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

Abstract

The invention provides a deep learning-based avalanche early warning model construction method and system, and the method comprises the steps: firstly obtaining multi-source environment monitoring data, including meteorological time sequence, topographic space and accumulated snow layer physical data, of a target region, carrying out the time dimension alignment of the meteorological time sequence data to generate a feature sequence, carrying out the meshing of the topographic space data to generate a feature set, and carrying out the construction of an avalanche early warning model; the method comprises the following steps: extracting parameters from accumulated snow layer physical data to generate a state vector, inputting a deep learning network model containing time sequence attention, spatial convolution and cross-modal interaction units, generating a fusion feature vector, constructing a training set based on historical avalanche event annotation data, performing dynamic weight optimization on the fusion feature vector, and generating an avalanche risk prediction model. And finally, receiving current monitoring data in real time, outputting a risk level and an early warning trigger threshold value by the avalanche risk prediction model, and generating a multi-level early warning signal when a real-time risk value exceeds the threshold value, thereby realizing accurate avalanche early warning.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a method and system for constructing an avalanche warning model based on deep learning. Background Art

[0002] In the field of avalanche warning technology, traditional avalanche warning methods often rely solely on meteorological or topographic data, neglecting the critical factor of the physical state of the snowpack. This results in inaccurate and incomplete warnings. Furthermore, due to a lack of effective integration and in-depth mining of multi-source data, traditional methods struggle to capture the complex nonlinear relationships in avalanche risk, limiting improvements in warning accuracy.

[0003] Furthermore, related 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 can also lead to the loss of key information, thus affecting the accuracy of early warnings. Summary of the Invention

[0004] In view of the above-mentioned problems, in combination with the first aspect of the present invention, an embodiment of the present invention provides a method for constructing an avalanche warning model based on deep learning, the method comprising: Acquiring multi-source environmental monitoring data of a target area, wherein the multi-source environmental monitoring data includes meteorological time series data, terrain spatial data, and snow layer physical data; Performing time dimension alignment processing on the meteorological time series data to generate a meteorological feature sequence, performing spatial gridding processing on the terrain spatial data to generate a terrain spatial feature set, and extracting physical parameters from the snow layer physical data to generate a snow state vector; Inputting the meteorological feature sequence, the terrain spatial feature set, and the snow state vector into a deep learning network model to generate a fused feature vector, wherein the deep learning network model includes a temporal attention unit, a spatial convolution unit, and a cross-modal interaction unit; Building a training set based on historical avalanche event annotated data, and using the training set to dynamically optimize the weight of the fused feature vector to generate an avalanche risk prediction model; The current environmental monitoring data is received in real time, and the risk level and warning trigger threshold of the target area are output through the avalanche risk prediction model. When the real-time risk value exceeds the warning trigger threshold, a multi-level warning signal is generated.

[0005] On the other hand, an embodiment of the present invention also provides an avalanche warning model construction system based on deep learning, including a processor and a machine-readable storage medium, wherein the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0006] Based on the above aspects, the embodiment of the present application obtains the meteorological time series data, terrain spatial data and snow layer physical data of the target area. On this basis, by aligning the time dimension of the meteorological time series data, spatial gridding of the terrain spatial data and extracting the physical parameters of the snow layer physical data, the key features in each data source are effectively mined, forming a meteorological feature sequence, a terrain spatial feature set and a snow state vector. Then, a deep learning network model including a time series attention unit, a spatial convolution unit and a cross-modal interaction unit is introduced, which can automatically learn and capture the complex nonlinear relationship between meteorology, terrain and snow state. Through cross-modal interaction, the deep fusion of multi-source features is achieved, and a representative fusion feature vector is generated, which not only improves the richness and accuracy of feature expression, but also enhances the comprehensive perception of avalanche risk factors. In the model training stage, the fusion feature vector is dynamically weighted based on the training set constructed based on the annotated data of historical avalanche events, so that the avalanche risk prediction model can adaptively adjust the contribution of each feature to the prediction results, thereby improving the generalization ability and prediction accuracy of the model. Ultimately, by receiving real-time environmental monitoring data and utilizing a trained avalanche risk prediction model to output the target area's risk level and warning trigger threshold, the system achieves real-time monitoring and dynamic early warning of avalanche risk. When the real-time risk value exceeds the warning trigger threshold, a multi-level warning signal is rapidly generated, providing timely and effective decision-making support to relevant departments and personnel, significantly reducing the potential losses and impacts of avalanche disasters. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 This is a schematic diagram of the execution flow of the method for constructing an avalanche warning model based on deep learning provided by an embodiment of the present invention.

[0008] Figure 2 Schematic diagram of exemplary hardware and software components of a deep learning-based avalanche warning model construction system provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0009] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 This is a flow chart of a method for constructing an avalanche warning model based on deep learning provided by an embodiment of the present invention. The method for constructing an avalanche warning model based on deep learning is introduced in detail below.

[0010] Step S110 , obtaining multi-source environmental monitoring data of the target area, wherein the multi-source environmental monitoring data includes meteorological time series data, terrain spatial data, and snow layer physical data.

[0011] For example, a pre-configured environmental monitoring system can be used to monitor avalanche-prone areas in a mountainous area. Specifically, multiple meteorological monitoring stations can be set up in the area. These stations can record meteorological data such as air pressure, temperature, snowfall, and wind speed every hour for a long period of time, forming a meteorological time series. For another example, topographic spatial data can be obtained through a combination of aerial mapping and ground surveying. Aerial mapping captures large-scale terrain contour data, while ground surveying teams use specialized instruments to accurately measure detailed terrain data at key locations, encompassing the topographic and geomorphological information of the entire target area. For physical data on snow accumulation, sensors can be embedded at various locations to monitor snow conditions. These sensors can capture information such as snow layer profile data. For example, sensors can be installed at different heights and in different directions on a hillside to comprehensively understand the snow accumulation.

[0012] 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.

[0013] In this embodiment, in the meteorological time series data obtained previously, there will occasionally be a situation where a certain meteorological monitoring station does not record data at a certain time point due to equipment failure or communication problems, that is, there is a missing time node. For example, at 10 o'clock in the morning of a certain day, a meteorological monitoring station located on the mountainside did not record the air pressure data at that time. Therefore, the missing time node can be compensated by linear interpolation based on the pressure gradient change rate and temperature change trend of adjacent time windows. Assuming that the air pressure at adjacent 9 o'clock and 11 o'clock is 1000 hPa and 1005 hPa respectively, and the temperature rises from -5°C at 9 o'clock to -3°C at 11 o'clock, a reasonable air pressure value is calculated based on the above data to compensate for the missing air pressure data at 10 o'clock in the morning, thereby generating a continuous meteorological data stream. Key meteorological indicators can then be extracted from this continuous meteorological data stream, such as accumulated snowfall (assuming the target area has accumulated 50 cm of snowfall over the past 24 hours); peak wind speed duration (e.g., wind speeds exceeding 15 m / s for 3 hours); and diurnal temperature fluctuations (e.g., 8°C from a daytime high of -2°C to a nighttime low of -10°C). These key meteorological indicators within each preset time unit (here, 12 hours) are then aggregated using a sliding window mechanism to generate temporally correlated meteorological feature segments. For example, accumulated snowfall, peak wind speed duration, and diurnal temperature fluctuations within the first 12 hours are aggregated 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. This model can, for example, capture the cumulative effect of snowfall in the previous 12 hours on temperature changes in the next 12 hours, as well as sudden meteorological changes. Ultimately, a meteorological feature sequence with multidimensional temporal context is generated.

[0014] Furthermore, the geographic coordinate system of the target area can be converted to a grid coordinate system, and spatial grid cells can be generated at a preset resolution of 10 meters by 10 meters. Within each spatial grid cell, the slope change rate is extracted using topographic survey instruments. For example, the slope change rate from the foot of a mountain to the top of a certain grid cell is 0.3. The slope aspect dispersion is determined by assuming that the slope aspect within the cell has a set dispersion from northeast to southeast. The surface roughness index is measured to be 0.5 due to factors such as rocks and vegetation within the grid cell. At the same time, the surface cover type is identified based on satellite remote sensing imagery data. It is found that the vegetation density distribution within a certain grid cell is 30%, and the exposed rock area accounts for 20%. These data are then mapped to the corresponding grid cell. A three-dimensional terrain topological relationship map is then constructed, and the terrain potential energy gradient is calculated based on the elevation difference between adjacent grid cells. For example, if the elevation difference between adjacent grid cells is 5 meters, the potential motion path intensity distribution is calculated based on the avalanche motion mechanics model. Finally, multi-scale feature extraction is performed on these terrain energy gradients and potential motion path intensity distributions through a dilated convolutional neural network to generate a terrain spatial feature set containing spatial correlation features.

