Avalanche early warning model construction method and system based on deep learning

Abstract

The application provides an avalanche early warning model construction method and system based on deep learning. First, multi-source environmental monitoring data including meteorological time series, terrain space and snow layer physical data of a target area are acquired. The time dimension of the meteorological time series data is aligned to generate a feature sequence, the terrain space data is gridded to generate a feature set, and the snow layer physical data is extracted to generate a state vector. Then, a deep learning network model containing time series attention, spatial convolution and cross-modal interaction units is inputted to generate a fusion feature vector. A training set is constructed based on historical avalanche event labeling data. The fusion feature vector is dynamically weighted and optimized to generate an avalanche risk prediction model. Finally, real-time monitoring data is received, the avalanche risk prediction model outputs a risk level and a warning trigger threshold, and a multi-level warning signal is generated when the real-time risk value exceeds the threshold, thereby realizing accurate avalanche early warning.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and more specifically, to a method and system for constructing an avalanche early warning model based on deep learning. Background Technology

[0002] In the field of avalanche early warning technology, traditional avalanche early warning methods often rely solely on meteorological or topographical data analysis, neglecting the crucial factor of the physical state of the snow layer. This results in inaccurate and incomplete warning results. Furthermore, due to the lack of effective integration and in-depth analysis of multi-source data, traditional methods struggle to capture the complex nonlinear relationships within avalanche risk, thus limiting the improvement of early warning accuracy.

[0003] Furthermore, the relevant technologies lack cross-modal data fusion mechanisms. For example, meteorological time-series data, topographic spatial data, and snow cover physical data are often processed separately, failing to fully utilize their complementarity and correlation. This not only wastes data resources but may also lead to the loss of key information, thereby affecting the accuracy of early warnings. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a method for constructing an avalanche early warning model based on deep learning, the method comprising:

[0005] Acquire multi-source environmental monitoring data for the target area, including meteorological time-series data, topographic spatial data, and snow cover physical data;

[0006] The meteorological time series data is aligned to the time dimension to generate a meteorological feature sequence; the topographic spatial data is spatially gridded to generate a topographic spatial feature set; and the physical parameters of the snow layer physical data are extracted to generate a snow state vector.

[0007] The meteorological feature sequence, topographic spatial feature set, and snow cover state vector are input into a deep learning network model to generate a fused feature vector. The deep learning network model includes a temporal attention unit, a spatial convolution unit, and a cross-modal interaction unit.

[0008] A training set is constructed based on historical avalanche event annotation data. The training set is then used to dynamically optimize the weights of the fused feature vector to generate an avalanche risk prediction model.

[0009] The system receives real-time environmental monitoring data and outputs the risk level and early warning trigger threshold of the target area through the avalanche risk prediction model. When the real-time risk value exceeds the early warning trigger threshold, a multi-level early warning signal is generated.

[0010] In another aspect, embodiments of the present invention also provide a deep learning-based avalanche early warning model construction system, including a processor and a machine-readable storage medium connected to the processor. The machine-readable storage medium is used to store programs, instructions, or code, and the processor is used to execute the programs, instructions, or code in the machine-readable storage medium to implement the above-described method.

[0011] Based on the above, this application's embodiments acquire meteorological time-series data, topographic spatial data, and snow layer physical data of the target area. On this basis, through temporal dimension alignment processing of the meteorological time-series data, spatial gridding processing of the topographic spatial data, and extraction of physical parameters from the snow layer physical data, key features from each data source are effectively mined, forming meteorological feature sequences, topographic spatial feature sets, and snow state vectors. Then, a deep learning network model including temporal attention units, spatial convolution units, and cross-modal interaction units is introduced. This model can automatically learn and capture the complex nonlinear relationships between meteorological, topographic, and snow state data. Through cross-modal interaction, deep fusion of multi-source features is achieved, generating representative fused feature vectors. This not only improves the richness and accuracy of feature representation but also enhances the comprehensive perception of avalanche risk factors. During the model training phase, the fused feature vectors are dynamically weighted and optimized based on a training set constructed from historical avalanche event annotation data. This allows the avalanche risk prediction model to adaptively adjust the contribution of each feature to the prediction results, thereby improving the model's generalization ability and prediction accuracy. Ultimately, by receiving real-time environmental monitoring data and utilizing a trained avalanche risk prediction model to output the risk level and warning trigger threshold for the target area, real-time monitoring and dynamic early warning of avalanche risk were achieved. When the real-time risk value exceeds the warning trigger threshold, multi-level warning signals can be rapidly generated, providing timely and effective decision support for relevant departments and personnel, and significantly reducing the potential losses and impacts of avalanche disasters. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the execution flow of the avalanche early warning model construction method based on deep learning provided in the embodiments of the present invention.

[0013] Figure 2 This is a schematic diagram of exemplary hardware and software components of the avalanche early warning model construction system based on deep learning provided in an embodiment of the present invention. Detailed Implementation

[0014] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating a method for constructing an avalanche early warning model based on deep learning, according to an embodiment of the present invention. The following is a detailed description of this method.

[0015] Step S110: Obtain multi-source environmental monitoring data of the target area, including meteorological time series data, topographic spatial data and snow layer physical data.

[0016] For example, in a mountainous area prone to avalanches, a corresponding environmental monitoring system can be pre-configured. Specifically, multiple meteorological monitoring stations can be set up in the area, recording meteorological data such as air pressure, temperature, snowfall, and wind speed every hour for an extended period, forming time-series meteorological data. Another example is topographic spatial data, which can be obtained through a combination of aerial mapping and ground surveying. Aerial mapping acquires large-area topographic contour data, while ground surveying teams use specialized instruments to precisely measure detailed topographic data at key locations, encompassing the topographic and geomorphological information of the entire target area. For snow cover physical data, sensors can be buried at different locations to detect snow accumulation. These sensors can acquire information such as snow cover profile data. For example, sensors can be set up at different heights and aspects of the hillside to comprehensively understand the snow cover situation.

[0017] Step S120: Perform time dimension alignment processing on the meteorological time series data to generate a meteorological feature sequence; perform spatial gridding processing on the terrain spatial data to generate a terrain spatial feature set; and extract physical parameters from the snow layer physical data to generate a snow state vector.

[0018] In this embodiment, occasionally, in the previously acquired meteorological time-series data, a meteorological monitoring station may fail to record data at a certain time point due to equipment failure or communication problems, resulting in missing time nodes. For example, at 10:00 AM on a certain day, a meteorological monitoring station located on a hillside may not record the air pressure data for that time. Therefore, linear interpolation compensation can be performed on the missing time node based on the rate of change of air pressure gradient and the temperature change trend of adjacent time windows. Assuming that the air pressure at adjacent times of 9:00 AM and 11:00 AM are 1000 hPa and 1005 hPa respectively, and the temperature rises from -5°C at 9:00 AM to -3°C at 11:00 AM, a reasonable air pressure value can be calculated based on the above data to compensate for the missing air pressure data at 10:00 AM, thereby generating a continuous meteorological data stream. Then, key meteorological indicators can be extracted from this continuous meteorological data stream, such as cumulative snowfall (assuming the cumulative snowfall in the target area reached 50 cm in the past 24 hours), peak wind speed duration (e.g., wind speeds exceeding 15 m / s for 3 hours), and diurnal temperature range fluctuation (e.g., a fluctuation of 8°C from a daytime high of -2°C to a nighttime low of -10°C). Next, these key meteorological indicators within each preset time unit (here set to 12 hours) are aggregated locally using a sliding window mechanism to generate time-correlated meteorological feature segments. For example, the cumulative snowfall, peak wind speed duration, and diurnal temperature range fluctuation within the first 12 hours are aggregated together to form a meteorological feature segment. Finally, a bidirectional long short-term memory network is used to model the temporal dependencies of these meteorological feature segments, such as capturing the cumulative effect of snowfall in the previous 12 hours on temperature changes in the next 12 hours and the association with sudden meteorological abrupt changes, ultimately generating a meteorological feature sequence containing multidimensional temporal context.

[0019] Furthermore, the geographic coordinate system of the target area can be converted to a raster coordinate system, and spatial grid cells can be generated at a preset resolution of 10m × 10m. Within each spatial grid cell, the slope change rate is extracted using topographic surveying instruments; for example, the slope change rate from the foot of the mountain to the summit within a certain grid cell is 0.3. Aspect dispersion is also extracted, assuming that the aspect dispersion within the cell is measured to be set from northeast to southeast. Surface roughness index is measured, and the surface roughness index within the grid cell due to factors such as rocks and vegetation is measured to be 0.5. Simultaneously, based on satellite remote sensing image data, the surface cover type is identified, revealing that the vegetation density distribution within a certain grid cell is 30%, and the exposed rock area accounts for 20%, and these data are mapped to the corresponding grid cells. Next, a three-dimensional terrain topology map is constructed, and the terrain energy gradient is calculated based on the elevation differences between adjacent grid cells; for example, if the elevation difference between adjacent grid cells is 5 meters, the potential movement path intensity distribution is calculated based on an avalanche kinetics model. Finally, multi-scale feature extraction of these terrain energy gradients and potential motion path intensity distributions is performed using a dilated convolutional neural network to generate a terrain spatial feature set containing spatial correlation features.

[0020] Furthermore, snow layer profile data can be acquired using a ground-based sensor array. For example, at a certain monitoring point, the snow layer thickness distribution can be 10 cm, 20 cm, and 30 cm at different depths from the surface; the crystal morphology can be classified as hexagonal; and the liquid water content can be 5%. The vertical density gradient can be calculated based on the snow layer thickness distribution, assuming a surface snow layer density of 0.1 g / cm³ and a deep snow layer density of 0.3 g / cm³. The interlayer shear strength coefficient can be determined by combining the crystal morphology classification; for example, based on the characteristics of hexagonal crystals, the interlayer shear strength coefficient is determined to be 0.2. The surface temperature field of the snow layer is retrieved based on thermal infrared imaging data, yielding a surface temperature of -5℃. The rate of change of the temperature gradient within the snow layer is calculated using the heat conduction equation, assuming a temperature change rate of -1℃ / cm from the surface to a depth of 10 cm. By integrating the interlayer shear strength coefficient, liquid water content, and temperature gradient rate of change, a snow mass stability assessment index can be constructed. Then, a nonlinear mapping of the snow stability assessment index is performed using a physical constraint neural network to generate a snow state vector that reflects the internal structural state of the snow.

