Ice lake outburst prediction method based on deep learning
By fusing multi-source data and using deep learning models, glacial lake features are extracted and a stability model is constructed, which solves the problems of data generalization and response lag in traditional glacial lake outburst prediction, and achieves high-precision outburst risk prediction and rapid emergency response.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-17
- Publication Date
- 2026-03-13
AI Technical Summary
Traditional glacial lake outburst flood prediction methods often rely on a single data source, have poor generalization ability, cannot adapt to different climate scenarios, lack risk classification and accurate early warning, and have delayed early warning response, making it difficult to meet the timeliness requirements of emergency response.
Multi-source remote sensing data and ground monitoring data were fused and preprocessed. Deep convolutional neural networks were used to extract the characteristics of glacial lake boundaries, ice thickness and upstream confluence. Combined with time-series hydrological data, a dynamic change model of glacial lake stability was constructed. The outflow risk level was classified and a risk map and emergency plan were generated and pushed to relevant terminals.
It significantly improves the accuracy of glacial lake outburst flood prediction, supports rapid response, enhances disaster prevention timeliness, and enables efficient utilization of multi-source features and risk-level early warning.
Smart Images
Figure CN121660187A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of glacial lake outburst prediction technology, and more particularly to a deep learning-based method for predicting glacial lake outbursts. Background Technology
[0002] Glacial lakes generally refer to lakes within glacial zones that are directly or indirectly connected to glaciers, commonly found in high-altitude mountainous areas. A glacial lake outburst flood occurs when glacial moraine dams become unstable due to meltwater pressure or impacts from ice avalanches or snow avalanches, causing a large volume of stored water to be released in a short period, resulting in a devastating flood and debris flow. This type of disaster is characterized by its suddenness and immense destructive power, seriously threatening the safety of downstream residents and infrastructure. To prevent such disasters, glacial lake outburst flood prediction utilizes remote sensing monitoring, on-site surveys, and mathematical models to comprehensively analyze key indicators such as lake area, water volume, dam structure, and glacier dynamics to assess the risk of outburst flooding, thereby providing a scientific basis for disaster early warning and risk prevention.
[0003] Deep reinforcement learning is a revolutionary interdisciplinary field in artificial intelligence that combines the perceptual capabilities of deep learning with the decision-making abilities of reinforcement learning. Its core lies in the agent learning optimal action strategies from reward signals obtained through continuous trial-and-error interaction with the environment. Deep learning neural networks act as the "brain," responsible for processing high-dimensional, complex raw input data (such as images and sensor data) and evaluating the value of actions or directly generating strategies based on this data.
[0004] Glacier outburst floods are a typical manifestation of glacier disasters against the backdrop of global warming. As glaciers retreat at an accelerated pace, the number and size of glacier lakes continue to expand, leading to frequent floods and debris flows that severely threaten downstream ecosystems and human safety. Based on existing technologies, traditional glacier outburst flood prediction methods often rely on single data sources, making it difficult to comprehensively characterize the state of glaciers. Furthermore, traditional empirical models or purely data-driven models lack physical constraints, exhibiting poor generalization ability under extreme weather conditions such as heavy precipitation and surges in glacial meltwater, resulting in insufficient prediction accuracy. Additionally, some prediction methods neglect dynamic fusion of multiple features, failing to adapt to the weight differences in different scenarios such as rainy / dry seasons. They also lack risk classification and precise delivery, leading to delayed early warning responses and failing to meet the timeliness requirements of emergency response. Therefore, this invention proposes a deep learning-based glacier outburst flood prediction method to address the problems existing in current technologies. Summary of the Invention
[0005] To address the aforementioned problems, the present invention aims to propose a deep learning-based method for predicting glacial lake outburst floods. This method solves the problems of traditional glacial lake outburst flood prediction methods, which often rely on a single data source, have poor generalization ability under extreme weather conditions such as heavy precipitation and glacial melt surges, and cannot adapt to the weight differences in different scenarios such as rainy / dry seasons. Furthermore, these methods lack risk classification and accurate push notifications, resulting in delayed early warning responses.