[0015] Furthermore, ground-based sensor arrays can be used to obtain snow layer profile data. For example, at a given monitoring point, the snow layer thickness distribution from the surface to the depth is 10 cm, 20 cm, 30 cm, and so on; the crystal morphology is classified as hexagonal, etc.; and the liquid water content is 5%. Based on the snow layer thickness distribution, the vertical density gradient is calculated, assuming a surface snow density of 0.1 g / cm³ and a deep snow density of 0.3 g / cm³. The interlayer shear strength coefficient is determined based on 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 snowpack is inverted based on thermal infrared imaging data, resulting in a surface temperature of -5°C. The temperature gradient rate within the snowpack is calculated using the heat conduction equation, assuming a temperature gradient of -1°C / cm³ from the surface to a depth of 10 cm. The interlayer shear strength coefficient, liquid water content, and temperature gradient rate are combined to construct a snow stability assessment index. Then, the snow stability evaluation index is nonlinearly mapped through a physical constraint neural network to generate a snow state vector reflecting the internal structural state of the snow.

[0016] Step S130: Input the meteorological feature sequence, terrain spatial feature set, and snow state vector into a deep learning network model to generate a fusion feature vector, wherein the deep learning network model includes a temporal attention unit, a spatial convolution unit, and a cross-modal interaction unit.

[0017] In this embodiment, a previously generated meteorological feature sequence can be input into the temporal attention unit of the deep learning network model. For example, different time nodes in the meteorological feature sequence contain different meteorological indicators, such as snowfall and wind speed at different times of the day. Thus, by calculating the attention weight distribution in the time dimension, it is assumed that during periods with a higher probability of avalanche occurrence, such as periods with heavy snowfall and fast wind speed, the corresponding meteorological indicators are given higher attention weights. In this way, the meteorological indicators at different time nodes in the meteorological feature sequence are dynamically weighted to generate weighted meteorological time series features.

[0018] The terrain spatial feature set can then be fed into the spatial convolution unit of the deep learning network model, where multi-level dilated convolution kernels are used to extract spatial contextual features from the terrain spatial feature set. For example, in an area with complex terrain, multi-level dilated convolution kernels can better capture the terrain correlations between adjacent grid cells, such as the slope and aspect relationships between a grid cell on a hillside and its neighbors, generating multi-scale terrain spatial features.

[0019] Next, the snow state vector and the weighted meteorological time series features are aligned and spliced in terms of feature dimensions. For example, the snow stability assessment index in the snow state vector and the snowfall, wind speed and other indicators in the weighted meteorological time series features are aligned and spliced in terms of dimensions to generate the first fusion intermediate feature.

[0020] The first fused intermediate features and the multi-scale terrain spatial features are then input into the cross-modal interaction unit. A cross-attention mechanism is used to calculate the cross-modal feature correlation between the first fused intermediate features and the multi-scale terrain spatial features. For example, linear projection is performed on the first fused intermediate features to generate a query vector sequence, and linear projection is performed on the multi-scale terrain spatial features to generate a key vector sequence and a value vector sequence. An attention score matrix is calculated between the query vector sequence and the key vector sequence. It is assumed that weighted aggregation of the value vector sequence based on this matrix will result in a spatially enhanced feature sequence that better reflects the relationship between snow cover state and terrain spatial features. The spatially enhanced feature sequence is then residually connected with the first fused intermediate features to generate a second fused intermediate feature. This second fused intermediate feature is then input into a multi-layer perceptron for nonlinear mapping to generate a high-dimensional interaction feature. Finally, the high-dimensional interaction feature is gated and fused with the snow state vector. The weights of the various dimensions in the high-dimensional interaction feature are dynamically adjusted based on the current value of the snow state vector. For example, when the snow stability assessment index in the snow state vector is low, higher weights are assigned to the terrain-related high-dimensional interaction feature dimensions to generate a cross-modal interaction feature.

[0021] Finally, the cross-modal interaction features are filtered using a weighted channel-attention approach to retain feature channels that are strongly correlated with avalanche triggering conditions. For example, these channels are directly related to snowfall, slope, and snow stability. This generates optimized cross-modal interaction features. These optimized cross-modal interaction features are then fed into a fully connected layer for nonlinear transformation and feature dimensionality reduction to generate a fused feature vector.

[0022] Step S140: constructing a training set based on historical avalanche event annotated data, and using the training set to dynamically optimize the weight of the fused feature vector to generate an avalanche risk prediction model.

[0023] In this example, historical avalanche event data from the target area over the past several years can be collected, detailing the time, location, and environmental factors of each avalanche. For example, an avalanche occurred at a specific location on a hillside, with cumulative snowfall reaching 80 centimeters, a wind speed of 20 meters per second, and a southeast slope. The data also includes whether an avalanche occurred (a binary label) and the timestamp of the actual avalanche trigger.

[0024] 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 value of the avalanche probability. Assume that for a certain input fused feature vector, the predicted value of the avalanche probability output by the first fully connected layer is 0.6; the second fully connected layer outputs the predicted value of the avalanche trigger time, for example, the predicted avalanche trigger time is 3 pm.

[0025] Next, we perform a cross-entropy loss calculation on the predicted avalanche probability and the binary classification labels in the historical avalanche event annotated data. For example, if an avalanche actually occurs but the predicted probability is low, a large classification loss will be generated. We also perform a mean squared error calculation on the predicted avalanche trigger time and the actual trigger timestamps in the historical avalanche event annotated data. If the predicted avalanche trigger time differs significantly from the actual trigger time, a large time regression loss will be generated.

[0026] The classification loss and temporal regression loss are then fed into the dynamic weight allocator, which generates a classification weight coefficient and a temporal weight coefficient based on the ratio of the classification loss to the temporal regression loss in the current training batch. For example, when the classification loss is large, the classification weight coefficient might be set to 0.7 and the temporal weight coefficient to 0.3.

[0027] Next, the classification loss is multiplied by the classification weight coefficient, and the temporal regression loss is multiplied by the temporal weight coefficient to generate a weighted classification loss component and a weighted temporal loss component. These weighted classification loss components and weighted temporal loss components are then added together to generate the joint optimization loss value.

[0028] Finally, backpropagation gradient updates are performed on the parameters of the first and second fully connected layers and the deep learning network model based on the joint optimization loss value, iteratively adjusting the mapping 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 a convergence threshold (e.g., 0.01) over a preset number of training iterations (e.g., 10), the parameters of the deep learning network model are frozen, generating an avalanche risk prediction model.

[0029] Step S150, receiving current environmental monitoring data in real time, outputting the risk level and warning trigger threshold of the target area through the avalanche risk prediction model, and generating a multi-level warning signal when the real-time risk value exceeds the warning trigger threshold.

[0030] In this embodiment, while the monitoring system in the target area is continuously operating, it can collect current meteorological time-series data, topographic spatial data, and snow cover physical data in real time. For current meteorological time-series data, the same meteorological monitoring station records data such as air pressure, temperature, snowfall, and wind speed in real time. Current topographic spatial data is updated through regular topographic measurements, and snow cover physical data is acquired in real time by embedded sensors.

[0031] The current meteorological time series data is aligned in time to generate a real-time meteorological feature sequence. For example, at a certain moment, the snowfall in the real-time meteorological feature sequence increases dramatically in a short period of time. The current terrain spatial data is spatially gridded to generate a real-time terrain spatial feature set. For example, new snowfall may cause a slight change in the slope change rate within a grid cell. The physical parameters of the current snow layer are extracted to generate a real-time snow state vector. For example, the liquid water content of the snow increases due to temperature changes.

[0032] The real-time meteorological feature sequence, real-time terrain spatial feature set, and real-time snow state vector are input into the avalanche risk prediction model, which outputs 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.

[0033] The risk level interval is divided according to the difference between the real-time risk value and the warning trigger threshold, and the risk level label corresponding to the risk level interval is 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.

[0034] When the real-time risk value exceeds the warning trigger threshold, a pre-set warning level mapping table is activated based on the risk level tag, generating a multi-level warning signal with 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 notifying nearby residents to evacuate to a safe area as soon as possible and restricting traffic in mountainous areas.

[0035] Based on the above steps, the embodiment of the present application obtains the meteorological time series data, terrain spatial data and snow layer physical data of the target area. On this basis, by aligning the time dimension of the meteorological time series data, spatial gridding of the terrain spatial data and extracting the physical parameters of the snow layer physical data, the key features in each data source are effectively mined, forming a meteorological feature sequence, a terrain spatial feature set and a snow state vector. Then, a deep learning network model including a time series attention unit, a spatial convolution unit and a cross-modal interaction unit is introduced, which can automatically learn and capture the complex nonlinear relationship between meteorology, terrain and snow state. Through cross-modal interaction, the deep fusion of multi-source features is achieved, and a representative fusion feature vector is generated, which not only improves the richness and accuracy of feature expression, but also enhances the comprehensive perception of avalanche risk factors. In the model training stage, the fusion feature vector is dynamically weighted based on the training set constructed based on the annotated data of historical avalanche events, so that the avalanche risk prediction model can adaptively adjust the contribution of each feature to the prediction results, thereby improving the generalization ability and prediction accuracy of the model. Ultimately, by receiving real-time environmental monitoring data and utilizing a trained avalanche risk prediction model to output the target area's risk level and warning trigger threshold, the system achieves real-time monitoring and dynamic early warning of avalanche risk. When the real-time risk value exceeds the warning trigger threshold, a multi-level warning signal is rapidly generated, providing timely and effective decision-making support to relevant departments and personnel, significantly reducing the potential losses and impacts of avalanche disasters.