[0021] Step S130: Input the meteorological feature sequence, terrain spatial feature set and snow cover state vector into the deep learning network model to generate a fused feature vector. The deep learning network model includes a temporal attention unit, a spatial convolution unit and a cross-modal interaction unit.

[0022] In this embodiment, the previously generated meteorological feature sequence can be input into the temporal attention unit of the deep learning network model. For example, different time points in the meteorological feature sequence contain different meteorological indicators, such as snowfall and wind speed at different times of a day. Therefore, by calculating the attention weight distribution in the time dimension, it is assumed that during periods with a higher probability of avalanches, such as periods with heavy snowfall and high wind speeds, the corresponding meteorological indicators are assigned higher attention weights. This dynamically weights the meteorological indicators at different time points in the meteorological feature sequence, generating weighted meteorological temporal features.

[0023] Then, the terrain spatial feature set can be input into the spatial convolutional unit of the deep learning network model, and the spatial context features of the terrain spatial feature set can be extracted using multi-level dilated convolutional kernels. For example, in a region with complex terrain, multi-level dilated convolutional kernels can better capture the terrain correlation between adjacent grid cells, such as capturing the correlation between a grid cell on a hillside and its adjacent cells in terms of slope, aspect, etc., and generating multi-scale terrain spatial features.

[0024] Next, the snow cover state vector and the weighted meteorological time series features are aligned and spliced ​​together in terms of feature dimensions. For example, the snow body stability assessment index in the snow cover state vector is aligned and spliced ​​together with the snowfall amount, wind speed and other indicators in the weighted meteorological time series features in terms of dimensions to generate the first fusion intermediate feature.

[0025] Then, the first fused intermediate feature and the multi-scale terrain spatial features are input into the cross-modal interaction unit. The cross-modal feature correlation degree between the first fused intermediate feature and the multi-scale terrain spatial features is calculated through a cross-attention mechanism. For example, the first fused intermediate feature is used to generate a query vector sequence through linear projection, and the multi-scale terrain spatial features are used to generate a key vector sequence and a value vector sequence through linear projection. The attention score matrix between the query vector sequence and the key vector sequence is calculated. It is assumed that after weighted aggregation of the value vector sequence based on this matrix, the resulting spatially enhanced feature sequence can better reflect the relationship between snow cover state and terrain spatial features. Then, the spatially enhanced feature sequence and the first fused intermediate feature are residually connected to generate the second fused intermediate feature. The second fused intermediate feature is then input into a multilayer perceptron for nonlinear mapping to generate high-dimensional interactive features. Finally, the high-dimensional interactive features are gatedly fused with the snow cover state vector. The weight ratio of each dimension in the high-dimensional interactive features is dynamically adjusted according to the current value of the snow cover state vector. For example, when the snow body stability evaluation index in the snow cover state vector is low, higher weights are assigned to the terrain-related high-dimensional interactive feature dimensions to generate cross-modal interactive features.

[0026] Finally, channel attention weighted filtering is applied to the cross-modal interaction features to retain feature channels strongly correlated with avalanche triggering conditions, such as those directly related to snowfall, slope, and snow mass stability, generating optimized cross-modal interaction features. These optimized features are then input into a fully connected layer for nonlinear transformation and feature dimensionality reduction to generate a fused feature vector.

[0027] Step S140: Construct a training set based on historical avalanche event annotation data, and use the training set to dynamically optimize the weights of the fused feature vector to generate an avalanche risk prediction model.

[0028] In this embodiment, historical avalanche event data from the past few years can be collected for the target area. This data records in detail the time, location, and various environmental factors of each avalanche. For example, one avalanche occurred at a specific location on a hillside, with a cumulative snowfall of 80 centimeters, a wind speed of 20 meters per second, and a southeast orientation, among other relevant data. The data is also marked with whether an avalanche occurred (binary label) and the actual timestamp of the avalanche triggering.

[0029] Then, the fused feature vector can be input into the first fully connected layer and the second fully connected layer of the deep learning network model, respectively. The first fully connected layer outputs the predicted avalanche occurrence probability. Suppose that for a certain input fused feature vector, the predicted avalanche occurrence probability output by the first fully connected layer is 0.6; the second fully connected layer outputs the predicted avalanche trigger time, for example, the predicted avalanche trigger time is 3 pm.

[0030] Next, cross-entropy loss is calculated by comparing the predicted avalanche occurrence probability with the binary classification labels in the historical avalanche event annotation data. For example, if an avalanche actually occurs but the predicted probability is low, a large classification loss value will be generated. Mean squared error is then calculated by comparing the predicted avalanche trigger time with the actual trigger timestamps in the historical avalanche event annotation data. If the predicted avalanche trigger time differs significantly from the actual trigger time, a large time regression loss value will be generated.

[0031] Then, the classification loss value and the temporal regression loss value are input into the dynamic weight allocator, which generates classification weight coefficients and temporal weight coefficients based on the ratio of classification loss value to temporal regression loss value in the current training batch. For example, when the classification loss value is large, the classification weight coefficient may be set to 0.7 and the temporal weight coefficient to 0.3.

[0032] Next, the classification loss value is multiplied by the classification weight coefficient, and the time regression loss value is multiplied by the time weight coefficient to generate a weighted classification loss component and a weighted time loss component. Then, the weighted classification loss component and the weighted time loss component are added together to generate the joint optimization loss value.

[0033] Finally, backpropagation gradient updates can be performed on the parameters of the first fully connected layer, the second fully connected layer, and the deep learning network model based on the joint optimization loss value, iteratively adjusting the mapping relationship between the fused feature vector and the avalanche risk prediction result. For example, if the joint optimization loss value is large, the model parameters will be adjusted significantly. When the decrease in the joint optimization loss value is less than the convergence threshold (e.g., 0.01) in a preset number of training iterations (e.g., 10), the parameters of the deep learning network model are frozen, generating the avalanche risk prediction model.

[0034] Step S150: Receive current environmental monitoring data in real time, output the risk level and warning trigger threshold of the target area through the avalanche risk prediction model, and generate a multi-level warning signal when the real-time risk value exceeds the warning trigger threshold.

[0035] In this embodiment, during the continuous operation of the monitoring system in the target area, real-time meteorological time-series data, topographic spatial data, and snow cover physical data can be collected. For the current meteorological time-series data, the same meteorological monitoring station records data such as air pressure, temperature, snowfall, and wind speed in real time. The current topographic spatial data is updated through periodic topographic measurements, and the snow cover physical data is acquired in real time by buried sensors.

[0036] The system performs time-dimensional alignment processing on current meteorological time-series data to generate real-time meteorological feature sequences. For example, at a certain moment, the snowfall in the real-time meteorological feature sequence may increase sharply in a short period of time. It also performs spatial gridding processing on current topographic spatial data to generate real-time topographic spatial feature sets. For example, new snowfall may cause slight changes in the slope rate within a certain grid cell. Finally, it extracts physical parameters from current snow layer physical data to generate real-time snow state vectors. For example, the liquid water content in snow may increase due to temperature changes.

[0037] The avalanche risk prediction model inputs real-time meteorological feature sequences, real-time terrain spatial feature sets, and real-time snow cover state vectors to output a real-time risk value for the target area and a dynamically updated warning trigger threshold. Assume the output real-time risk value is 0.7 and the warning trigger threshold is 0.6.

[0038] Risk level intervals are divided based on the difference between the real-time risk value and the warning trigger threshold, and risk level labels for the corresponding risk level intervals are generated. Here, since the real-time risk value is greater than the warning trigger threshold, it may be classified as a high-risk level label.

[0039] When the real-time risk value exceeds the warning trigger threshold, a preset warning level mapping table is activated based on the risk level label, generating a multi-level warning signal containing different alarm frequencies and emergency response instructions. For example, at a high-risk level, the alarm frequency is set to once every 10 minutes, and the emergency response instructions may include measures such as notifying nearby residents to evacuate to safe areas as soon as possible and restricting traffic in mountainous areas.

[0040] Based on the above steps, this embodiment of the application acquires meteorological time-series data, topographic spatial data, and snow layer physical data of the target area. On this basis, through temporal dimension alignment processing of the meteorological time-series data, spatial gridding processing of the topographic spatial data, and physical parameter extraction of the snow layer physical data, key features from each data source are effectively mined, forming meteorological feature sequences, topographic spatial feature sets, and snow state vectors. Then, a deep learning network model containing temporal attention units, spatial convolution units, and cross-modal interaction units is introduced. This model can automatically learn and capture the complex nonlinear relationships between meteorological, topographic, and snow state data. Through cross-modal interaction, deep fusion of multi-source features is achieved, generating representative fused feature vectors. This not only improves the richness and accuracy of feature representation but also enhances the comprehensive perception of avalanche risk factors. During the model training phase, the fused feature vectors are dynamically weighted and optimized based on a training set constructed from historical avalanche event labeled data. This allows the avalanche risk prediction model to adaptively adjust the contribution of each feature to the prediction results, thereby improving the model's generalization ability and prediction accuracy. Ultimately, by receiving real-time environmental monitoring data and utilizing a trained avalanche risk prediction model to output the risk level and warning trigger threshold for the target area, real-time monitoring and dynamic early warning of avalanche risk were achieved. When the real-time risk value exceeds the warning trigger threshold, multi-level warning signals can be rapidly generated, providing timely and effective decision support for relevant departments and personnel, and significantly reducing the potential losses and impacts of avalanche disasters.

[0041] In one possible implementation, step S120 includes:

[0042] Step S121: Identify missing time nodes in the meteorological time series data, and perform linear interpolation compensation on the missing time nodes based on the pressure gradient change rate and temperature change trend of adjacent time windows to generate a continuous meteorological data stream.