[0006] To achieve the objectives of this invention, the invention is implemented through the following technical solution: a deep learning-based method for predicting glacial lake outburst floods, comprising the following steps:
[0007] Step 1: Acquire multi-source remote sensing data and ground monitoring data related to the glacial lake, and perform standardized preprocessing to obtain multi-dimensional feature data that can be used for multi-dimensional feature extraction;
[0008] Step 2: Based on the preprocessed multi-dimensional feature data, use a deep convolutional neural network to extract the boundary features, ice thickness features, and upstream confluence features of the glacial lake;
[0009] Step 3: Combining time-series hydrological data, a dynamic change model of glacial lake stability is constructed using extracted boundary features, ice thickness features, and upstream confluence features to predict the probability of glacial lake outburst within a preset timeframe.
[0010] Step 4: Based on the predicted probability of glacial lake outburst, classify the risk level of glacial lake outburst, generate a risk map and targeted emergency response plan, and push them to relevant user terminals.
[0011] Further improvements are made in the following aspects: In step one, the multi-source remote sensing data includes lidar data, satellite optical images and synthetic aperture radar data, the ground monitoring data includes real-time meteorological data and hydrological observation data, and the standardization preprocessing includes spatial registration, temporal alignment, noise removal and missing data imputation.
[0012] A further improvement is made in step two, where the deep convolutional neural network employs a multi-branch parallel architecture. Each branch receives preprocessed multi-source data as input, outputs corresponding features, and passes them to the main feature fusion module. Specifically, this includes:
[0013] Spatial feature extraction branch: U-Net network with encoder-decoder structure;
[0014] Time series analysis branch: Using a bidirectional long short-term memory network, historical meteorological data and hydrological data are input, long-term dependencies of the data are captured, and a prediction curve of future glacial lake water volume changes is output.
[0015] Upstream confluence characteristic branch: Based on the elevation zoning model, the contributions of glacial meltwater, snowmelt and precipitation are calculated. After normalizing each contribution, it is input into the fully connected layer to generate a feature vector reflecting the intensity of upstream water inflow.
[0016] Feature fusion module: It receives spatial features, temporal features and upstream confluence feature vectors through a gated loop unit, dynamically adjusts the weights of each feature, and outputs a fused multi-dimensional feature vector.
[0017] Further improvements are made in the following aspects: In step two, the boundary features are accurately identified by fusing spectral features and multi-scale spatial features; the ice thickness features are obtained by analyzing SAR data, radar echo signals, and water color changes in satellite optical images; and the upstream confluence features are generated by calculating the contributions of glacial meltwater, snowmelt, and precipitation.
[0018] Further improvements are made in the following: During the training of the deep convolutional neural network, a dataset is constructed using publicly available remote sensing image datasets and ice lake-specific labeled data. Then, a hybrid loss function is used to support the model optimization, as shown in the formula:
[0019]
[0020] in, Cross-entropy loss for boundary classification, The mean square error loss in water quantity prediction The correlation loss between upstream confluence features and stability coefficient is used. The training strategy is to initialize the VGG16 encoder with transfer learning, perform Adam optimization with an initial learning rate of 0.001, and decrease the learning rate by 0.1 every 10 rounds.
[0021] A further improvement is made in step three, where the construction process of the dynamic change model for glacial lake stability is as follows: based on the fused upstream confluence characteristics, the real-time water storage of the glacial lake is calculated; combined with the rate of water level change, the expansion trend of the glacial lake is analyzed; then, the rate of change of ice thickness is retrieved through SAR data; combined with the height and slope of the moraine dam monitored by lidar, the anti-sliding stability of the dam is evaluated; then, the rate of change of water volume, the rate of change of ice thickness, and the dam stability coefficient are used as inputs to input a deep learning model with physical constraints, and the probability of glacial lake collapse within the next 72 hours is output.
[0022] A further improvement is made in step four, when classifying risk levels, a default probability threshold is preset. If the probability is within 10% of the threshold, it is classified as low risk; if it is 10-30% higher than the threshold, it is classified as medium risk; and if it is 30% higher than the threshold, it is classified as high risk.
[0023] Further improvements are made in step four, when generating the risk map, based on the DEM and the flood path simulation, the potential flood areas and the scope of impact are marked, and reservoir capacity estimation and flood peak flow prediction are superimposed. Then, the risk map, emergency plan and real-time monitoring data are pushed to the emergency management department terminal and the residents' APP through the cloud platform.