[0036] In a possible implementation, step S120 includes: Step S121 , identifying missing time nodes in the meteorological time series data, performing linear interpolation compensation on the missing time nodes based on the pressure gradient change rate and temperature change trend of adjacent time windows, and generating a continuous meteorological data stream.

[0037] 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 may occasionally be missing. For example, a meteorological monitoring station located on a mountainside may have missing time nodes in its meteorological time series data between 9:00 and 11:00 a.m. on a certain day. In this case, linear interpolation compensation is performed by analyzing the pressure gradient change rate and temperature change trend of adjacent time windows to generate a continuous meteorological data stream. Assume that at 8:00 a.m., the air pressure is 1002 hPa and the temperature is -3°C. At 12:00 p.m., the air pressure is 1006 hPa and the temperature is -1°C. First, calculate the pressure gradient change rate. From 8:00 to 12:00, a total of 4 hours, the air pressure changed by 1006-1002=4 hPa, so the hourly pressure gradient change rate is 4÷4=1 hPa / hour. In terms of temperature change trend, the temperature rose by -1-(-3)=2°C in 4 hours, and the hourly temperature change rate is 2÷4=0.5°C / hour. For the missing time node of 10:00, two hours have passed since 8:00. Based on the pressure gradient change rate, the estimated pressure at 10:00 is 1002 + 1 × 2 = 1004 hPa. Based on the temperature change rate, the estimated temperature at 10:00 is -3 + 0.5 × 2 = -2°C. By interpolating and compensating for the missing time nodes in this way, a continuous meteorological data stream is obtained.

[0038] Step S122: extract key meteorological indicators from the continuous meteorological data stream, wherein the key meteorological indicators include accumulated snowfall, duration of peak wind speed, and fluctuation amplitude of diurnal temperature difference.

[0039] For example, in terms of cumulative snowfall, data from various monitoring points across the target area over the past week totaled 80 centimeters. Analysis of wind speed data from each monitoring station revealed that wind speeds exceeding 12 meters per second lasted for five hours. The amplitude of the diurnal temperature fluctuation is calculated based on a 24-hour period, with a daytime maximum temperature of -1°C and a nighttime minimum temperature of -9°C. The amplitude of the diurnal temperature fluctuation is -1-(-9)=8°C.

[0040] In step S123 , the key meteorological indicators in each preset time unit are subjected to local feature aggregation through a sliding window mechanism to generate meteorological feature segments with time correlation.

[0041] In this embodiment, the preset time unit can be set to 12 hours. Within the first 12 hours, the cumulative snowfall is 30 cm, the peak wind speed duration is 2 hours, and the day and night temperature fluctuation amplitude is 6°C. Aggregating the above indicators together can form a meteorological feature segment.

[0042] Step S124 , performing temporal dependency modeling on the meteorological feature segments through a bidirectional long short-term memory network, capturing the forward cumulative effect and the backward mutation association, and generating a meteorological feature sequence containing a multi-dimensional time context.

[0043] For example, the accumulated snowfall over the previous 12 hours will affect the temperature over the next 12 hours, a forward cumulative effect. If a sudden cold front arrives, causing a sharp increase in wind speed, this is a backward mutational association. Bidirectional long short-term memory networks can capture these relationships, generating a meteorological feature sequence with multidimensional temporal context.

[0044] Step S125 : converting the geographic coordinate system of the target area into a grid coordinate system, and generating spatial grid units according to a preset resolution.

[0045] For example, for a target area, its geographic coordinate system can be converted into a grid coordinate system, and spatial grid cells can be generated according to a preset resolution of 20 meters × 20 meters.

[0046] Step S126 , extracting the slope change rate, slope aspect dispersion and surface roughness index in each spatial grid unit as terrain attribute data.

[0047] For example, a spatial grid cell located in the middle of a hillside was measured using a topographic survey instrument to determine the slope gradient. The elevation difference from one end of the cell to the other was 4 meters, and the horizontal distance was 20 meters, resulting in a slope gradient of 4 ÷ 20 = 0.2. Regarding aspect dispersion, the slope angle was measured, revealing a range from 30 degrees northeast to 40 degrees northeast, with a defined degree of dispersion. The surface roughness index, due to the presence of rocks and vegetation within the cell, was measured and calculated to be 0.4.

[0048] Step S127: Identify the surface cover type based on satellite remote sensing image data, and map the vegetation density distribution and the proportion of exposed rock area to the corresponding grid cells.

[0049] For example, within a certain grid unit, through analysis of remote sensing images, it was found that the vegetation coverage area accounts for 40% of the total area of the unit, so the vegetation density distribution is 40%, and the rock exposed area accounts for 10%. The above data are then mapped to the corresponding grid unit.

[0050] Step S128: construct a three-dimensional terrain topology relationship map, calculate the terrain potential gradient based on the elevation difference between adjacent grid cells, and generate the potential movement path intensity distribution in combination with the avalanche movement mechanics model.

[0051] For example, assuming the elevation of one grid cell is 100 meters, the elevation of the adjacent grid cell is 105 meters, and the horizontal distance is 20 meters, the avalanche dynamics model first calculates the elevation difference to be 105-100 = 5 meters, and the terrain energy gradient to be 5÷20 = 0.25. The avalanche dynamics model is then used to generate a potential movement path intensity distribution. The potential movement path intensity distribution for this grid cell and its adjacent grid cells is then calculated based on factors such as terrain slope, aspect, and land cover type.

[0052] Step S129: Perform multi-scale feature extraction on the terrain energy gradient and potential motion path intensity distribution through a dilated convolutional neural network to generate a terrain spatial feature set containing spatial correlation features.

[0053] For example, the dilated convolutional neural network can capture terrain features at different scales. For example, at a large scale, it can capture the impact of the overall direction of the mountain range on the avalanche path, and at a small scale, it can capture the impact of small changes in local terrain within the grid cell on avalanches, thereby generating a terrain spatial feature set containing spatial correlation features.

[0054] Step S1210: Acquire snow layer profile data through a ground sensor array, wherein the snow layer profile data includes snow layer thickness distribution, crystal morphology classification, and liquid water content.

[0055] For example, at a monitoring point near the summit, snow profile data showed a snow thickness distribution of 15 cm from the surface (0-10 cm), 25 cm from a depth of 10-20 cm, and 30 cm from a depth of 20-30 cm. Microscopic observation and analysis confirmed the crystal morphology to be dendritic. Liquid water content, measured by specialized instruments, was 8%.

[0056] Step S1211, calculating the vertical density gradient according to the snow layer thickness distribution, and determining the interlayer shear strength coefficient in combination with the crystal morphology classification.

[0057] For example, assuming the surface snow density is 0.15 g / cm³ and the snow density at a depth of 30 cm is 0.3 g / cm³, the snow layer thickness from the surface to a depth of 30 cm is 60 cm. The vertical density gradient is calculated as (0.3 - 0.15) ÷ 30 = 0.005 g / cm³ / cm. 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.

[0058] Step S1212: Invert the snow surface temperature field based on the thermal infrared imaging data, and calculate the temperature gradient change rate inside the snow layer using the heat conduction equation.

[0059] For example, if a thermal infrared imager measures a snow surface temperature of -4°C, the heat conduction equation can be used to calculate the rate of change of the temperature gradient within the snowpack. Assuming the temperature is -6°C at a depth of 10 cm, the heat conduction equation calculates the rate of change of the temperature gradient as (-6 - (-4)) ÷ 10 = -0.2°C / cm.

[0060] Step S1213: The interlaminar shear strength coefficient, liquid water content, and temperature gradient change rate are integrated to construct a snow stability evaluation index.

[0061] 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 evaluation index is calculated as 0.25×0.4+0.08×0.3+(-0.2)×0.3=0.014.

[0062] Step S1214: Perform nonlinear mapping on the snow stability evaluation index through a physical constraint neural network to generate a snow state vector reflecting the internal structural state of the snow.

[0063] For example, the physical constraint neural network established a mapping relationship based on a large amount of existing snow experimental data, and input the snow stability assessment index 0.014 into the physical constraint neural network. After the neuron calculation and nonlinear function processing within the physical constraint neural network, a snow state vector reflecting the internal structural state of the snow was generated. The snow state vector contains various characteristic information of the internal structure of the snow, such as the stability of the snow, internal stress distribution and other related information.

[0064] In a possible implementation, step S130 includes: In step S131, the meteorological feature sequence is input into the temporal attention unit of the deep learning network model, and the meteorological indicators of different time nodes in the meteorological feature sequence are dynamically weighted by calculating the attention weight distribution in the time dimension to generate weighted meteorological temporal features.

