avalanche hazard classification model, which is a support vector machine (SVM) model with radial basis function (RBF) as the kernel function. Its decision function expression is as follows: ； in, This represents the standardized feature vector of the input. Represents support vectors, This represents the category label corresponding to the support vector. Represents the Lagrange multipliers. Indicates the bias term. This represents the RBF kernel function.

10. The intelligent identification method for ice avalanche hazards based on multi-source remote sensing and ground coordination according to claim 9, characterized in that, Based on the decision function output value, the intelligent identification result of the ice avalanche hazard area is obtained, including: like If so, the corresponding glacier unit or spatial grid is determined to be a potential ice avalanche zone; like If so, it is determined to be a stable region; After processing all input grids, a spatial distribution map of ice avalanche hazards at the regional scale is generated, along with a confidence score and a list of key contributing features for each hazard area.

Images