[0024] The beneficial effects of this invention are as follows: This invention breaks through the limitations of traditional single data sources by fusing and preprocessing multi-source remote sensing (LiDAR, optical imagery, synthetic aperture radar) and ground monitoring data, solving the problems of inconsistent spatial location and asynchronous time of data, and providing high-quality input for multi-source feature extraction. Furthermore, by designing a three-branch deep network for spatial feature extraction, temporal analysis, and upstream confluence features, and combining a gating mechanism to dynamically adjust the weights of each branch, the utilization rate of glacial lake boundary, water volume, and confluence features is significantly improved. At the same time, a physical-data hybrid model is used to integrate water volume-water level coupling and dam stability assessment, taking into account both model interpretability and data-driven accuracy. In addition, a risk decision-making closed loop is used to realize the prediction of failure probability, graded early warning, and multi-terminal solution push, supporting rapid linkage from prediction to response. Compared with traditional methods, the prediction accuracy is higher, significantly enhancing the timeliness of disaster prevention and response. Attached Figure Description
[0025] Figure 1 This is a flowchart illustrating the deep learning-based method for predicting glacial lake outbursts according to the present invention. Detailed Implementation
[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0027] It should be noted that the technical means not described in detail in the following embodiments are all conventional means in the field, are not the key points of the invention, and will not be elaborated upon.
[0028] See Figure 1 This embodiment provides a deep learning-based method for predicting glacial lake outburst crises, including the following steps:
[0029] Step 1: Multi-source data acquisition and preprocessing
[0030] Multi-source remote sensing data and ground monitoring data related to the glacial lake were acquired and standardized preprocessed to obtain multi-dimensional feature data that can be used for multi-dimensional feature extraction. The multi-source remote sensing data includes LiDAR data, satellite optical imagery and synthetic aperture radar (SAR) data, and the ground monitoring data includes real-time meteorological data (temperature, precipitation, wind speed) and hydrological observation data (water level, flow rate).
[0031] The standardization preprocessing of multi-dimensional feature data in this embodiment includes spatial registration, temporal alignment, noise removal, and missing data imputation, specifically:
[0032] LiDAR data processing: A high-precision digital elevation model (DEM) is generated through point cloud filtering and interpolation. This model is used to analyze the topographic slope, valley morphology, and spatial distribution of glacial moraine in the glacial lake basin, providing basic topographic data for subsequent analysis of glacial moraine morphology and basin elevation zonation.
[0033] Satellite optical image processing: Based on multispectral bands, water color indices (such as NDWI) are extracted. Combined with the dual polarization (HH / HV) echo characteristics of SAR data, ice structure and thickness changes are analyzed through ice-water classification algorithms. A radar incident angle correction model is introduced to eliminate the interference of terrain on SAR signals. By fusing NDWI and SAR dual polarization (radar features), the accuracy of ice thickness inversion is improved. Incident angle correction ensures the reliability of SAR data under different terrain conditions.
[0034] Ground monitoring data processing: Real-time collection of temperature, precipitation, wind speed, and water level data through IoT sensors; removal of high-frequency noise through Kalman filtering; storage in time-series database according to time windows; elimination of high-frequency noise (such as sensor jitter) in real-time monitoring data; and storage by hour to adapt to the input requirements of time-series models.
[0035] Multi-source data fusion: Based on the DEM geographic coordinates, spatial registration and temporal alignment are performed on remote sensing images and ground monitoring data. Wavelet transform is used to remove random noise from remote sensing images, and linear interpolation is used to fill in missing water level or meteorological data. The DEM is used as the geographic reference to unify spatial coordinates (avoiding "same object in different locations") and unify timestamps (avoiding "synchronous data from different times"). Wavelet transform + linear interpolation ensures data integrity and provides high-quality input for feature extraction in step two.
[0036] Step 2: Multi-dimensional Feature Extraction
[0037] Based on the multi-dimensional feature data preprocessed in step one, a deep convolutional neural network (CNN) is used to extract the boundary features, ice thickness features, and upstream confluence features of the glacial lake. The boundary features are accurately identified by fusing spectral features and multi-scale spatial features. The ice thickness features are obtained by analyzing the changes in water color in SAR data radar echo signals and satellite optical images. The upstream confluence features are generated by calculating the contributions of glacial meltwater, snowmelt, and precipitation.