[0065] In this embodiment, the meteorological feature sequence includes meteorological indicators such as snowfall, wind speed, and temperature at different time nodes. When calculating the attention weight distribution in the time dimension, weights can be assigned based on the degree of influence of different meteorological indicators on the possibility of avalanche occurrence. For example, during periods when avalanches are more likely to occur, such as periods with heavy snowfall and low temperatures, relevant meteorological indicators will be given higher attention weights. Taking a certain meteorological feature sequence as an example, at one time node, the snowfall is 15 cm / hour, the wind speed is 10 m / s, and the temperature is -5°C. 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 set to 0.3, and the weight of temperature is set to 0.3. Then, after dynamically weighting the meteorological indicators at this time node, 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, thereby generating a weighted meteorological time series feature.

[0066] Step S132: Input the terrain spatial feature set into the spatial convolution unit of the deep learning network model, use a multi-level dilated convolution kernel to extract spatial context features from the terrain spatial feature set, capture the terrain correlation between adjacent grid cells, and generate multi-scale terrain spatial features.

[0067] In this embodiment, in the terrain spatial feature set of the target area, each spatial grid unit contains information such as slope change rate, slope discreteness, surface roughness index, and surface cover type. For example, for adjacent grid cells in a certain area, the slope change rate of a grid cell is 0.2, the slope discreteness is a certain value, the surface roughness index is 0.4, and the vegetation density distribution is 40%; the corresponding values of adjacent grid cells are 0.3, different discreteness values, 0.5, and 30%, respectively. The multi-level void convolution kernel can capture the terrain correlation between these adjacent grid cells at different scales. On a large scale, you may focus on the overall terrain trend of the mountain range, and on a small scale, you may focus on the impact of local terrain changes within the grid cell on the avalanche path. In this way, multi-scale terrain spatial features are generated.

[0068] Step S133: Align and splice the snow state vector and the weighted meteorological time series features in terms of feature dimensions to generate a first fusion intermediate feature.

[0069] In this embodiment, the snow state vector includes comprehensive snow stability assessment indicators, such as interlayer shear strength coefficient, liquid water content, and temperature gradient change rate, calculated based on physical data of the snow layer. The weighted meteorological time series features include weighted meteorological indicators at different time points. The two are then aligned and concatenated. For example, the snow 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.

[0070] In step S134, 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 between the first fused intermediate feature and the multi-scale terrain spatial feature is calculated through a cross-attention mechanism to generate a cross-modal interaction feature.

[0071] In a possible implementation, step S134 includes: Step S1341 : linearly projecting the first fused intermediate features to generate a query vector sequence, and linearly projecting the multi-scale terrain spatial features to generate a key vector sequence and a value vector sequence.

[0072] Step S1342: Calculate the attention score matrix between the query vector sequence and the key vector sequence, perform weighted aggregation on the value vector sequence, and generate a spatial enhancement feature sequence.

[0073] Step S1343: Perform a residual connection on the spatial enhancement feature sequence and the first fusion intermediate feature to generate a second fusion intermediate feature.

[0074] Step S1344: input the second fused intermediate features into a multi-layer perceptron for nonlinear mapping to generate high-dimensional interactive features.

[0075] Step S1345: performing gated fusion on the high-dimensional interaction feature and the snow state vector, dynamically adjusting the weight ratio of each dimension in the high-dimensional interaction feature according to the current value of the snow state vector, and generating the cross-modal interaction feature.

[0076] In this embodiment, when calculating the attention score matrix between the query vector sequence and the key vector sequence, a certain feature component in the first fused intermediate feature and the related component in the multi-scale terrain spatial feature are taken as an example. Assume that the snowfall component in the first fused 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 the existing calculation rules. The value vector sequence is weighted aggregated. For example, for each value vector component in the multi-scale terrain spatial feature, different weights are assigned according to the calculated attention score, and then weighted summation is performed to generate a spatial enhancement feature sequence.

[0077] For example, if a feature value in the spatially enhanced feature sequence is 5, and the corresponding feature value in the first fused intermediate feature is 3, the residual connection results in 5 + 3 = 8, generating the second fused intermediate feature. This second fused intermediate feature is then input into a multilayer perceptron for nonlinear mapping. Neurons within the multilayer perceptron perform operations on the input second fused intermediate feature based on predefined weights and activation functions. For example, a neuron receives an input value of 8, multiplies it by a weight of 0.5, and outputs 4. This is then activated by an activation function (such as a ReLU function). If 4 is greater than 0, the output is 4. After multiple layers of these operations, a high-dimensional interactive feature is generated.

[0078] Assume the snow stability evaluation index in the snow state vector is 0.014. A low value indicates poor snow stability. In this case, higher weights are assigned to high-dimensional interaction features related to avalanche triggering, such as terrain slope and snowfall. For example, a dimension related to snowfall, originally weighted at 0.3, is associated with snowfall. Based on the value of the snow state vector, this weight is adjusted to 0.4, and the weights of the other dimensions are adjusted accordingly, thus generating cross-modal interaction features.

[0079] Step S135 , performing channel attention weighted screening on the cross-modal interaction features, retaining feature channels that are strongly correlated with avalanche triggering conditions, and generating optimized cross-modal interaction features.

[0080] For example, one channel relates the shear strength coefficient between snow layers to terrain slope, while another relates liquid water content to wind speed. Analysis revealed that the channel relating the shear strength coefficient between snow layers to terrain slope has a stronger influence on avalanche triggering, so it was assigned a higher weight, while the channel relating liquid water content to wind speed was assigned a lower weight. This approach preserves the characteristic channels that are strongly correlated with avalanche triggering conditions, generating optimized cross-modal interaction features.

[0081] Step S136: Input the optimized cross-modal interaction features into a fully connected layer for nonlinear transformation and feature dimensionality reduction to generate the fused feature vector.

[0082] In this embodiment, 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 its inputs according to the set weights, and then performs a nonlinear transformation (such as a sigmoid function). For example, a neuron receives a weighted sum of 5 inputs, which is calculated by the sigmoid function to obtain a value between 0 and 1. After performing such operations on multiple neurons, the nonlinear transformation of features is achieved and the feature dimension is reduced, ultimately generating a fused feature vector. This fused feature vector integrates multiple aspects of information such as weather, topography, and snow cover, and has been optimized, filtered, and transformed.

[0083] In a possible implementation, step S140 includes: 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 a predicted value of the avalanche occurrence probability, and the second fully connected layer outputs a predicted value of the avalanche trigger time.

[0084] For example, for a specific fused feature vector, which integrates information previously processed from multiple data sources, such as weather, topography, and snow cover, the first fully connected layer, after calculation and processing by its internal neurons, outputs a predicted avalanche probability. For example, if this value is 0.4, it indicates a 40% chance of an avalanche. Based on the same fused feature vector, the second fully connected layer outputs a predicted avalanche trigger time, predicting, for example, that an avalanche will trigger at 5 PM.

[0085] Step S142 , performing cross entropy loss calculation on the predicted value of the avalanche occurrence probability and the binary classification labels in the historical avalanche event annotated data to generate a classification loss value.

[0086] The historical avalanche event annotation data records in detail the occurrence of avalanches in the past, where the binary classification label indicates whether an avalanche occurred, 1 for occurrence, and 0 for non-occurrence. Suppose that in a certain historical event, an avalanche actually occurred, that is, the binary classification label is 1, and the current predicted probability of an avalanche is 0.4. The cross-entropy loss calculation process is as follows: First, calculate the logarithm of the probability of the true label. Since the true label is 1, the logarithm of the probability is ln(1)=0. Then calculate the logarithm of the predicted probability, that is, ln(0.4), which is approximately -0.916. The cross-entropy loss value is -(0×ln(0.4)+(1-1)×ln(1-0.4)), and the calculated result is approximately 0.916, which is the classification loss value.

[0087] Step S143 , performing mean square error calculation on the avalanche triggering time prediction value and the actual triggering timestamp in the historical avalanche event annotation data to generate a time regression loss value.

[0088] For example, the predicted avalanche trigger time is 5 PM, while the actual trigger time, as noted in historical events, is 4 PM. Converting these two times to minutes, 5 PM is 300 minutes, and 4 PM is 240 minutes. The mean squared error (MSE) calculation process is as follows: First, calculate the difference between the predicted and actual values: 300 - 240 = 60 minutes. Then, square this difference to obtain 60 × 60 = 3600. The mean squared error is the average of these squared values (since there is only one sample), resulting in a temporal regression loss of 3600.

[0089] Step S144: input the classification loss value and the time regression loss value into a dynamic weight allocator, and generate a classification weight coefficient and a time weight coefficient according to the proportional relationship between the classification loss value and the time regression loss value in the current training batch.

[0090] In this example, the classification loss is approximately 0.916, and the temporal regression loss is 3600. Calculating their ratio, the classification loss accounts for a very small proportion of the total loss (classification loss + temporal regression loss), which is approximately 0.000254. The temporal regression loss accounts for a very small proportion of the total loss, which is 3600 ÷ (0.916 + 3600), which is approximately 0.999746. Based on this ratio, the classification weight coefficient and the temporal weight coefficient are generated. The classification weight coefficient might be set to 0.000254, and the temporal weight coefficient to 0.999746.