[0043] In this embodiment, meteorological time-series data is recorded by multiple meteorological monitoring stations at fixed time intervals. However, during long-term meteorological monitoring, data gaps may occasionally occur. For example, a meteorological monitoring station located on a mountainside may have missing time nodes in its meteorological time-series data between 9:00 AM and 11:00 AM on a certain day. In this case, linear interpolation compensation is performed by analyzing the rate of change of air pressure gradient and the temperature change trend of adjacent time windows to generate a continuous meteorological data stream. Assume that the air pressure is 1002 hPa and the temperature is -3℃ at 8:00 AM, and the air pressure is 1006 hPa and the temperature is -1℃ at 12:00 PM. First, calculate the rate of change of air pressure gradient. From 8:00 AM to 12:00 PM, a total of 4 hours, the air pressure changes by 1006 - 1002 = 4 hPa, so the hourly rate of change of air pressure gradient is 4 ÷ 4 = 1 hPa / hour. Regarding the temperature change trend, the temperature rises by -1 - (-3) = 2℃ in 4 hours, so the hourly rate of change of temperature is 2 ÷ 4 = 0.5℃ / hour. For the missing time point of 10:00, which is two hours after 8:00, the estimated air pressure at 10:00 is 1002 + 1 × 2 = 1004 hPa based on the rate of change of the air pressure gradient; and the estimated temperature at 10:00 is -3 + 0.5 × 2 = -2℃ based on the rate of change of temperature. By interpolating and compensating for the missing time point in this way, a continuous meteorological data stream is obtained.

[0044] Step S122: Extract key meteorological indicators from the continuous meteorological data stream. The key meteorological indicators include cumulative snowfall, peak wind speed duration, and diurnal temperature range fluctuation.

[0045] For example, regarding cumulative snowfall, the cumulative snowfall was 80 cm after summarizing the snowfall data from all monitoring points in the target area over the past week. Regarding the duration of peak wind speeds, analysis of wind speed data from each monitoring station revealed that wind speeds exceeding 12 m / s lasted for 5 hours. The diurnal temperature range was calculated using a 24-hour cycle, with a daytime high of -1℃ and a nighttime low of -9℃, resulting in a diurnal temperature range of -1 - (-9) = 8℃.

[0046] Step S123: The key meteorological indicators within each preset time unit are aggregated locally using a sliding window mechanism to generate meteorological feature segments with time correlation.

[0047] In this embodiment, a preset time unit can be set to 12 hours. In the first 12 hours, the cumulative snowfall is 30 centimeters, the peak wind speed lasts for 2 hours, and the diurnal temperature range fluctuates by 6°C. By combining the above indicators, a meteorological feature segment can be formed.

[0048] Step S124: The meteorological feature fragments are modeled for temporal dependence using a bidirectional long short-term memory network to capture the correlation between forward cumulative effect and backward mutation, and to generate a meteorological feature sequence containing multidimensional temporal context.

[0049] For example, the cumulative snowfall in the previous 12 hours will affect the temperature in the next 12 hours; this is the forward cumulative effect. If a cold front suddenly arrives, the wind speed will increase sharply; this is the backward abrupt correlation. These relationships can be captured using bidirectional long short-term memory networks, thereby generating meteorological feature sequences that contain multidimensional temporal context.

[0050] Step S125: Convert the geographic coordinate system of the target area into a raster coordinate system and generate spatial grid cells according to a preset resolution.

[0051] For example, for a target area, its geographic coordinate system can be converted into a raster coordinate system, and spatial grid cells can be generated at a preset resolution of 20 meters × 20 meters.

[0052] Step S126: Extract the slope change rate, aspect dispersion, and surface roughness index as terrain attribute data within each spatial grid cell.

[0053] Taking a spatial grid cell located in the middle of a hillside as an example, the slope change rate was measured using a topographic surveying instrument. From one end of the cell to the other, the height difference is 4 meters, and the horizontal distance is 20 meters; the slope change rate is 4 ÷ 20 = 0.2. Regarding aspect dispersion, by measuring the change in aspect angle, it was found that the aspect angle changes from 30 degrees east of north to 40 degrees east of north from one side of the cell, indicating a predetermined dispersion. As for the surface roughness index, due to the presence of rocks and vegetation within the cell, the measured and calculated surface roughness index is 0.4.

[0054] Step S127: Identify land cover type based on satellite remote sensing image data, and map vegetation density distribution and the proportion of exposed rock area to the corresponding grid cells.

[0055] For example, within a certain grid cell, analysis of remote sensing imagery reveals that vegetation cover accounts for 40% of the total area of ​​the cell, so the vegetation density distribution is 40%, and the area of ​​exposed rocks accounts for 10%. These data are then mapped to the corresponding grid cells.

[0056] Step S128: Construct a three-dimensional terrain topology map, calculate the terrain energy gradient based on the elevation difference of adjacent grid cells, and generate the potential motion path intensity distribution by combining the avalanche motion mechanics model.

[0057] For example, assuming a grid cell has an elevation of 100 meters, an adjacent grid cell has an elevation of 105 meters, and a horizontal distance of 20 meters, according to the avalanche kinetic model, the elevation difference is first calculated as 105 - 100 = 5 meters, and the terrain energy gradient is 5 ÷ 20 = 0.25. Then, combining this with the avalanche kinetic model, the potential motion path intensity distribution is generated. Based on factors such as terrain slope, aspect, and surface cover type, the potential motion path intensity distribution in this grid cell and its adjacent grid cells is calculated.

[0058] Step S129: Multi-scale feature extraction is performed on the terrain energy gradient and potential motion path intensity distribution using a dilated convolutional neural network to generate a terrain spatial feature set containing spatial correlation features.

[0059] For example, dilated convolutional neural networks can capture terrain features at different scales. On a large scale, they can capture the influence of the overall orientation of mountains on avalanche paths, while on a small scale, they can capture the influence of small changes in local terrain within grid cells on avalanches, thereby generating a set of terrain spatial features containing spatial correlation features.

[0060] Step S1210: Obtain snow layer profile data through a ground sensor array. The snow layer profile data includes snow layer thickness distribution, crystal morphology classification, and liquid water content.

[0061] For example, at a monitoring point near the mountaintop, snow layer profile data showed the following snow thickness distribution: 15 cm for the surface 0-10 cm depth, 25 cm for the 10-20 cm depth, and 30 cm for the 20-30 cm depth. Regarding crystal morphology, microscopic observation and analysis determined the crystal morphology to be dendritic. The liquid water content, measured with specialized instruments, was 8%.

[0062] Step S1211: Calculate the vertical density gradient based on the snow layer thickness distribution, and determine the interlayer shear strength coefficient in combination with the crystal morphology classification.

[0063] For example, assuming the surface snow density is 0.15 g / cm³, the snow density at a depth of 30 cm is 0.3 g / cm³, and the snow thickness from the surface to the 30 cm depth is 60 cm, the vertical density gradient is calculated as (0.3 - 0.15) ÷ 30 = 0.005 g / cm³. The interlaminar shear strength coefficient is determined by combining crystal morphology classification. For dendritic crystals, for example, the interlaminar shear strength coefficient can be determined to be 0.25.

[0064] Step S1212: The surface temperature field of the snow is inverted based on thermal infrared imaging data, and the rate of change of temperature gradient inside the snow layer is calculated using the heat conduction equation.

[0065] For example, the surface temperature of the snow is obtained as -4℃ using a thermal infrared imager. Then, the rate of change of the temperature gradient inside the snow layer is calculated using the heat conduction equation. Assuming the temperature is -6℃ at a depth of 10 cm, according to the heat conduction equation, the rate of change of the temperature gradient is (-6 - (-4)) ÷ 10 = -0.2℃ / cm.

[0066] Step S1213: Combine the interlayer shear strength coefficient, liquid water content and temperature gradient change rate to construct a snow body stability evaluation index.

[0067] For example, the calculation method can be to combine the interlayer shear strength coefficient of 0.25, the liquid water content of 8%, and the temperature gradient change rate of -0.2℃ / cm according to the set weights. For example, the weight of the interlayer shear strength coefficient is 0.4, the weight of the liquid water content is 0.3, and the weight of the temperature gradient change rate is 0.3. The snow stability assessment index is calculated as 0.25×0.4+0.08×0.3+(-0.2)×0.3=0.014.

[0068] Step S1214: The snow stability evaluation index is nonlinearly mapped through a physical constraint neural network to generate a snow state vector that reflects the internal structure of the snow.

[0069] For example, the physical constraint neural network establishes a mapping relationship based on a large amount of existing snow accumulation experimental data. The snow stability evaluation index of 0.014 is input into the physical constraint neural network. After the neuron operation and nonlinear function processing inside the physical constraint neural network, a snow accumulation state vector reflecting the internal structure of the snow is generated. This snow accumulation state vector contains various feature information of the internal structure of the snow, such as the stability of the snow and the internal stress distribution.

[0070] In one possible implementation, step S130 includes:

[0071] Step S131: Input the meteorological feature sequence into the temporal attention unit of the deep learning network model, and dynamically weight the meteorological indicators at different time nodes in the meteorological feature sequence by calculating the attention weight distribution in the time dimension to generate weighted meteorological temporal features.

[0072] In this embodiment, the meteorological feature sequence includes meteorological indicators such as snowfall, wind speed, and temperature at different time points. When calculating the attention weight distribution in the time dimension, the weights can be allocated according to the degree of influence of different meteorological indicators on the probability of avalanche occurrence. For example, during periods when avalanches are more likely to occur, such as periods with heavy snowfall and low temperatures, the relevant meteorological indicators will be assigned higher attention weights. Taking a certain meteorological feature sequence as an example, at one time point, the snowfall is 15 cm / h, the wind speed is 10 m / s, and the temperature is -5℃. After analysis by the temporal attention unit, since snowfall and low temperature have a significant impact on the occurrence of avalanches, the weight of snowfall is set to 0.4, the weight of wind speed is 0.3, and the weight of temperature is 0.3. Then, after dynamically weighting the meteorological indicators at this time point, the weighted snowfall is 15 × 0.4 = 6, the weighted wind speed is 10 × 0.3 = 3, and the weighted temperature is -5 × 0.3 = -1.5, thus generating a weighted meteorological temporal feature.