[0038] The deep convolutional neural network in this embodiment adopts a multi-branch parallel architecture. Each branch takes preprocessed multi-source data as input, outputs corresponding features, and passes them to the main feature fusion module. Specifically, it includes:
[0039] Spatial feature extraction branch: The U-Net network with encoder-decoder structure is adopted. The encoder part uses the ImageNet pre-trained VGG16 model to extract shallow, medium and deep convolutional features. The decoder upsamples through transposed convolution, fuses multi-scale feature maps, and outputs pixel-level classification probability maps of ice lake boundaries after Softmax activation. By utilizing the shallow edge features, medium texture features and deep semantic features of the pre-trained VGG16, combined with the multi-scale fusion of the U-Net decoder, pixel-level recognition of ice lake boundaries is achieved.
[0040] Temporal analysis branch: A bidirectional long short-term memory network (BiLSTM) is used to input meteorological and hydrological data from the past 7 days, capture the long-term dependencies of the data, and output the predicted curve of glacial lake water volume change for the next 72 hours. The bidirectional structure captures the "past-present" dependency of meteorological / hydrological data (such as the lagged effect of previous precipitation on the current water level), outputs the water volume change trend, and provides temporal context for upstream confluence characteristics.
[0041] Upstream confluence characteristic branch: Based on the elevation zoning model, the contributions of glacial meltwater, snowmelt, and precipitation are calculated. These contributions are normalized and input into the fully connected layer to generate a feature vector reflecting the intensity of upstream water inflow. This quantifies the complex meltwater / precipitation contributions into a normalized feature vector, avoiding the dimensionality curse of the original data. The core calculation of the elevation zoning model uses a degree-day factor model to quantify the meltwater effect: the contribution of glacial meltwater is expressed by the formula... calculate( The contribution of glacial meltwater, For temperature, As the glacier diurnal factor, To correspond to the area of the elevation zone, use the subscript " " is an abbreviation for "Glacier". The contribution of snowmelt is calculated using a modified diurnal factor model and incorporates a snow depth correction coefficient.
[0042] Feature fusion module: It receives spatial features, temporal features and upstream confluence feature vectors through gated recurrent units (GRU), dynamically adjusts the weights of each feature, and outputs a fused multi-dimensional feature vector as input to the stability model in step three. It dynamically adjusts the weights of each branch through a gating mechanism (e.g., increasing the weight of temporal features during the rainy season and increasing the weight of confluence features during the dry season) to solve the problem that static fusion cannot adapt to different scenarios and improve feature utilization.
[0043] The training process of this deep convolutional neural network includes:
[0044] Dataset construction: Publicly available remote sensing image datasets and glacial lake-specific labeled data were used;
[0045] Loss function design: A hybrid loss function is adopted.
[0046]
[0047] in Cross-entropy loss for boundary classification, The mean square error loss in water quantity prediction The correlation loss between upstream confluence characteristics and stability coefficient is considered. It is the total loss value of the model training, which is used to comprehensively guide the model optimization by weighting and fusing the losses of the three sub-tasks: boundary classification, water volume prediction, and stability assessment.
[0048] Training strategy: Initialize the VGG16 encoder using transfer learning, optimize with Adam at an initial learning rate of 0.001, decay by 0.1 every 10 rounds, and introduce gradient clipping and early stopping mechanisms to prevent overfitting;
[0049] Step 3: Dynamic Modeling of Glacial Lake Stability and Prediction of Breach Probability
[0050] By combining time-series hydrological data, a dynamic change model of ice lake stability is constructed using the boundary features, ice thickness features and upstream confluence features extracted in step two. By analyzing the ice structure stability threshold and water level change trend, the probability of ice lake collapse within a preset time period (e.g., 72 hours) is predicted.
[0051] In this embodiment, the construction of the dynamic change model of glacial lake stability includes:
[0052] Glacial lake water volume-water level coupling analysis: Based on the upstream confluence characteristics after fusion, the real-time water storage of the glacial lake is calculated (V=A×h, where A is the lake surface area and h is the average water depth), and the expansion trend of the glacial lake is analyzed in combination with the water level change rate.
[0053] Ice layer and dam stability assessment: The ice layer thickness change rate is retrieved by inverting SAR data, and the dam's anti-sliding stability is assessed by combining the height and slope of the glacial moraine dam monitored by lidar.
[0054] Calculation of glacial lake failure probability: Taking the rate of change of water volume, the rate of change of ice thickness, and the dam stability coefficient as inputs, a deep learning model with physical constraints (the loss function includes the mean square error of water volume prediction and a penalty term for violating the stability threshold) is input, and the output is the probability of glacial lake failure in the next 72 hours.