[0091] Step S145 : multiplying the classification loss value by the classification weight coefficient, and multiplying the time regression loss value by the time weight coefficient to generate a weighted classification loss component and a weighted time loss component.

[0092] Step S146: Add the weighted classification loss component and the weighted time loss component to generate a joint optimization loss value.

[0093] In this example, the classification loss value of 0.916 is multiplied by the classification weight coefficient of 0.000254, resulting in a weighted classification loss component of approximately 0.000232. The temporal regression loss value of 3600 is multiplied by the temporal weight coefficient of 0.999746, resulting in a weighted temporal loss component of approximately 3599.0856. The weighted classification loss component and the weighted temporal loss component are then added together to generate the joint optimization loss value. That is, 0.000232 + 3599.0856 = 3599.085832.

[0094] Step S147: Back-propagation gradient update is performed on the parameters of the first fully connected layer, the second fully connected layer, and the deep learning network model according to the joint optimization loss value, and the mapping relationship between the fused feature vector and the avalanche risk prediction result is iteratively adjusted.

[0095] During backpropagation, starting from the output layer (the first and second fully connected layers), the gradient of each parameter is calculated based on the joint optimization loss. For example, for a neuron connection weight in the first fully connected layer, the gradient is calculated based on the joint optimization loss and the weight's influence on the output. If this weight has a significant impact on the output avalanche probability prediction, the weight will be adjusted more significantly based on the joint optimization loss during backpropagation. Similar gradient calculations and updates are performed for other parameters in the second fully connected layer and deep learning network model. Through multiple iterations, these parameters are continuously adjusted to achieve a more accurate mapping from the fused feature vector to the avalanche risk prediction result.

[0096] Step S148: When the decrease in the joint optimization loss value in a preset number of consecutive training iterations is less than a convergence threshold, the parameters of the deep learning network model are frozen to generate the avalanche risk prediction model.

[0097] Assume that the preset number of times is 10 and the convergence threshold is 0.001. During the training process, observe the joint optimization loss value after each iteration. For example, the joint optimization loss value of the first iteration is 3599.085832, and after the second iteration it becomes 3590.08, a decrease of (3599.085832-3590.08) ÷ 3599.085832, which is a large decrease. As the number of iterations increases, suppose that by the 11th iteration, it is found that the decrease in the joint optimization loss value in the last 10 iterations is less than 0.001. At this time, the parameters of the deep learning network model are frozen and no longer updated. In this way, an avalanche risk prediction model is generated. This avalanche risk prediction model has undergone dynamic weight optimization based on historical data and can more accurately predict the probability of avalanche occurrence and trigger time.

[0098] In a possible implementation, step S150 includes: Step S151 , collecting the current meteorological time series data, current terrain spatial data and current snow layer physical data of the target area in real time.

[0099] 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, 5 cm of snowfall over the past hour, and a wind speed of 8 m / s. For current terrain spatial data, regular topographic measurements and updates are performed, such as using high-precision measuring instruments to remeasure terrain attribute data within each spatial grid cell. For example, the newly measured slope change rate within a grid cell is 0.25, the slope aspect dispersion has a new value, and the surface roughness index has changed to 0.45. In addition, the surface cover type is updated based on new satellite remote sensing image data, revealing that the vegetation density distribution has changed to 35%, and the exposed rock area accounts for 15%. For the current physical data of the snow layer, the ground sensor array continuously obtains snow layer profile data. At a certain monitoring point, the snow layer thickness distribution is 12 cm for the surface 0-10 cm, 20 cm for the depth of 10-20 cm, and 25 cm for the depth of 20-30 cm. The crystal morphology is classified as hexagonal crystal system, and the liquid water content is 6%.

[0100] 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.

[0101] Collected meteorological time series data may contain data discontinuities, such as a brief interruption in data transmission at a monitoring station at a specific moment. A continuous meteorological data stream is generated by identifying missing time nodes and compensating for them through linear interpolation based on the pressure gradient change rate and temperature trend of adjacent time windows. Suppose, in the aforementioned meteorological data, data from a monitoring station is missing between 9:00 AM and 10:00 AM. At the adjacent 8:00 AM, the pressure is 1002 hPa and the temperature is -3°C. At 11:00 AM, the pressure is 1004 hPa and the temperature is -1°C. Calculating the pressure gradient change rate, the pressure changed by 1004 - 1002 = 2 hPa over three hours, resulting in an hourly pressure gradient change rate of 2 ÷ 3, approximately 0.67 hPa / hour. Regarding the temperature trend, the temperature rose by -1 - (-3) = 2°C over three hours, resulting in an hourly temperature change rate of 2 ÷ 3, approximately 0.67°C / hour. For the missing time node of 10:00, two hours have passed since 8:00. Based on the pressure gradient change rate, the estimated pressure at 10:00 is 1002 + 0.67 × 2 = 1003.34 hPa. Based on the temperature change rate, the estimated temperature at 10:00 is -3 + 0.67 × 2 = -1.66°C. After this interpolation and compensation, a continuous meteorological data stream is obtained. Key meteorological indicators are then extracted from this continuous meteorological data stream, such as the current accumulated snowfall of 5 cm (the amount of snowfall in the past hour), the duration of the peak wind speed (assuming the current wind speed of 8 m / s has not reached its peak, the duration is within the current hour), and the amplitude of the diurnal temperature fluctuation (assuming the current daytime temperature is -2°C and the nighttime temperature is -6°C, the fluctuation amplitude is -2 - (-6) = 4°C). The key meteorological indicators within each preset time unit (e.g., 12 hours) are then aggregated using a sliding window mechanism to generate temporally correlated meteorological feature segments, such as the accumulated snowfall within 12 hours, the duration of the peak wind speed, and the amplitude of the diurnal temperature fluctuation. Finally, a bidirectional long short-term memory network is used to model the temporal dependencies of meteorological feature fragments, capture the forward cumulative effect and backward mutation association, and generate a real-time meteorological feature sequence containing multidimensional temporal context.

[0102] In this embodiment, the geographic coordinate system of the target area is converted into a grid coordinate system, and spatial grid cells are generated according to a preset resolution (such as 20 meters × 20 meters). In each spatial grid cell, the slope change rate, slope aspect discreteness 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 slope aspect discreteness has a new value after remeasurement, and the surface roughness index becomes 0.45. Based on the new satellite remote sensing image data, the surface cover type is identified, and the vegetation density distribution of 35% and the rock exposed area of 15% are mapped to the corresponding grid cell. A three-dimensional terrain topological relationship diagram is constructed, and the terrain potential gradient is calculated based on the elevation difference between adjacent grid cells. For example, if the elevation of one grid cell is 100 meters, the elevation of the adjacent grid cell is 103 meters, and the horizontal distance is 20 meters, the terrain potential gradient is (103-100) ÷ 20 = 0.15. The potential motion path intensity distribution is generated by combining the avalanche motion mechanics model. The multi-scale feature extraction of the terrain energy gradient and the potential motion path intensity distribution is performed through the void convolutional neural network to generate a real-time terrain spatial feature set containing spatial correlation features.

[0103] 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 by the ground sensor array shows that at a certain monitoring point, the snow layer thickness distribution is 12 cm from the surface (0-10 cm), 20 cm from a depth of 10-20 cm, and 25 cm from a depth of 20-30 cm. The crystal morphology is hexagonal, and the liquid water content is 6%. Based on the snow layer thickness distribution, the vertical density gradient is calculated. Assuming the surface snow layer density is 0.12 g / cm³ and the snow layer density at a depth of 30 cm is 0.25 g / cm³, the snow layer thickness from the surface to a depth of 30 cm is 57 cm. The vertical density gradient is calculated as (0.25-0.12)÷30≈0.0043 g / cm³ / cm³. The interlaminar shear strength coefficient is determined by combining crystal morphology classification. For hexagonal crystals, the interlaminar shear strength coefficient is 0.22 based on existing experimental data and experience. The snow surface temperature field is inverted based on thermal infrared imaging data. Assuming a surface temperature of -3°C, the temperature gradient rate within the snow layer is calculated using the heat conduction equation. For example, if the temperature is -4°C at a depth of 10 cm, the temperature gradient rate is (-4 - (-3)) ÷ 10 = -0.1°C / cm. A snow stability index is constructed by integrating the interlayer shear strength coefficient, liquid water content, and temperature gradient rate. Assuming a weight of 0.4 for the interlayer shear strength coefficient, 0.3 for the liquid water content, and 0.3 for the temperature gradient rate, the snow stability index is calculated as 0.22 × 0.4 + 0.06 × 0.3 + (-0.1) × 0.3 = 0.064. A physical constraint neural network is used to perform nonlinear mapping of the snow stability index, generating a real-time snow state vector reflecting the internal structural state of the snowpack.

[0104] Step S153: input the real-time meteorological feature sequence, the real-time terrain spatial feature set and the real-time snow 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.

[0105] For example, after internal calculation, the avalanche risk prediction model outputs a real-time risk value of 0.55, and the warning trigger threshold is 0.5.