[0073] Step S132: Input the terrain spatial feature set into the spatial convolutional unit of the deep learning network model, use multi-level dilated convolutional kernels to extract spatial context features from the terrain spatial feature set, capture the terrain correlation between adjacent grid units, and generate multi-scale terrain spatial features.

[0074] In this embodiment, within the topographic spatial feature set of the target area, each spatial grid cell contains information such as slope change rate, aspect dispersion, surface roughness index, and surface cover type. For example, for adjacent grid cells in a certain area, one grid cell has a slope change rate of 0.2, a constant aspect dispersion, a surface roughness index of 0.4, and a vegetation density distribution of 40%; the corresponding values ​​for adjacent grid cells are 0.3, different dispersion values, 0.5, and 30%, respectively. Multi-level dilated convolutional kernels can capture the topographic correlations between these adjacent grid cells at different scales. At a large scale, the focus may be on the overall topographic trend of the mountain range, while at a small scale, the focus may be on the impact of local topographic changes within a grid cell on avalanche paths. In this way, multi-scale topographic spatial features are generated.

[0075] Step S133: Align and concatenate the snow cover state vector with the weighted meteorological time series features to generate the first fused intermediate feature.

[0076] In this embodiment, the snow state vector includes comprehensive snow body stability assessment indicators such as interlayer shear strength coefficient, liquid water content, and temperature gradient change rate, calculated based on snow layer physical data. The weighted meteorological time series features include meteorological indicators weighted at different time points. The two are then aligned and concatenated according to their feature dimensions. For example, the snow body stability assessment indicators in the snow state vector are combined with indicators such as snowfall, wind speed, and temperature in the weighted meteorological time series features in a predetermined order and manner to generate the first fused intermediate feature.

[0077] Step S134: Input the first fused intermediate feature and the multi-scale terrain spatial feature into the cross-modal interaction unit, calculate the cross-modal feature correlation degree between the first fused intermediate feature and the multi-scale terrain spatial feature through a cross-attention mechanism, and generate cross-modal interaction features.

[0078] In one possible implementation, step S134 includes:

[0079] Step S1341: Generate a query vector sequence from the first fused intermediate features through linear projection, and generate a key vector sequence and a value vector sequence from the multi-scale terrain spatial features through linear projection.

[0080] Step S1342: Calculate the attention score matrix between the query vector sequence and the key vector sequence, and perform weighted aggregation on the value vector sequence to generate a spatially enhanced feature sequence.

[0081] Step S1343: The spatial enhancement feature sequence is residually connected with the first fusion intermediate feature to generate the second fusion intermediate feature.

[0082] Step S1344: Input the second fused intermediate features into a multilayer perceptron for nonlinear mapping to generate high-dimensional interactive features.

[0083] Step S1345: Gated fusion of the high-dimensional interaction features and the snow state vector is performed, and the weight ratio of each dimension in the high-dimensional interaction features is dynamically adjusted according to the current value of the snow state vector to generate the cross-modal interaction features.

[0084] In this embodiment, when calculating the attention score matrix between the query vector sequence and the key vector sequence, we take a feature component in the first fusion intermediate feature and a related component in the multi-scale terrain spatial feature as an example. Assume that the snowfall component in the first fusion intermediate feature is related to the slope change rate component in the multi-scale terrain spatial feature. If, in a certain area, the snowfall is 15 cm / hour and the slope change rate is 0.2, the attention score between them is calculated according to existing calculation rules. The value vector sequences are then weighted and aggregated. For example, for each value vector component in the multi-scale terrain spatial feature, different weights are assigned based on the calculated attention score, and then a weighted sum is performed to generate a spatially enhanced feature sequence.

[0085] For example, if a feature value in the spatial augmentation feature sequence is 5, and the corresponding feature value in the first fusion intermediate feature is 3, the result after residual connection is 5 + 3 = 8, generating a second fusion intermediate feature. This second fusion intermediate feature is then input into a multilayer perceptron for nonlinear mapping. Neurons within the multilayer perceptron perform operations on the input second fusion intermediate feature according to set weights and activation functions. For example, if a neuron receives an input value of 8, multiplying it by a weight of 0.5 yields 4, which is then processed by an activation function (such as the ReLU function). If 4 is greater than 0, the output is 4. After multiple layers of such operations, a high-dimensional interactive feature is generated.

[0086] Assuming the snow stability evaluation index in the snow cover state vector is 0.014, a low value indicates poor snow cover stability. In this case, higher weights are assigned to dimensions closely related to avalanche triggering, such as terrain slope and snowfall, in the high-dimensional interaction features. For example, one dimension in the high-dimensional interaction features is related to snowfall, with an original weight of 0.3. Based on the value of the snow cover state vector, this weight is adjusted to 0.4, and the weights of other dimensions are adjusted accordingly, thereby generating cross-modal interaction features.

[0087] Step S135: Perform channel attention weighted filtering on the cross-modal interaction features, retain feature channels that are strongly correlated with avalanche triggering conditions, and generate optimized cross-modal interaction features.

[0088] For example, one channel relates the shear strength coefficient between snow layers to topographic slope, while another relates liquid water content to wind speed. Analysis revealed that the channel relating the shear strength coefficient between snow layers to topographic slope has a stronger impact on avalanche triggering, so this channel is assigned a higher weight, while the channel relating liquid water content to wind speed is assigned a lower weight. In this way, feature channels strongly correlated with avalanche triggering conditions are preserved, generating optimized cross-modal interaction features.

[0089] Step S136: Input the optimized cross-modal interaction features into the fully connected layer for nonlinear transformation and feature dimensionality reduction to generate the fused feature vector.

[0090] In this embodiment, the neurons in the fully connected layer are connected to each feature of the optimized cross-modal interaction feature. Each neuron performs a weighted sum of the inputs according to set weights, and then performs a non-linear transformation (such as the sigmoid function). For example, if a neuron receives an input with a weighted sum of 5, the sigmoid function will produce a value between 0 and 1. After such operations by multiple neurons, the non-linear transformation of the features is achieved, and the feature dimensionality is reduced, ultimately generating a fused feature vector. This fused feature vector integrates information from multiple aspects such as meteorology, terrain, and snow cover, and has undergone optimization, filtering, and transformation.

[0091] In one possible implementation, step S140 includes:

[0092] Step S141: Input the fused feature vector into the first fully connected layer and the second fully connected layer of the deep learning network model respectively. The first fully connected layer outputs the predicted avalanche occurrence probability value, and the second fully connected layer outputs the predicted avalanche trigger time value.

[0093] For example, a specific fused feature vector integrates information obtained from processing data on meteorology, topography, snow cover, and other factors. The first fully connected layer, after calculation and processing by its internal neurons, outputs a predicted avalanche probability value; assuming this value is 0.4, it means there is a 40% chance of an avalanche occurring. The second fully connected layer, based on the same fused feature vector, outputs a predicted avalanche trigger time value, for example, predicting that the avalanche will trigger at 5 PM.

[0094] Step S142: Perform cross-entropy loss calculation on the predicted avalanche occurrence probability value and the binary classification labels in the historical avalanche event annotation data to generate a classification loss value.

[0095] Historical avalanche event annotation data records in detail the occurrence of past avalanches. Binary labels indicate whether an avalanche occurred: 1 for occurrence and 0 for non-occurrence. Assume that in a certain historical event, an avalanche actually occurred (binary label: 1), and the currently predicted avalanche probability is 0.4. The cross-entropy loss calculation process is as follows: First, calculate the logarithm of the true label probability. Since the true label is 1, the logarithm is ln(1) = 0. Then, calculate the logarithm of the predicted probability, ln(0.4), which is approximately -0.916. The cross-entropy loss value is -(0×ln(0.4)+(1-1)×ln(1-0.4)), with a result of approximately 0.916. This is the classification loss value.

[0096] Step S143: Calculate the mean square error between the predicted avalanche trigger time and the actual trigger timestamp in the historical avalanche event annotation data to generate a time regression loss value.

[0097] For example, if the predicted avalanche trigger time is 5 PM, while the actual trigger time marked in historical event data is 4 PM, converting these times to minutes, 5 PM is 300 minutes and 4 PM is 240 minutes. The mean squared error is calculated as follows: first, calculate the difference between the predicted and actual values, i.e., 300 - 240 = 60 minutes. Then, square this difference, obtaining 60 × 60 = 3600. The mean squared error is the average of these squared values ​​(since there is only one sample), so the time regression loss value is 3600.

[0098] Step S144: Input the classification loss value and the time regression loss value into the dynamic weight allocator, and generate classification weight coefficient and time weight coefficient according to the ratio of the classification loss value and the time regression loss value in the current training batch.

[0099] In this example, the classification loss is approximately 0.916, and the time regression loss is 3600. Calculating their proportions, the classification loss accounts for 0.916 ÷ (0.916 + 3600) of the total loss, which is very small, approximately 0.000254. The time regression loss accounts for 3600 ÷ (0.916 + 3600), approximately 0.999746. Based on this proportion, classification weight coefficients and time weight coefficients are generated. The classification weight coefficient might be set to 0.000254, and the time weight coefficient to 0.999746.

[0100] Step S145: Multiply the classification loss value by the classification weight coefficient, and multiply the time regression loss value by the time weight coefficient to generate a weighted classification loss component and a weighted time loss component.

[0101] Step S146: Add the weighted classification loss component and the weighted time loss component to generate a joint optimization loss value.

[0102] In this embodiment, the classification loss value of 0.916 is multiplied by the classification weight coefficient of 0.000254 to obtain a weighted classification loss component of approximately 0.000232. The time regression loss value of 3600 is multiplied by the time weight coefficient of 0.999746 to obtain a weighted time loss component of approximately 3599.0856. Then, the weighted classification loss component and the weighted time loss component are added together to generate the joint optimization loss value. That is, 0.000232 + 3599.0856 = 3599.085832.

[0103] Step S147: Based on the joint optimization loss value, backpropagation gradient update is performed on the parameters of the first fully connected layer, the second fully connected layer, and the deep learning network model to iteratively adjust the mapping relationship between the fused feature vector and the avalanche risk prediction result.