[0055] Step 4: Risk Classification and Early Warning Decision Output
[0056] Based on the predicted failure probability obtained in step three, risk levels are divided into low-risk, medium-risk, and high-risk categories. A risk map and targeted emergency response plans are generated and pushed to relevant user terminals, including:
[0057] Risk Level Classification: Preset failure probability thresholds (e.g., low risk: ≤10%, medium risk: 10-30%, high risk: >30%) are used to match different response plans. Based on these failure probability thresholds, plans are matched (e.g., high risk triggers "personnel evacuation + engineering rescue," medium risk triggers "strengthened monitoring + material preparation") to achieve a "prediction-decision" closed loop. The failure probability thresholds are set based on the following logic: First, through historical failure event data statistics and physical model simulations, the critical values of water volume change rate, ice erosion rate, and dam stability parameters during glacial lake failures of different scales are analyzed. Combined with the exposure of downstream disaster-bearing bodies (e.g., population density, infrastructure vulnerability) and emergency response capabilities, the quantitative boundary of the risk level is determined. Second, machine learning models are used to train on multi-source data (e.g., ice thickness retrieved from SAR, dam slope monitored by LiDAR, and meteorological and hydrological time-series data). Cross-validation is used to identify the probability distribution patterns of characteristic variables and actual failure events. ROC curve analysis is used to determine the optimal threshold combination under different false alarm and false negative rates. At the same time, the threshold setting must meet the principle of graded response to emergency management needs. For example, the low-risk threshold (≤10%) corresponds to routine monitoring and early warning, the medium-risk threshold (10%-30%) triggers local evacuation preparation, and the high-risk threshold (>30%) initiates emergency evacuation, so as to ensure that the implementation of the plan matches the pace of risk evolution.
[0058] Risk map generation: Based on DEM and breach path simulation, potential breach areas and impact ranges are marked, and reservoir capacity estimates and peak flow predictions are overlaid. DEM provides the topographic basis, breach path simulation (such as based on MIKE11 / MIKE21 hydrodynamic model) marks potential inundation areas, and reservoir capacity (affecting total flood volume) and peak flow (affecting damage intensity) are overlaid to intuitively display the spatial distribution of risk.
[0059] Early warning push: Risk maps, emergency plans and real-time monitoring data are pushed to emergency management department terminals and residents' APP through the cloud platform. It supports multi-language and multi-format downloads and multi-terminal support (emergency department PC / tablet, residents' APP) to ensure information reach. Multi-language / format (GeoJSON for GIS system, PDF for paper report) adapts to different user needs and realizes "precise early warning + rapid response".
[0060] The deep learning-based ice lake outburst prediction method in this embodiment supports lightweight deployment and real-time prediction, as detailed below:
[0061] Model compression: The model is compressed using knowledge distillation technology, retaining more than 95% of the accuracy; the student model uses depthwise separable convolution to reduce computational complexity;
[0062] Edge deployment: The compressed model is converted into TensorRT format and deployed on edge computing devices, achieving a single frame image processing time of ≤1.5s;
[0063] Online updates: Edge data is collected regularly through a federated learning framework to update the cloud model and push it to the edge, enabling continuous evolution.
[0064] This embodiment of the deep learning-based glacial lake outburst prediction method integrates reinforcement learning to dynamically optimize model parameters: the state is defined as the current glacial lake water volume, ice thickness, dam stability coefficient, and weather forecast; the action is the adjustment amount of model parameters; the reward function encourages accurate prediction and punishes false alarms and missed alarms; and the agent is trained through a proximal policy optimization algorithm to learn the optimal parameter adjustment strategy.
[0065] The deep learning-based glacial lake outburst prediction method in this embodiment also constructs a glacial lake outburst knowledge graph to assist decision-making: it integrates historical outburst events, geological conditions, and climate data to construct an "entity-relationship" network, analyzes entity associations through graph attention networks, and generates emergency plans by matching similar historical cases.