[0106] Step S154 , dividing the risk level intervals according to the difference between the real-time risk value and the warning trigger threshold, and generating risk level labels corresponding to the risk level intervals.

[0107] 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 classification rules, for example, a difference between 0 and 0.1 is low risk, between 0.1 and 0.3 is medium risk, and above 0.3 is high risk. Here, 0.05 is between 0 and 0.1, so a low risk level label is generated.

[0108] Step S155: 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 including different alarm frequencies and emergency response instructions.

[0109] For example, because the risk level is low, the preset warning level mapping table indicates that the corresponding alarm frequency may be once every 30 minutes. Emergency response instructions may include notifying mountain personnel to strengthen patrols in key areas and providing safety reminders to nearby tourists. If the risk level is medium, the alarm frequency may be increased to once every 15 minutes, and emergency response instructions may include restricting movement in certain dangerous areas. At a high risk level, the alarm frequency may be reduced to once every 5 minutes, and emergency response instructions may include 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.

[0110] In a possible implementation, after step S150, the method further includes: Step S310 : collecting real-time surface deformation monitoring data and weather radar reflectivity data of the target area to generate a dynamic environment verification data set.

[0111] For real-time surface deformation monitoring data, high-precision deformation monitoring instruments installed within the target area continuously record minute surface changes. For example, at key locations, such as hillsides with significant variations in height and slope, these monitoring instruments can accurately measure vertical and horizontal surface displacement. Suppose, at a given moment, a monitoring point on a mountainside records a vertical displacement of 0.5 cm and a horizontal displacement of 1 cm. Simultaneously, weather radar reflectivity data is continuously acquired. Weather radar transmits electromagnetic waves and receives reflected waves to obtain various atmospheric information. Reflectivity data reflects the distribution of different atmospheric substances and is also important for determining the state of the snowpack. By analyzing weather radar reflectivity data, reflectivity values at different altitudes can be obtained. These values are correlated with snow characteristics such as moisture content and density. By integrating surface deformation monitoring data with weather radar reflectivity data, a dynamic environmental validation dataset is generated.

[0112] In step S320, the dynamic environment verification data set is input into a pre-trained snow movement recognition model to extract the snow layer displacement rate and surface crack propagation trajectory to generate a snow structure anomaly index.

[0113] In this embodiment, the snow movement recognition model is pre-trained using a large amount of snow movement-related data. When fed with a dynamic environment validation dataset, the snow movement recognition model can extract information such as the snow layer displacement rate and surface crack propagation trajectory, thereby generating a snow structure anomaly index. Taking the snow layer displacement rate as an example, this is calculated by analyzing the displacement and time information in the surface deformation monitoring data. Assume that over a period of one hour, the vertical displacement of the aforementioned mountainside monitoring point increases from 0 cm to 0.5 cm, and the horizontal displacement increases from 0 cm to 1 cm. 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. Regarding the surface crack propagation trajectory, information such as the crack propagation direction and speed is indirectly inferred by analyzing snow surface image data (which can be obtained from specialized snow surface monitoring equipment) or by analyzing surface deformation data. This information, such as the snow layer displacement rate and surface crack propagation trajectory, is combined to generate a snow structure anomaly index according to a predefined algorithm.

[0114] Step S330: performing a spatiotemporal matching analysis based on the snow structure anomaly index and the risk level output by the avalanche risk prediction model to generate a model prediction deviation coefficient.

[0115] Suppose the avalanche risk prediction model previously outputs a medium risk level, with corresponding expected ranges for key factors such as snow thickness and terrain slope. However, the snow structure anomaly indicator indicates rapid snow displacement and extensive surface crack propagation, which differ from the snow conditions expected for a medium risk level. When calculating spatiotemporal matching, both temporal synchrony and spatial correlation must be considered. For example, if the avalanche risk prediction model predicts high snow stability in a particular area, but the snow structure anomaly indicator indicates high snow displacement in that area, there is a spatial mismatch. In terms of time, if the avalanche risk prediction model, based on historical data and current meteorological, topographic, and snow cover data, predicts a low avalanche risk for a period of time, but the snow structure anomaly indicator indicates rapidly deteriorating snow conditions, there is a temporal mismatch. These temporal and spatial mismatches are combined to generate a model prediction bias coefficient. Assume the model prediction bias coefficient is 0.3 (this value is derived from specific calculation results and is provided for illustrative purposes only).

[0116] Step S340: When the model prediction deviation coefficient exceeds a preset threshold, the terrain scanning device is activated to perform three-dimensional point cloud reconstruction on the target area, generate a surface micro-topography change map, and compare the surface micro-topography change map with the characteristic terrain template before the historical avalanche event to identify the coordinates of potential trigger points.

[0117] Assuming the preset threshold is 0.2, and because the previously calculated model prediction error coefficient of 0.3 is greater than 0.2, the terrain scanning device is activated. These terrain scanning devices, such as high-precision measurement equipment like lidar, comprehensively scan the target area, acquiring a large amount of terrain point cloud data. Then, using data processing algorithms, they generate a surface microtopography change map. This surface microtopography change map details subtle topographic changes, such as changes in hillock height and gully depth. The surface microtopography change map is then compared with a characteristic terrain template from before avalanche events. This characteristic terrain template is derived by analyzing pre-avalanche terrain data from areas where avalanches occurred. It includes terrain features such as specific slope and aspect combinations and surface roughness. By comparing the surface microtopography change map with 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 those characteristic of pre-avalanche events, such as steepening slopes and increased surface roughness, the coordinates of that area are identified as potential trigger points.

[0118] Step S350: Adjust the warning trigger threshold of the avalanche risk prediction model according to the potential trigger point coordinates, 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 transmit the updated feature weight back to the cross-modal interaction unit for feature recalibration.

[0119] Assuming the area corresponding to the identified potential trigger point coordinates is a region with a high probability of avalanche occurrence, the warning trigger threshold needs to be lowered to improve warning accuracy. For example, if the original warning trigger threshold is 0.5, based on the risk associated with the potential trigger point coordinates, the warning trigger threshold is adjusted to 0.45, and an adaptive threshold update instruction is generated. Based on this adaptive threshold update instruction, the feature weight allocation strategy for the current environmental monitoring data is modified. For example, the weights of monitoring data closely related to avalanche occurrence, such as liquid water content in the physical data of the snow layer and slope in the spatial terrain data, are increased near the potential trigger point coordinates. The updated feature weights are then fed back to the cross-modal interaction unit for feature recalibration. The cross-modal interaction unit has already integrated multiple data sources, including meteorological feature sequences, spatial terrain feature sets, and snow state vectors. By recalibrating the feature weights, the cross-modal interaction unit can better integrate these data.

[0120] Step S360: Use the recalibrated cross-modal interaction unit to output a new fusion feature vector to drive the avalanche risk prediction model to generate incremental learning parameters, and perform gradient interpolation fusion on the incremental learning parameters and the original model parameters to generate an environment-adaptive avalanche risk prediction model.

[0121] In this embodiment, the recalibrated cross-modal interaction unit processes the input data according to the new feature weights and outputs a new fused feature vector, which contains the weighted multi-source data. This fused feature vector is then fed into the avalanche risk prediction model, which then generates incremental learning parameters based on the new input data. These incremental learning parameters reflect the direction and magnitude of model adjustments required due to environmental changes, such as the discovery of potential trigger points.

[0122] The incremental learning parameters are then 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 of the incremental learning parameters is 0.1, the two are fused using the gradient interpolation fusion algorithm to obtain a new weight value, such as 0.35. This fusion operation is performed on all parameters in the avalanche risk prediction model to generate an environmentally adaptive avalanche risk prediction model. This environmentally adaptive avalanche risk prediction model can better adapt to actual environmental changes in the target area and improve the accuracy of avalanche risk prediction.

[0123] In step S370, the parameters of the environment-adaptive avalanche risk prediction model are synchronized to all online warning terminals through the edge computing node to complete the closed-loop iteration of the warning logic.

[0124] In this embodiment, edge computing nodes are located close to the target area and have the ability to quickly process data. They can send new model parameters to various online warning terminals distributed around the target area, such as at the entrance to a tourist attraction or near residential areas in mountainous areas. Once these warning terminals receive the new model parameters, they can use the updated, environmentally adaptive avalanche risk prediction model to provide more accurate avalanche risk predictions and warnings, thus completing the closed-loop iteration process of the entire warning logic and continuously improving the accuracy and effectiveness of avalanche risk predictions and warnings.

[0125] For example, in one possible implementation, the method further includes: After generating the snow structure anomaly index, perform the following steps: Step S410: construct a three-dimensional motion vector field based on the displacement rate of the snow layer, extract anisotropic displacement distribution characteristics, and couple the anisotropic displacement distribution characteristics with the terrain potential gradient in the terrain spatial feature set to generate a potential energy-motion correlation matrix.