[0104] During backpropagation, starting from the output layers (the first and second fully connected layers), the gradient of each parameter is calculated based on the joint optimization loss. For example, for a connection weight of a neuron in the first fully connected layer, the gradient is calculated based on the joint optimization loss and the degree of influence of this weight on the output. If this weight has a significant impact on the predicted avalanche probability, then the adjustment of this weight based on the joint optimization loss will also be substantial during backpropagation. Similarly, similar gradient calculations and updates are performed for other parameters in the second fully connected layer and the deep learning network model. Through multiple iterations, these parameters are continuously adjusted to make the mapping relationship between the fused feature vector and the avalanche risk prediction result more accurate.

[0105] Step S148: When the decrease in the joint optimization loss value is less than the convergence threshold during a preset number of consecutive training iterations, the parameters of the deep learning network model are frozen to generate the avalanche risk prediction model.

[0106] Assuming a preset number of iterations of 10 and a convergence threshold of 0.001, the joint optimization loss value is observed after each iteration during training. For example, the joint optimization loss value is 3599.085832 in the first iteration and becomes 3590.08 after the second iteration, a decrease of (3599.085832-3590.08)÷3599.085832, which is a significant decrease. As the number of iterations increases, assuming that by the 11th iteration, the decrease in the joint optimization loss value in the most recent 10 iterations is less than 0.001, the parameters of the deep learning network model are frozen and no longer updated. This generates an avalanche risk prediction model. This avalanche risk prediction model has undergone dynamic weight optimization based on historical data and can accurately predict the probability and trigger time of avalanches.

[0107] In one possible implementation, step S150 includes:

[0108] Step S151: Real-time collection of current meteorological time-series data, current topographic spatial data, and current snow layer physical data of the target area.

[0109] For current meteorological time-series data, multiple meteorological monitoring stations in the target area continuously record various meteorological data, such as air pressure, temperature, snowfall, and wind speed. For example, at a certain moment, a meteorological monitoring station on a mountainside recorded a current air pressure of 1003 hPa, a temperature of -2°C, a snowfall of 5 cm in the past hour, and a wind speed of 8 m / s. For current topographic spatial data, regular topographic measurements and updates are performed. For example, high-precision measuring instruments are used to remeasure the topographic attribute data within each spatial grid cell. For instance, a newly measured slope change rate of 0.25 was obtained within a certain grid cell, the aspect dispersion has a new value, the surface roughness index has changed to 0.45, and the surface cover type has been updated based on new satellite remote sensing image data, revealing that the vegetation density distribution has changed to 35%, and the proportion of exposed rock area is 15%. For the current physical data of the snow layer, the ground sensor array continuously acquires snow layer profile data. At a certain monitoring point, the snow layer thickness distribution is as follows: the snow layer thickness is 12 cm at a depth of 0-10 cm, the snow layer thickness is 20 cm at a depth of 10-20 cm, and the snow layer thickness is 25 cm at a depth of 20-30 cm. The crystal morphology is classified as hexagonal, and the liquid water content is 6%.

[0110] Step S152: Perform time dimension alignment processing on the current meteorological time series data to generate a real-time meteorological feature sequence; perform spatial gridding processing on the current terrain spatial data to generate a real-time terrain spatial feature set; and extract physical parameters from the current snow layer physical data to generate a real-time snow state vector.

[0111] In the collected meteorological time-series data, there may be discontinuities, such as a brief interruption in data transmission at a certain monitoring station at a specific moment. By identifying these missing time points and performing linear interpolation compensation based on the pressure gradient change rate and temperature change trend of adjacent time windows, a continuous meteorological data stream can be generated. Assume that in the previously mentioned meteorological data, data from 9:00 AM to 10:00 AM is missing from a certain monitoring station. The adjacent data shows that at 8:00 AM, the pressure was 1002 hPa and the temperature was -3℃, and at 11:00 AM, the pressure was 1004 hPa and the temperature was -1℃. Calculating the pressure gradient change rate, the pressure changed by 1004 - 1002 = 2 hPa in 3 hours, with an hourly pressure gradient change rate of 2 ÷ 3 ≈ 0.67 hPa / hour. Regarding the temperature change trend, the temperature rose by -1 - (-3) = 2℃ in 3 hours, with an hourly temperature change rate of 2 ÷ 3 ≈ 0.67℃ / hour. For the missing time point of 10:00, two hours have passed since 8:00. According to the rate of change of air pressure gradient, the air pressure at 10:00 is estimated to be 1002 + 0.67 × 2 = 1003.34 hPa; according to the rate of change of temperature, the temperature at 10:00 is estimated to be -3 + 0.67 × 2 = -1.66℃. After such interpolation compensation, a continuous meteorological data stream is obtained. Then, key meteorological indicators are extracted from this continuous meteorological data stream, such as the current cumulative snowfall of 5 cm (snowfall in the past hour), the duration of peak wind speed (assuming the current wind speed is 8 m / s and has not reached its peak, the duration is within the current hour), and the diurnal temperature range fluctuation (if the current temperature is -2℃ for daytime, and the nighttime temperature is -6℃, the fluctuation range is -2 - (-6) = 4℃). Then, the key meteorological indicators in each preset time unit (e.g., 12 hours) are aggregated locally using a sliding window mechanism to generate meteorological feature segments with time correlation, such as aggregating the cumulative snowfall, peak wind speed duration, and diurnal temperature range fluctuation within 12 hours. Finally, a bidirectional long short-term memory network is used to model the temporal dependence of meteorological feature fragments, capture the forward cumulative effect and the backward abrupt change association, and generate a real-time meteorological feature sequence containing multi-dimensional temporal context.

[0112] In this embodiment, the geographic coordinate system of the target area is converted to a raster coordinate system, and spatial grid cells are generated according to a preset resolution (e.g., 20m × 20m). Within each spatial grid cell, the slope change rate, aspect dispersion, and surface roughness index are recalculated as terrain attribute data. Taking the previously mentioned grid cell as an example, the newly measured slope change rate is 0.25, the aspect dispersion has a new value after remeasurement, and the surface roughness index becomes 0.45. Based on the new satellite remote sensing image data, the land cover type is identified, and the vegetation density distribution of 35% and the proportion of exposed rock area of ​​15% are mapped to the corresponding grid cells. A three-dimensional terrain topology map is constructed, and the terrain energy gradient is calculated according to the elevation difference between adjacent grid cells. For example, if one grid cell has an elevation of 100m, the adjacent grid cell has an elevation of 103m, and the horizontal distance is 20m, the terrain energy gradient is (103-100)÷20=0.15. By combining the avalanche motion mechanics model to generate the potential motion path intensity distribution, and by using a hollow convolutional neural network to extract multi-scale features from the terrain energy gradient and the potential motion path intensity distribution, a real-time terrain spatial feature set containing spatial correlation features is generated.

[0113] Then, physical parameters are extracted from the current snow layer physical data to generate a real-time snow state vector. Snow layer profile data acquired through a ground sensor array shows that at a certain monitoring point, the snow layer thickness distribution is as follows: 12 cm at a depth of 0-10 cm, 20 cm at a depth of 10-20 cm, and 25 cm at a depth of 20-30 cm. The crystal morphology is hexagonal, and the liquid water content is 6%. The vertical density gradient is calculated based on the snow layer thickness distribution. Assuming the surface snow layer density is 0.12 g / cm³, the snow layer density at a depth of 30 cm is 0.25 g / cm³, and the total snow layer thickness from the surface to the 30 cm depth is 57 cm, the calculated vertical density gradient is (0.25-0.12)÷30≈0.0043 g / cm³ / cm. The interlaminar shear strength coefficient is determined based on crystal morphology classification. For hexagonal crystals, the interlaminar shear strength coefficient is determined to be 0.22 based on existing experimental data and experience. The surface temperature field of the snow was retrieved based on thermal infrared imaging data. Assuming a surface temperature of -3℃, the rate of change of the temperature gradient within the snow layer was calculated using the heat conduction equation. For example, at a depth of 10 cm, the temperature is -4℃, and the rate of change of the temperature gradient is (-4 - (-3)) ÷ 10 = -0.1℃ / cm. A snow stability assessment index was constructed by integrating the interlayer shear strength coefficient, liquid water content, and rate of change of the temperature gradient. Assuming a weight of 0.4 for the interlayer shear strength coefficient, 0.3 for the liquid water content, and 0.3 for the rate of change of the temperature gradient, the snow stability assessment index was calculated as 0.22 × 0.4 + 0.06 × 0.3 + (-0.1) × 0.3 = 0.064. A physical constraint neural network was used to perform a nonlinear mapping on the snow stability assessment index, generating a real-time snow state vector reflecting the internal structural state of the snow.

[0114] Step S153: Input the real-time meteorological feature sequence, real-time terrain spatial feature set and real-time snow cover state vector into the avalanche risk prediction model, and output the real-time risk value of the target area and the dynamically updated warning trigger threshold.

[0115] For example, after internal calculations, the avalanche risk prediction model outputs a real-time risk value of 0.55 and a warning trigger threshold of 0.5.

[0116] Step S154: Divide the risk level range according to the difference between the real-time risk value and the early warning trigger threshold, and generate a risk level label corresponding to the risk level range.

[0117] For example, since the real-time risk value of 0.55 is greater than the warning trigger threshold of 0.5, the difference is 0.55 - 0.5 = 0.05. According to the pre-set risk level classification rules, for example, a difference between 0 and 0.1 is a low risk level, between 0.1 and 0.3 is a medium risk level, and above 0.3 is a high risk level. Here, 0.05 is between 0 and 0.1, so a low risk level label is generated.

[0118] Step S155: When the real-time risk value exceeds the warning trigger threshold, activate the preset warning level mapping table according to the risk level label to generate a multi-level warning signal containing different alarm frequencies and emergency response instructions.