[0066] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A deep learning-based method for predicting glacial lake outburst floods, characterized in that, Includes the following steps: Step 1: Acquire multi-source remote sensing data and ground monitoring data related to the glacial lake, and perform standardized preprocessing to obtain multi-dimensional feature data that can be used for multi-dimensional feature extraction; Step 2: Based on the preprocessed multi-dimensional feature data, use a deep convolutional neural network to extract the boundary features, ice thickness features, and upstream confluence features of the glacial lake; Step 3: Combining time-series hydrological data, a dynamic change model of glacial lake stability is constructed using extracted boundary features, ice thickness features, and upstream confluence features to predict the probability of glacial lake outburst within a preset timeframe. Step 4: Based on the predicted probability of glacial lake outburst, classify the risk level of glacial lake outburst, generate a risk map and targeted emergency response plan, and push them to relevant user terminals.
2. The deep learning-based method for predicting glacial lake outburst floods according to claim 1, characterized in that: In step one, the multi-source remote sensing data includes lidar data, satellite optical images and synthetic aperture radar data, the ground monitoring data includes real-time meteorological data and hydrological observation data, and the standardization preprocessing includes spatial registration, temporal alignment, noise removal and missing data imputation.
3. The deep learning-based method for predicting glacial lake outburst floods according to claim 1, characterized in that: In step two, the deep convolutional neural network adopts a multi-branch parallel architecture. Each branch takes in preprocessed multi-source data, outputs corresponding features, and passes them to the main feature fusion module. Specifically, this includes: Spatial feature extraction branch: U-Net network with encoder-decoder structure; Time series analysis branch: Using a bidirectional long short-term memory network, historical meteorological data and hydrological data are input, long-term dependencies of the data are captured, and a prediction curve of future glacial lake water volume changes is output. Upstream confluence characteristic branch: Based on the elevation zoning model, the contributions of glacial meltwater, snowmelt and precipitation are calculated. After normalizing each contribution, it is input into the fully connected layer to generate a feature vector reflecting the intensity of upstream water inflow. Feature fusion module: It receives spatial features, temporal features and upstream confluence feature vectors through a gated loop unit, dynamically adjusts the weights of each feature, and outputs a fused multi-dimensional feature vector.
4. The deep learning-based method for predicting glacial lake outburst floods according to claim 1, characterized in that: In step two, the boundary features are accurately identified by fusing spectral features and multi-scale spatial features. The ice thickness features are obtained by analyzing SAR data, radar echo signals, and water color changes in satellite optical images. The upstream confluence features are generated by calculating the contributions of glacial meltwater, snowmelt, and precipitation.
5. The deep learning-based method for predicting glacial lake outburst floods according to claim 1, characterized in that: During training, the deep convolutional neural network is constructed using publicly available remote sensing image datasets and ice lake-specific labeled data. A hybrid loss function is then used to optimize the model, as shown in the formula: in, For boundary classification, cross-entropy loss, The mean square error loss in water quantity prediction The correlation loss between upstream confluence features and stability coefficient is used. The training strategy is to initialize the VGG16 encoder with transfer learning, perform Adam optimization with an initial learning rate of 0.001, and decrease the learning rate by 0.1 every 10 rounds.
6. The deep learning-based method for predicting glacial lake outburst floods according to claim 1, characterized in that: In step three, the construction process of the dynamic change model of glacial lake stability is as follows: based on the fused upstream confluence characteristics, the real-time water storage of the glacial lake is calculated, and combined with the water level change rate, the expansion trend of the glacial lake is analyzed. Then, the ice thickness change rate is inverted through SAR data, and combined with the height and slope of the glacial moraine dam monitored by lidar, the anti-sliding stability of the dam body is evaluated. Then, the water volume change rate, ice thickness change rate, and dam stability coefficient are used as inputs to input a deep learning model with physical constraints, and the probability of glacial lake collapse in the next 72 hours is output.
7. The deep learning-based method for predicting glacial lake outburst floods according to claim 1, characterized in that: In step four, when classifying risk levels, a default probability threshold is preset. If the probability is within 10% of the threshold, it is classified as low risk; if it is 10-30% higher than the threshold, it is classified as medium risk; and if it is 30% higher than the threshold, it is classified as high risk.
8. The deep learning-based method for predicting glacial lake outburst floods according to claim 1, characterized in that: In step four, when generating the risk map, the potential breach areas and impact ranges are marked based on the DEM and breach path simulation, and reservoir capacity estimation and peak flow prediction are overlaid. Then, the risk map, emergency plan and real-time monitoring data are pushed to the emergency management department terminal and residents' APP through the cloud platform.