[0126] In this embodiment, since the snow layer displacement rate has already been obtained when analyzing snow structure anomaly indicators, for example, the vertical and horizontal displacement rates of the snow layer at different monitoring points. For a monitoring point with a vertical displacement rate of 0.5 cm / hour and a horizontal displacement rate of 1 cm / hour, a three-dimensional motion vector field is constructed with this monitoring point as the origin, the vertical direction as the z-axis, and the horizontal directions as the x-axis and y-axis. In this environmentally adaptive avalanche risk prediction vector field, each point has a vector representing the direction and velocity of the snow layer's movement. Anisotropic displacement distribution characteristics are then extracted. Anisotropy refers to different displacement characteristics in different directions. For example, in a certain hillside area, due to factors such as terrain and wind direction, the displacement rate of snow in the direction down the slope (x-axis) may be greater than that in the direction perpendicular to the slope (y-axis). This difference in displacement rate in different directions is the anisotropic displacement distribution characteristic.

[0127] The terrain potential energy gradient, a collection of spatial features, reflects the impact of terrain on the potential motion of snowpack. For example, within a grid cell, the terrain potential energy gradient is 0.15, and the anisotropic displacement distribution in that area shows a large displacement trend of snowpack in a certain direction. Considering how the terrain potential energy gradient affects snowpack motion under anisotropic conditions, a potential energy-motion correlation matrix is generated. Each element in this potential energy-motion correlation matrix represents the correlation between potential energy and motion under specific terrain and snowpack displacement characteristics.

[0128] Step S420, based on the potential energy-motion correlation matrix, calculates the stress concentration index of the snow body in different altitude zones, identifies the boundary of the stress-exceeding area, and activates the drone cluster to perform infrared thermal imaging scanning on the specified altitude zone according to the boundary of the stress-exceeding area to obtain the temperature field distribution data inside the snow.

[0129] In this embodiment, the target area can be divided into different altitude zones based on altitude, for example, with each zone consisting of 100 meters. 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 at a certain altitude zone, based on the correlation relationship in the potential energy-motion correlation matrix and relevant information such as a snow density of 0.2 grams per cubic centimeter, a series of calculations yield a snow stress concentration index of 0.4. The boundaries of the stress-exceeding region are identified, namely, the boundaries of the region where the snow stress concentration index exceeds a preset critical value are found. Assuming the preset critical value is 0.3, the region with a snow stress concentration index of 0.4 is identified as a stress-exceeding region, and the boundary of this stress-exceeding region is determined.

[0130] After receiving the command, the drone swarm flies to the altitude zone corresponding to the stress-exceeding zone, for example, 1500-1600 meters above sea level. The drones' onboard infrared thermal imaging equipment scans the snow, acquiring data on the internal temperature distribution. Due to snow's thermal conductivity and other properties, temperatures vary at different depths. Infrared thermal imaging scans can reveal the temperature distribution from the snow surface to a specific depth. For example, consider a region with a surface temperature of -3°C and a temperature of -5°C at a depth of 10 cm.

[0131] In step S430, the temperature field distribution data inside the snow and the temperature gradient change rate in the snow state vector are subjected to data assimilation processing to generate a revised snow stability evaluation index. The revised snow stability evaluation index is input into a physical constraint neural network for forward reasoning to update the numerical distribution of the snow state vector.

[0132] In this example, the temperature gradient rate of change in the snow state vector was previously calculated based on a pre-defined set of parameters, such as a temperature gradient of -0.2°C / cm from the snow surface to a certain depth. Now, with more detailed data on the temperature distribution within the snowpack, this data is integrated with the original temperature gradient rate. Assume that a weighted average method is used to assign different weights to data at different depths based on their importance, to calculate a new temperature-related index. This is then combined with other factors in the snow state vector (such as the interlayer shear strength coefficient and liquid water content) 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.

[0133] The physical constraint neural network was previously trained using a large amount of snow data. It can calculate the corrected snow stability index input using its internal neuron connection weights and activation functions. For example, a neuron in the physical constraint neural network receives a corrected snow stability index of 0.055, multiplies it by the connection weight of 0.5, and obtains 0.0275. This value is then passed through an activation function (such as a ReLU function). If 0.0275 is greater than 0, the value is output as 0.0275. After multiple layers of neuron calculations, a new value is obtained, thereby updating the numerical distribution of the snow state vector.

[0134] In step S440, the updated snow state vector is re-input into the cross-modal interaction unit, triggering real-time recalculation of the fused feature vector, and adjusting the fully connected layer weights of the avalanche risk prediction model according to the recalculated fused feature vector to generate a dynamically optimized risk probability output.

[0135] The cross-modal interaction unit has previously fused multiple data sources, including meteorological feature sequences, terrain spatial feature sets, and the snow state vector. Now, the updated snow state vector is input and feature fusion and interaction calculations are re-performed. For example, previously, the cross-modal interaction unit performed operations such as cross-attention calculations on the snow state vector and other features. These calculations are now re-performed based on the updated snow state vector, triggering real-time recalculation of the fused feature vector.

[0136] The weights of the fully connected layer of the avalanche risk prediction model are fixed or based on previously calculated results during training and use. Based on the recalculated fused feature vector, for example, if the importance of certain features changes, the neuron connection weights in the fully connected layer are adjusted accordingly. For example, 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 calculations in the fully connected layer, a dynamically optimized risk probability output is generated. For example, the original risk probability output was 0.55, but after adjustment, it becomes 0.6.

[0137] Step S450, the risk probability output after dynamic optimization is combined with the original warning signal for decision fusion to generate a confidence-weighted final warning, etc., and the multi-level signal release strategy of the warning terminal is controlled according to the final warning level, and the data stream generated by the optimization process is synchronously written into the training set for storage.

[0138] In this embodiment, the original warning signal is previously generated based on the results of the avalanche risk prediction model output, such as the warning signal corresponding to the previous risk probability output of 0.55. The dynamically optimized risk probability output of 0.6 is now fused with the original warning signal. Assuming that a weighted fusion method is used to assign different weights according to the confidence of the two, for example, the confidence of the dynamically optimized risk probability output is 0.6, and the confidence of the original warning signal is 0.4, the final warning level is calculated. 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 more stringent measures, such as speeding up the organization of the evacuation of surrounding residents.

[0139] Throughout the optimization process, from analyzing snow structure anomalies to adjusting the final warning level, a series of data is generated, including the distribution of the internal temperature field of the snowpack, the updated snow state vector, and the recalculated fused feature vector. This data is then written into the training set according to the specified format and rules for use in subsequent model training and optimization, further improving the accuracy and adaptability of the avalanche risk prediction model.

[0140] Figure 2 A schematic diagram illustrates exemplary hardware and software components of a deep learning-based avalanche warning model construction system 100 that can implement the concepts of the present application, as provided in some embodiments of the present application. For example, the processor 120 can be used in the deep learning-based avalanche warning model construction system 100 and used to perform the functions of the present application.

[0141] The deep learning-based avalanche warning model construction system 100 can be a general-purpose server or a special-purpose server, both of which can be used to implement the deep learning-based avalanche warning model construction method of the present 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 processing load.

[0142] For example, the avalanche warning model construction system 100 based on deep learning may include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and different forms of storage media 140, such as a disk, ROM, or RAM, or any combination thereof. Exemplarily, the avalanche warning model construction system 100 based on deep learning may also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The method of the present application can be implemented according to these program instructions. The avalanche warning model construction system 100 based on deep learning also includes an input / output (I / O) interface 150 between the computer and other input and output devices.

[0143] For ease of explanation, only one processor is described in the deep learning-based avalanche warning model construction system 100. However, it should be noted that the deep learning-based avalanche warning model construction system 100 in the present application may also include multiple processors, so the steps performed by one processor described in the present application may also be performed jointly or individually by multiple processors. For example, if the processor of the deep learning-based avalanche warning model construction system 100 executes step A and step B, it should be understood that step A and step B may also be performed jointly by two different processors or individually in one processor. For example, the first processor executes step A, the second processor executes step B, or the first processor and the second processor execute steps A and B together.

[0144] In addition, an embodiment of the present invention also provides a readable storage medium, in which computer-executable instructions are preset. When a processor executes the computer-executable instructions, the above-mentioned method for constructing an avalanche warning model based on deep learning is implemented.

[0145] It should be noted that in order to simplify the description of the present invention and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, multiple features are sometimes combined into one embodiment, figure or description thereof.

Claims

1. A method for constructing an avalanche warning model based on deep learning, characterized in that: The method comprises: Acquiring multi-source environmental monitoring data of a target area, wherein the multi-source environmental monitoring data includes meteorological time series data, terrain spatial data, and snow layer physical data; Performing time dimension alignment processing on the meteorological time series data to generate a meteorological feature sequence, performing spatial gridding processing on the terrain spatial data to generate a terrain spatial feature set, and extracting physical parameters from the snow layer physical data to generate a snow state vector; Inputting the meteorological feature sequence, the terrain spatial feature set, and the snow state vector into a deep learning network model to generate a fused feature vector, wherein the deep learning network model includes a temporal attention unit, a spatial convolution unit, and a cross-modal interaction unit; Building a training set based on historical avalanche event annotated data, and using the training set to dynamically optimize the weight of the fused feature vector to generate an avalanche risk prediction model; The current environmental monitoring data is received in real time, and the risk level and warning trigger threshold of the target area are output through the avalanche risk prediction model. When the real-time risk value exceeds the warning trigger threshold, a multi-level warning signal is generated.