[0119] For example, because this is a low-risk area, the alarm frequency in the preset warning level mapping table might be once every 30 minutes. Emergency response instructions might include notifying mountain staff to strengthen patrols of key areas and providing safety reminders to nearby tourists. For a medium-risk area, the alarm frequency might increase to once every 15 minutes, and emergency response instructions would include restricting personnel activities in certain dangerous areas. At a high-risk level, the alarm frequency might become once every 5 minutes, and emergency response instructions would include the organized evacuation of surrounding residents. This completes the entire process from real-time data collection to risk assessment, and then to generating corresponding warning signals based on the risk situation.

[0120] In one possible implementation, after step S150, the method further includes:

[0121] Step S310: Collect real-time surface deformation monitoring data and meteorological radar reflectivity data of the target area to generate a dynamic environment verification dataset.

[0122] For real-time surface deformation monitoring data, high-precision deformation monitoring instruments deployed within the target area continuously record minute changes in the surface. For example, in key locations, such as areas with significant variations in height and slope on hillsides, these instruments can accurately measure the vertical and horizontal displacement of the surface. Suppose that at a certain moment, a monitoring point located on a hillside records a vertical displacement of 0.5 cm and a horizontal displacement of 1 cm. Simultaneously, meteorological radar reflectivity data is continuously acquired. Meteorological radar obtains various information about the atmosphere by emitting electromagnetic waves and receiving reflected waves. Reflectivity data reflects the distribution of different substances in the atmosphere and is also important for determining the state of snow cover. By analyzing meteorological radar reflectivity data, reflectivity values ​​at different altitudes can be obtained, and these values ​​are related to characteristics such as snow moisture content and density. Integrating surface deformation monitoring data and meteorological radar reflectivity data generates a dynamic environment verification dataset.

[0123] Step S320: Input the dynamic environment verification dataset into the pre-trained snow motion recognition model, extract the snow layer displacement rate and surface crack propagation trajectory, and generate snow structure anomaly indicators.

[0124] In this embodiment, the snow movement recognition model is pre-trained using a large amount of snow movement-related data. When a dynamic environment verification dataset is input, the snow movement recognition model can extract information such as snow layer displacement rate and surface crack propagation trajectory, thereby generating snow structure anomaly indicators. Taking snow layer displacement rate as an example, displacement and time information from surface deformation monitoring data are analyzed and calculated. Assume that within a certain period, the vertical displacement of the previously mentioned hillside monitoring point increases from 0 cm to 0.5 cm, and the horizontal displacement increases from 0 cm to 1 cm, with this period lasting 1 hour. In the vertical direction, the displacement rate is 0.5 cm ÷ 1 hour = 0.5 cm / hour; in the horizontal direction, the displacement rate is 1 cm ÷ 1 hour = 1 cm / hour. For the surface crack propagation trajectory, information such as the crack propagation direction and speed is inferred indirectly from image data of the snow surface (which can be obtained from specialized snow surface monitoring equipment) or by analyzing surface deformation data. These snow layer displacement rates and surface crack propagation trajectories are then combined, and a snow structure anomaly indicator is generated according to a pre-defined algorithm.

[0125] Step S330: Perform a spatiotemporal matching degree analysis based on the snow structure anomaly index and the risk level output by the avalanche risk prediction model to generate the model prediction deviation coefficient.

[0126] Suppose the avalanche risk prediction model previously output a medium-risk level, with corresponding expected ranges for key factors such as snow cover thickness and terrain slope. However, anomaly indicators in the snow structure show a rapid snow layer displacement rate and a large extent of surface crack expansion, which differs from the snow condition expected for a medium-risk level. When calculating the spatiotemporal matching degree, temporal synchronicity and spatial correlation must be considered. For example, spatially, if the avalanche risk prediction model considers a region to have high snow stability, but the snow structure anomaly indicator shows a high snow layer displacement rate in that region, this constitutes a spatial mismatch. Temporally, if the avalanche risk prediction model predicts a low avalanche risk in the future based on historical data and current meteorological, topographical, and snow cover data, but the snow structure anomaly indicator shows a rapidly deteriorating snow condition, this constitutes a temporal mismatch. These temporal and spatial mismatches are considered to generate a model prediction bias coefficient. Assume the model's prediction bias coefficient is 0.3 (this value is derived from specific calculations and is only an example).

[0127] Step S340: When the model prediction deviation coefficient exceeds a preset threshold, activate the terrain scanning device to perform three-dimensional point cloud reconstruction of the target area, generate a surface micro-topography change map, and compare the surface micro-topography change map with the feature terrain template before the occurrence of historical avalanche events to identify the coordinates of potential trigger points.

[0128] Assuming a preset threshold of 0.2, the terrain scanning equipment is activated because the previously calculated model prediction bias coefficient of 0.3 is greater than 0.2. These terrain scanning devices can be high-precision measuring equipment such as lidar, which comprehensively scan the target area, acquiring a large amount of terrain point cloud data, and then generating a surface micro-topographic change map through data processing algorithms. This surface micro-topographic change map records detailed changes in the surface terrain, such as changes in hill height and gully depth. Next, the surface micro-topographic change map is compared with a characteristic terrain template from before historical avalanche events. The characteristic terrain template from before historical avalanche events is obtained by analyzing and summarizing the terrain data of areas where avalanches occurred in the past, including terrain features such as specific slope and aspect combinations, and surface roughness. By comparing the surface micro-topographic change map and the characteristic terrain template, the coordinates of potential trigger points are identified. For example, if the terrain changes in a certain area are similar to the characteristic terrain before a historical avalanche event, such as steeper slopes and increased surface roughness, the coordinates of that area are identified as the coordinates of potential trigger points.

[0129] Step S350: Adjust the early warning trigger threshold of the avalanche risk prediction model according to the coordinates of the potential trigger point, generate an adaptive threshold update instruction, correct the feature weight allocation strategy of the current environmental monitoring data based on the adaptive threshold update instruction, and send the updated feature weights back to the cross-modal interaction unit for feature recalibration.

[0130] Assuming the identified potential trigger point coordinates correspond to areas with a high probability of avalanche occurrence, the warning trigger threshold needs to be lowered to improve the accuracy of the warning. For example, if the original warning trigger threshold was 0.5, it can be adjusted to 0.45 based on the risk level corresponding to the potential trigger point coordinates, while simultaneously generating an adaptive threshold update instruction. Based on this adaptive threshold update instruction, the feature weight allocation strategy for the current environmental monitoring data is revised. For instance, for monitoring data near the potential trigger point coordinates, such as liquid water content in snow cover physical data and slope in topographic spatial data—features closely related to avalanche occurrence—their weights are increased. The updated feature weights are then fed back to the cross-modal interaction unit for feature recalibration. The cross-modal interaction unit has already fused multi-source data, including meteorological feature sequences, topographic spatial feature sets, and snow cover state vectors. By recalibrating the feature weights, the cross-modal interaction unit can better fuse this data.

[0131] Step S360: The new fusion feature vector is output by the recalibrated cross-modal interaction unit to drive the avalanche risk prediction model to generate incremental learning parameters. The incremental learning parameters are then fused with the original model parameters by gradient interpolation to generate an environmentally adaptive avalanche risk prediction model.

[0132] In this embodiment, the recalibrated cross-modal interaction unit processes the input data according to the new feature weights, outputting a new fused feature vector that contains multi-source data information with adjusted weights. This vector is then input into the avalanche risk prediction model, which generates incremental learning parameters based on the new input data. These incremental learning parameters reflect the direction and magnitude of model adjustments needed due to environmental changes (such as the discovery of potential trigger points).

[0133] Then, the incrementally learned parameters are fused with the original model parameters using gradient interpolation to generate an environmentally adaptive avalanche risk prediction model. For example, if the weight of a neuron connection in the original model parameters is 0.3, and the corresponding adjustment value in the incrementally learned parameters is 0.1, the two are fused using a gradient interpolation fusion algorithm to obtain a new weight value, such as 0.35. Performing this fusion operation on all parameters in the avalanche risk prediction model generates the environmentally adaptive avalanche risk prediction model. This environmentally adaptive avalanche risk prediction model can better adapt to the actual environmental changes in the target area, improving the accuracy of avalanche risk prediction.

[0134] Step S370: The parameters of the environmentally adaptive avalanche risk prediction model are synchronized to all online early warning terminals through edge computing nodes to complete the closed-loop iteration of the early warning logic.

[0135] In this embodiment, the edge computing nodes are located close to the target area and have the ability to process data quickly. They can send new model parameters to various online early warning terminals, which are distributed around the target area, such as at the entrance of tourist attractions or near mountain settlements. When these early warning terminals receive the new model parameters, they can use the updated environmentally adaptive avalanche risk prediction model to perform more accurate avalanche risk prediction and early warning, thus completing the closed-loop iterative process of the entire early warning logic and continuously improving the accuracy and effectiveness of avalanche risk prediction and early warning.

[0136] For example, in one possible implementation, the method further includes:

[0137] After generating the snow structure anomaly index, perform the following steps:

[0138] Step S410: Construct a three-dimensional motion vector field based on the snow layer displacement rate, extract anisotropic displacement distribution features, and couple the anisotropic displacement distribution features with the terrain potential energy gradient in the terrain spatial feature set to generate a potential energy-motion correlation matrix.

[0139] In this embodiment, since the snow layer displacement rate was already obtained during the previous analysis of snow structure anomaly indicators, such as the vertical and horizontal displacement rates of the snow layer obtained at different monitoring points, a three-dimensional motion vector field is constructed with the monitoring point as the origin, the vertical direction as the z-axis, and the horizontal directions as the x and y axes, respectively. In this environmentally adaptive avalanche risk prediction vector field, each point has a vector representing the direction and magnitude of the snow layer's movement. Then, anisotropic displacement distribution characteristics are extracted. Anisotropy refers to the different characteristics of displacement in different directions. For example, in a certain hillside area, due to the influence of factors such as terrain and wind direction, the snow layer's displacement rate in the downward direction along the hillside (x-axis direction) may be greater than that in the direction perpendicular to the hillside (y-axis direction). This difference in displacement rate in different directions is the anisotropic displacement distribution characteristic.

[0140] The geopotential energy gradient, a concentration of topographic spatial features, reflects the influence of topography on the potential kinetic energy of snow cover. For example, within a certain grid cell, the geopotential energy gradient is 0.15, while the anisotropic displacement distribution characteristics of this region indicate a significant displacement trend of snow cover in a certain direction. Considering how the geopotential energy gradient affects the anisotropic motion of snow cover, a potential energy-motion correlation matrix is ​​generated. Each element in this matrix represents the correlation between potential energy and motion under specific topographic and snow cover displacement characteristics.

[0141] Step S420: Calculate the snow stress concentration index for different altitude zones based on the potential energy-motion correlation matrix, identify the boundary of the stress exceeding the limit region, and activate the UAV cluster to perform infrared thermal imaging scanning on the specified altitude zone according to the boundary of the stress exceeding the limit region to obtain the temperature field distribution data inside the snow.

[0142] In this embodiment, the target area can be divided into different altitude zones according to elevation, for example, each zone is 100 meters high. For each altitude zone, the snow stress concentration index is calculated based on the data in the potential energy-motion correlation matrix and the physical properties of the snow (such as density, which can be obtained from previous snow layer physical data). Assuming that in a certain altitude zone, based on the correlation in the potential energy-motion correlation matrix and relevant information such as a snow density of 0.2 g / cm³, a series of calculations yields a snow stress concentration index of 0.4. The boundary of the stress-exceeding region is then identified, i.e., the boundary of the region where the snow stress concentration index exceeds a certain preset critical value. Assuming the preset critical value is 0.3, then the region with a snow stress concentration index of 0.4 is identified as the stress-exceeding region, and the boundary range of this stress-exceeding region is determined.

[0143] After receiving instructions, the drone swarm flies to the altitude zone corresponding to the area exceeding stress limits, such as an altitude of 1500-1600 meters. The infrared thermal imaging equipment onboard the drones scans the snow to obtain data on the internal temperature field distribution. Due to the thermal conductivity and other properties of snow, the temperature varies at different depths. Infrared thermal imaging can reveal the temperature distribution from the snow surface to a certain depth. For example, assuming the surface temperature of snow in a certain area is -3℃, the temperature at a depth of 10 cm is -5℃, and so on, this data represents the temperature at different depths.

[0144] Step S430: The temperature field distribution data inside the snow cover and the temperature gradient change rate in the snow cover state vector are assimilated to generate a corrected snow body stability evaluation index. The corrected snow body stability evaluation index is then input into a physical constraint neural network for forward inference to update the numerical distribution of the snow cover state vector.

[0145] In this embodiment, the temperature gradient change rate in the snow state vector was previously calculated based on a predefined value, such as -0.2℃ / cm from the snow surface to a certain depth. Now, with more detailed data on the internal temperature field distribution of the snow, this data is fused with the original temperature gradient change rate. Assuming a weighted averaging method is used, assigning different weights based on the importance of data at different depths, a new temperature-related index is calculated. This index is then combined with other factors in the snow state vector (such as interlayer shear strength coefficient, liquid water content, etc.) to recalculate the snow stability assessment index. For example, the original snow stability assessment index was 0.064, and after data assimilation, the revised snow stability assessment index is 0.055.

[0146] The physically constrained neural network (PCN) was previously trained on a large amount of snow-related data. It can process the corrected snow stability assessment index input according to the internal neuron connection weights and activation functions. For example, a neuron in the PCN receives the corrected snow stability assessment index 0.055, multiplies it by a connection weight of 0.5 to obtain 0.0275, and then passes it through an activation function (such as the ReLU function). If 0.0275 is greater than 0, it outputs 0.0275. After processing by multiple layers of neurons, a new value is obtained, thereby updating the numerical distribution of the snow state vector.

[0147] Step S440: The updated snow cover state vector is re-input into the cross-modal interaction unit to trigger the real-time recalculation of the fused feature vector, and the weights of the fully connected layer of the avalanche risk prediction model are adjusted according to the recalculated fused feature vector to generate a dynamically optimized risk probability output.

[0148] The cross-modal interaction unit has previously fused multi-source data, including meteorological feature sequences, topographic spatial feature sets, and snow cover state vectors. Now, with the updated snow cover state vector as input, feature fusion and interaction calculations are re-performed. For example, previously in the cross-modal interaction unit, the snow cover state vector underwent cross-attention mechanism calculations with other features; now, these calculations are re-performed based on the updated snow cover state vector, thus triggering real-time recalculation of the fused feature vector.

[0149] The weights of the fully connected layers in the avalanche risk prediction model were fixed or based on previous calculations during training and use. Now, based on the recalculated fused feature vector—for example, if the importance of some features has changed—the neuron connection weights in the fully connected layers are adjusted accordingly. Suppose a neuron originally had a connection weight of 0.3; due to the change in the fused feature vector, its weight is adjusted to 0.35 based on the new calculation. After calculation by the fully connected layers, a dynamically optimized risk probability output is generated. For example, the original risk probability output was 0.55, and after adjustment, it becomes 0.6.

[0150] Step S450: The dynamically optimized risk probability output is fused with the original early warning signal to generate a confidence-weighted final early warning, and the multi-level signal release strategy of the early warning terminal is controlled according to the final early warning level. Simultaneously, the data stream generated during the optimization process is written into the training set for storage.

[0151] In this embodiment, the original warning signal is generated based on the previous output of the avalanche risk prediction model, for example, the warning signal corresponding to a previous risk probability output of 0.55. Now, the dynamically optimized risk probability output of 0.6 is fused with the original warning signal. Assuming a weighted fusion method is used, different weights are assigned according to the confidence levels of the two signals, for example, the confidence level of the dynamically optimized risk probability output is 0.6, and the confidence level of the original warning signal is 0.4, to calculate the final warning level. If the final warning level is a high-risk level, then in the multi-level signal release strategy of the warning terminal, the alarm frequency will be increased to once every 5 minutes, and the emergency response instructions will include stricter measures, such as accelerating the evacuation of surrounding residents.

[0152] Throughout the optimization process, from analyzing anomaly indicators of snow cover structure to adjusting the final warning level, a series of data were generated, such as the internal temperature field distribution data of snow cover, the updated snow cover state vector, and the recalculated fusion feature vector. These data were then written into the training set according to a defined format and rules for use in subsequent model training and optimization, further improving the accuracy and adaptability of the avalanche risk prediction model.

[0153] Figure 2 The illustration shows exemplary hardware and software components of a deep learning-based avalanche early warning model building system 100 that can implement the ideas of this application, according to some embodiments of this application. For example, a processor 120 can be used in the deep learning-based avalanche early warning model building system 100 and to perform the functions in this application.

[0154] The deep learning-based avalanche early warning model construction system 100 can be a general-purpose server or a special-purpose server; both can be used to implement the deep learning-based avalanche early warning model construction method of this application. Although only one server is shown in this application, for convenience, the functions described in this application can be implemented in a distributed manner on multiple similar platforms to balance the load.

[0155] For example, the deep learning-based avalanche early warning model building system 100 may include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and various forms of storage media 140, such as a disk, ROM, or RAM, or any combination thereof. Exemplarily, the deep learning-based avalanche early warning model building system 100 may also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The methods of this application can be implemented according to these program instructions. The deep learning-based avalanche early warning model building system 100 also includes an input / output (I / O) interface 150 between the computer and other input / output devices.

[0156] For ease of explanation, only one processor is described in the deep learning-based avalanche early warning model construction system 100. However, it should be noted that the deep learning-based avalanche early warning model construction system 100 of this application may also include multiple processors. Therefore, the steps executed by one processor as described in this application may also be executed jointly by multiple processors or individually. For example, if the processor of the deep learning-based avalanche early warning model construction system 100 executes steps A and B, it should be understood that steps A and B may also be executed jointly by two different processors or individually by one processor. For example, the first processor executes step A, the second processor executes step B, or the first processor and the second processor jointly execute steps A and B.

[0157] Furthermore, this embodiment of the invention also provides a readable storage medium, wherein computer-executable instructions are preset in the readable storage medium, and when the processor executes the computer-executable instructions, the above-mentioned method for constructing an avalanche early warning model based on deep learning is implemented.