2. The method for constructing an avalanche warning model based on deep learning according to claim 1, characterized in that: The step of performing time dimension alignment processing on the meteorological time series data to generate a meteorological feature sequence includes: Identifying missing time nodes in the meteorological time series data, and performing 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; Extracting key meteorological indicators from the continuous meteorological data stream, wherein the key meteorological indicators include accumulated snowfall, duration of peak wind speed, and fluctuation amplitude of diurnal temperature difference; The key meteorological indicators in each preset time unit are aggregated locally through a sliding window mechanism to generate meteorological feature fragments with time correlation; The temporal dependency modeling of the meteorological feature fragments is performed through a bidirectional long short-term memory network, the forward cumulative effect and the backward mutation association are captured, and a meteorological feature sequence containing a multidimensional time context is generated.

3. The method for constructing an avalanche warning model based on deep learning according to claim 2, characterized in that: The 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 into a grid coordinate system and generate spatial grid cells according to the preset resolution; The slope change rate, aspect dispersion and surface roughness index are extracted as terrain attribute data in each spatial grid cell; Identify land cover types based on satellite remote sensing image data and map vegetation density distribution and rock exposed area ratio to corresponding grid cells; Construct a three-dimensional terrain topology map, calculate the terrain potential gradient based on the elevation difference between adjacent grid cells, and generate the potential movement path intensity distribution based on the avalanche movement mechanics model; Multi-scale feature extraction is performed on the terrain energy gradient and potential motion path intensity distribution through a void convolutional neural network to generate a terrain spatial feature set containing spatial correlation features.

4. The method for constructing an avalanche 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: Acquiring snow layer profile data through a ground sensor array, wherein the snow layer profile data includes snow layer thickness distribution, crystal morphology classification, and liquid water content; Calculating the vertical density gradient based on the snow layer thickness distribution and determining the interlayer shear strength coefficient in combination with the crystal morphology classification; The snow surface temperature field is inverted based on thermal infrared imaging data, and the temperature gradient change rate inside the snow layer is calculated using the heat conduction equation. The interlayer shear strength coefficient, liquid water content and temperature gradient change rate are integrated to construct a snow stability evaluation index; The snow stability evaluation index is nonlinearly mapped through a physical constraint neural network to generate a snow state vector reflecting the internal structural state of the snow.

5. The method for constructing an avalanche warning model based on deep learning according to claim 1, characterized in that: The step of inputting the meteorological feature sequence, the terrain spatial feature set, and the snow state vector into a deep learning network model to generate a fusion feature vector includes: Inputting the meteorological feature sequence into the time series attention unit of the deep learning network model, dynamically weighting the meteorological indicators at different time nodes in the meteorological feature sequence by calculating the attention weight distribution in the time dimension, and generating weighted meteorological time series features; Inputting the terrain spatial feature set into the spatial convolution unit of the deep learning network model, performing spatial context feature extraction on the terrain spatial feature set using a multi-level dilated convolution kernel, capturing the terrain correlation between adjacent grid cells, and generating multi-scale terrain spatial features; Aligning and splicing the snow state vector and the weighted meteorological time series features in their feature dimensions to generate a first fusion intermediate feature; Inputting the first fused intermediate feature and the multi-scale terrain spatial feature into the cross-modal interaction unit, calculating the cross-modal feature correlation between the first fused intermediate feature and the multi-scale terrain spatial feature through a cross-attention mechanism, and generating a cross-modal interaction feature; Performing channel attention weighted screening on the cross-modal interaction features, retaining feature channels that are strongly correlated with avalanche triggering conditions, and 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 warning model based on deep learning according to claim 5, characterized in that: The step of calculating the cross-modal feature correlation between the first fused intermediate feature and the multi-scale terrain spatial feature through a cross-attention mechanism to generate a cross-modal interaction feature includes: Generate a query vector sequence by linearly projecting the first fused intermediate features, and generate a key vector sequence and a value vector sequence by linearly projecting the multi-scale terrain spatial features; Calculating an attention score matrix between the query vector sequence and the key vector sequence, performing weighted aggregation on the value vector sequence, and generating a spatially enhanced feature sequence; Performing a residual connection between the spatial enhancement feature sequence and the first fused intermediate feature to generate a second fused intermediate feature; Inputting the second fused intermediate features into a multi-layer perceptron for nonlinear mapping to generate high-dimensional interactive features; The high-dimensional interaction feature is gated and fused with the snow state vector, and the weight ratio of each dimension in the high-dimensional interaction feature is dynamically adjusted according to the current value of the snow state vector to generate the cross-modal interaction feature.

7. The method for constructing an avalanche warning model based on deep learning according to claim 1, characterized in that: The step of dynamically weighting the fused feature vector using the training set to generate an avalanche risk prediction model includes: Inputting 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 a predicted value of avalanche occurrence probability, and the second fully connected layer outputs a predicted value of avalanche trigger time; Performing cross entropy loss calculation on the predicted value of avalanche occurrence probability and the binary classification labels in the historical avalanche event annotated data to generate a classification loss value; Performing mean square error calculation on the predicted avalanche trigger time value and the actual trigger timestamp in the historical avalanche event annotated data to generate a time regression loss value; Input the classification loss value and the time regression loss value into a dynamic weight allocator, and generate a classification weight coefficient and a time weight coefficient according to a proportional relationship between the classification loss value and the time regression loss value in a current training batch; Multiplying the classification loss value by the classification weight coefficient, and multiplying the time regression loss value by the time weight coefficient to generate a weighted classification loss component and a weighted time loss component; Adding the weighted classification loss component and the weighted time loss component to generate a joint optimization loss value; Performing back-propagation gradient updates on the parameters of the first fully connected layer, the second fully connected layer, and the deep learning network model according to the joint optimization loss value, and iteratively adjusting the mapping relationship between the fused feature vector and the avalanche risk prediction result; When the decrease in the joint optimization loss value in a preset number of consecutive training iterations is less than a convergence threshold, the parameters of the deep learning network model are frozen to generate the avalanche risk prediction model.

8. The method for constructing an avalanche warning model based on deep learning according to claim 1, characterized in that: The method includes receiving current environmental monitoring data in real time, outputting the risk level and warning trigger threshold of the target area through the avalanche risk prediction model, and generating a multi-level warning signal when the real-time risk value exceeds the warning trigger threshold, including: Real-time collection of current meteorological time series data, current terrain spatial data and current snow layer physical data of the target area; Performing time dimension alignment processing on the current meteorological time series data to generate a real-time meteorological feature sequence, performing spatial gridding processing on the current terrain spatial data to generate a real-time terrain spatial feature set, and extracting physical parameters from the current snow layer physical data to generate a real-time snow state vector; Inputting the real-time meteorological feature sequence, the real-time terrain spatial feature set and the real-time snow state vector into the avalanche risk prediction model, outputting the real-time risk value of the target area and the dynamically updated warning trigger threshold; Divide the risk level intervals according to the difference between the real-time risk value and the warning trigger threshold, and generate risk level labels corresponding to the risk level intervals; 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.

9. The method for constructing an avalanche warning model based on deep learning according to claim 1, characterized in that: After generating a multi-level warning signal when the real-time risk value exceeds the warning triggering threshold, the method further includes: Collecting real-time surface deformation monitoring data and meteorological radar reflectivity data of the target area to generate a dynamic environment verification data set; Input the dynamic environment verification dataset into the pre-trained snow movement recognition model to extract the snow layer displacement rate and surface crack propagation trajectory to generate snow structure anomaly indicators; Performing a spatiotemporal matching analysis based on the snow structure anomaly index and the risk level output by the avalanche risk prediction model to generate a model prediction deviation coefficient; When the model prediction deviation coefficient exceeds a preset threshold, the terrain scanning device is activated 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 characteristic terrain template before the historical avalanche event to identify the coordinates of potential trigger points; Adjusting the warning trigger threshold of the avalanche risk prediction model according to the potential trigger point coordinates, generating an adaptive threshold update instruction, correcting the feature weight allocation strategy of the current environmental monitoring data based on the adaptive threshold update instruction, and transmitting the updated feature weight back to the cross-modal interaction unit for feature recalibration; The recalibrated cross-modal interaction unit is used to output a new fusion feature vector to drive the avalanche risk prediction model to generate incremental learning parameters, and the incremental learning parameters are gradient interpolated and fused with the original model parameters to generate an environment-adaptive avalanche risk prediction model; The parameters of the environment-adaptive avalanche risk prediction model are synchronized to all online warning terminals through edge computing nodes to complete the closed-loop iteration of the warning logic.

10. A deep learning-based avalanche warning model construction system, characterized in that: The deep learning-based avalanche warning model construction system includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the deep learning-based avalanche warning model construction method described in any one of claims 1 to 9 above.

Images