[0158] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A method for constructing an avalanche early warning model based on deep learning, characterized in that, The method includes: Acquire multi-source environmental monitoring data for the target area, including meteorological time-series data, topographic spatial data, and snow cover physical data; The meteorological time series data is aligned to the time dimension to generate a meteorological feature sequence; the topographic spatial data is spatially gridded to generate a topographic spatial feature set; and the physical parameters of the snow layer physical data are extracted to generate a snow state vector. The meteorological feature sequence, topographic spatial feature set, and snow cover state vector are input into a deep learning network model to generate a fused feature vector. The deep learning network model includes a temporal attention unit, a spatial convolution unit, and a cross-modal interaction unit. A training set is constructed based on historical avalanche event annotation data. The training set is then used to dynamically optimize the weights of the fused feature vector to generate an avalanche risk prediction model. The system receives real-time environmental monitoring data and outputs the risk level and warning trigger threshold of the target area through the avalanche risk prediction model. When the real-time risk value exceeds the warning trigger threshold, a multi-level warning signal is generated. The system receives real-time environmental monitoring data, outputs the risk level and early warning trigger threshold of the target area through the avalanche risk prediction model, and generates multi-level early warning signals when the real-time risk value exceeds the early warning trigger threshold, including: Real-time collection of current meteorological time-series data, current topographic spatial data, and current snow cover physical data of the target area; The current meteorological time series data is aligned in time dimension to generate a real-time meteorological feature sequence; the current terrain spatial data is spatially gridded to generate a real-time terrain spatial feature set; and the current snow layer physical data is physical parameter extracted to generate a real-time snow state vector. The real-time meteorological feature sequence, real-time terrain spatial feature set, and real-time snow cover state vector are input into the avalanche risk prediction model, and the real-time risk value of the target area and the dynamically updated warning trigger threshold are output. The risk level range is divided according to the difference between the real-time risk value and the warning trigger threshold, and a risk level label corresponding to the risk level range is generated. When the real-time risk value exceeds the warning trigger threshold, a preset warning level mapping table is activated according to the risk level label to generate a multi-level warning signal containing different alarm frequencies and emergency response instructions. After generating a multi-level early warning signal when the real-time risk value exceeds the early warning trigger threshold, the method further includes: Real-time surface deformation monitoring data and meteorological radar reflectivity data of the target area are collected to generate a dynamic environment verification dataset. The dynamic environment verification dataset is input into the pre-trained snow motion recognition model to extract the snow layer displacement rate and surface crack propagation trajectory, and generate snow structure anomaly indicators. Based on the spatiotemporal matching degree analysis of the snow structure anomaly index and the risk level output by the avalanche risk prediction model, a model prediction deviation coefficient is generated. When the model prediction deviation coefficient exceeds a preset threshold, the terrain scanning device is activated to reconstruct the three-dimensional point cloud of the target area, generate a surface micro-topography change map, and compare the surface micro-topography change map with the feature terrain template before the occurrence of historical avalanche events to identify the coordinates of potential trigger points. The warning trigger threshold of the avalanche risk prediction model is adjusted according to the coordinates of the potential trigger points, an adaptive threshold update instruction is generated, the feature weight allocation strategy of the current environmental monitoring data is corrected based on the adaptive threshold update instruction, and the updated feature weights are sent back to the cross-modal interaction unit for feature recalibration. The new fusion feature vector is output by the recalibrated cross-modal interaction unit to drive the avalanche risk prediction model to generate incremental learning parameters. The incremental learning parameters are then fused with the original model parameters by gradient interpolation to generate an environmentally adaptive avalanche risk prediction model. The parameters of the environmentally adaptive avalanche risk prediction model are synchronized to all online early warning terminals through edge computing nodes, completing the closed-loop iteration of the early warning logic; The method further includes: After generating the snow structure anomaly index, perform the following steps: A three-dimensional motion vector field is constructed based on the snow layer displacement rate, anisotropic displacement distribution features are extracted, and the anisotropic displacement distribution features are coupled with the terrain potential energy gradient in the terrain spatial feature set to generate a potential energy-motion correlation matrix. Based on the potential energy-motion correlation matrix, the stress concentration index of snow body in different altitude zones is calculated, the boundary of stress over-limit area is identified, and the UAV cluster is activated to perform infrared thermal imaging scanning on the specified altitude zone according to the boundary of stress over-limit area to obtain the temperature field distribution data inside the snow. The data of the temperature field distribution inside the snow is assimilated with the rate of change of the temperature gradient in the snow state vector to generate a corrected snow stability evaluation index. The corrected snow stability evaluation index is then input into a physical constraint neural network for forward inference to update the numerical distribution of the snow state vector. The updated snow cover state vector is re-input into the cross-modal interaction unit, triggering real-time recalculation of the fused feature vector. The weights of the fully connected layer of the avalanche risk prediction model are adjusted according to the recalculated fused feature vector to generate a dynamically optimized risk probability output. The dynamically optimized risk probability output is fused with the original early warning signal to generate a confidence-weighted final early warning, and the multi-level signal release strategy of the early warning terminal is controlled according to the final early warning level. The data stream generated during the optimization process is simultaneously written into the training set for storage.

2. The method for constructing an avalanche early warning model based on deep learning according to claim 1, characterized in that, The step of performing time-dimensional alignment processing on the meteorological time-series data to generate a meteorological feature sequence includes: The missing time nodes in the meteorological time series data are identified, and linear interpolation compensation is performed on the missing time nodes based on the pressure gradient change rate and temperature change trend of adjacent time windows to generate a continuous meteorological data stream. Key meteorological indicators are extracted from the continuous meteorological data stream, including cumulative snowfall, peak wind speed duration, and diurnal temperature range fluctuation. Key meteorological indicators within each preset time unit are aggregated locally using a sliding window mechanism to generate time-related meteorological feature fragments. The meteorological feature fragments are modeled temporally using a bidirectional long short-term memory network to capture the correlation between forward cumulative effects and backward abrupt changes, generating meteorological feature sequences containing multidimensional temporal context.

3. The method for constructing an avalanche early warning model based on deep learning according to claim 2, characterized in that, The step of performing spatial gridding processing on the terrain spatial data to generate a terrain spatial feature set includes: Convert the geographic coordinate system of the target area to a raster coordinate system and generate spatial grid cells according to a preset resolution; Within each spatial grid cell, slope change rate, aspect dispersion, and surface roughness index are extracted as terrain attribute data. Based on satellite remote sensing image data, land cover types are identified, and vegetation density distribution and the proportion of exposed rock area are mapped to corresponding grid cells; Construct a three-dimensional terrain topology map, calculate the terrain energy gradient based on the elevation difference of adjacent grid cells, and generate the intensity distribution of potential motion paths by combining the avalanche motion mechanics model. Multi-scale feature extraction of the terrain energy gradient and potential motion path intensity distribution is performed using a dilated convolutional neural network to generate a terrain spatial feature set containing spatial correlation features.

4. The method for constructing an avalanche early warning model based on deep learning according to claim 3, characterized in that, The step of extracting physical parameters from the physical data of the snow layer to generate a snow state vector includes: Snow layer profile data is acquired through a ground sensor array, including snow layer thickness distribution, crystal morphology classification, and liquid water content. The vertical density gradient is calculated based on the snow layer thickness distribution, and the interlayer shear strength coefficient is determined in combination with the crystal morphology classification. The surface temperature field of snow was retrieved based on thermal infrared imaging data, and the rate of change of temperature gradient inside the snow layer was calculated using the heat conduction equation. By integrating the interlayer shear strength coefficient, liquid water content, and temperature gradient change rate, a snow body stability assessment index is constructed. The snow stability assessment index is nonlinearly mapped by a physical constraint neural network to generate a snow state vector that reflects the internal structure of the snow.

5. The method for constructing an avalanche early warning model based on deep learning according to claim 1, characterized in that, The step of inputting the meteorological feature sequence, topographic spatial feature set, and snow cover state vector into a deep learning network model to generate a fused feature vector includes: The meteorological feature sequence is input into the temporal attention unit of the deep learning network model. By calculating the attention weight distribution in the time dimension, the meteorological indicators at different time nodes in the meteorological feature sequence are dynamically weighted to generate weighted meteorological time-series features. The terrain spatial feature set is input into the spatial convolutional unit of the deep learning network model, and the spatial context features of the terrain spatial feature set are extracted using multi-level dilated convolutional kernels to capture the terrain correlation between adjacent grid units and generate multi-scale terrain spatial features. The snow cover state vector and the weighted meteorological time series features are aligned and concatenated according to their feature dimensions to generate the first fusion intermediate feature; The first fused intermediate feature and the multi-scale terrain spatial feature are input into the cross-modal interaction unit, and the cross-modal feature correlation degree between the first fused intermediate feature and the multi-scale terrain spatial feature is calculated through a cross-attention mechanism to generate cross-modal interaction features. Channel attention weighted filtering is applied to the cross-modal interaction features to retain feature channels that are strongly correlated with avalanche triggering conditions, thereby generating optimized cross-modal interaction features; The optimized cross-modal interaction features are input into a fully connected layer for nonlinear transformation and feature dimensionality reduction to generate the fused feature vector.

6. The method for constructing an avalanche early warning model based on deep learning according to claim 5, characterized in that, The step of calculating the cross-modal feature correlation degree between the first fused intermediate feature and the multi-scale terrain spatial feature through a cross-attention mechanism to generate cross-modal interactive features includes: The first fused intermediate features are used to generate a query vector sequence through linear projection, and the multi-scale terrain spatial features are used to generate a key vector sequence and a value vector sequence through linear projection. Calculate the attention score matrix between the query vector sequence and the key vector sequence, and perform weighted aggregation on the value vector sequence to generate a spatially enhanced feature sequence; The spatial enhancement feature sequence is residually concatenated with the first fusion intermediate feature to generate the second fusion intermediate feature; The second fused intermediate feature is input into a multilayer perceptron for nonlinear mapping to generate high-dimensional interactive features; The high-dimensional interaction features are gated and fused with the snow state vector. The weight ratio of each dimension in the high-dimensional interaction features is dynamically adjusted according to the current value of the snow state vector to generate the cross-modal interaction features.

7. The method for constructing an avalanche early warning model based on deep learning according to claim 1, characterized in that, The step of dynamically optimizing the weights of the fused feature vector using the training set to generate an avalanche risk prediction model includes: The fused feature vector is input into the first fully connected layer and the second fully connected layer of the deep learning network model, respectively. The first fully connected layer outputs the predicted avalanche occurrence probability, and the second fully connected layer outputs the predicted avalanche trigger time. The cross-entropy loss is calculated by comparing the predicted avalanche occurrence probability with the binary classification labels in the historical avalanche event annotation data to generate a classification loss value. The mean squared error is calculated by comparing the predicted avalanche trigger time with the actual trigger timestamps in the historical avalanche event annotation data to generate a time regression loss value. The classification loss value and the time regression loss value are input into the dynamic weight allocator, and classification weight coefficients and time weight coefficients are generated according to the ratio of the classification loss value to the time regression loss value in the current training batch. Multiply the classification loss value by the classification weight coefficient, and multiply the time regression loss value by the time weight coefficient to generate a weighted classification loss component and a weighted time loss component. The weighted classification loss component and the weighted time loss component are added together to generate a joint optimization loss value; Based on the joint optimization loss value, backpropagation gradient updates are performed on the parameters of the first fully connected layer, the second fully connected layer, and the deep learning network model to iteratively adjust the mapping relationship between the fused feature vector and the avalanche risk prediction result. When the decrease in the joint optimization loss value is less than the convergence threshold during a preset number of consecutive training iterations, the parameters of the deep learning network model are frozen to generate the avalanche risk prediction model.

8. A deep learning-based avalanche early warning model construction system, characterized in that, The deep learning-based avalanche early warning model construction system includes a processor and a memory, the memory and the processor being connected, the memory being used to store programs, instructions or code, and the processor being used to execute the programs, instructions or code in the memory to implement the deep learning-based avalanche early warning model construction method according to any one of claims 1-7.

Images