Mountain debris flow dynamic early warning method based on deep learning

Abstract

The invention discloses a mountainous area debris flow dynamic early warning method based on deep learning, and relates to the technical field of geological disaster early warning, and the method comprises the steps: S1, marking a rainwater collection conventional region covered by a rainfall region as a rainwater collection risk region; s2, recording a real-time influence factor set of potential debris flow in each rainwater harvesting risk area; s3, inputting the real-time influence factor set as an iteration input value into the deep learning network for iteration, and setting an iteration output value as a prediction influence factor set; and S4, carrying out risk assessment on the recorded data of the prediction influence factor set by using an abnormal assessment strategy, screening out all rainwater collection risk areas meeting the risk assessment level, and marking the rainwater collection risk areas as debris flow risk areas. Compared with the prior art, the method utilizes the deep learning network to process the complex data set, covers more potential risk areas, and has the advantages of improving the accuracy of mountain debris flow early warning and recognizing the debris flow risk mountain areas more accurately.

Description

Technical Field

[0001] The present invention relates to the technical field of geological disaster early warning, and in particular to a dynamic early warning method for debris flow in mountainous areas based on deep learning. Background Art

[0002] Debris flows in mountainous areas mainly occur in valleys or on mountain slopes. These areas have complex terrain and steep slopes, which are conducive to the formation and movement of debris flows. The mountainous terrain allows water to quickly converge and carry a large amount of loose soil and rock materials to form debris flows. The terrain classification of debris flows in mountainous areas mainly includes valley type and hillside type. These two types of debris flows differ in formation conditions, distribution characteristics, and degree of harm. Debris flows in mountainous areas are extremely destructive natural disasters for both humans and nature because they often occur in areas with large terrain differences. In order to effectively reduce the losses caused by debris flow disasters, the establishment of a dynamic early warning system is crucial.

[0003] In the debris flow disaster in mountainous areas, heavy rainfall in the rainy season is an important factor in the occurrence of debris flow disasters. In the existing data processing technology for dynamic early warning of debris flow in mountainous areas, when dynamic early warning is carried out for debris flow in target mountainous areas with large mountain range and complex landforms, when there are multiple rainfall sources such as local rainstorms and comprehensive rainfall in the basin that do not cover the entire area, the accuracy of debris flow prediction is still relatively low, or more monitoring and early warning resources are required. On the one hand, when different types of rainfall occur simultaneously in mountainous areas with different height differences, the impact on the formation of debris flow is different, and the current prediction and early warning methods mainly tend to treat rainfall conditions and mountain areas as a whole, which affects the prediction response speed in mountainous areas with more debris flow prone areas, or requires more monitoring costs; on the other hand, in mountainous areas with large mountain ranges, the existing prediction methods have limited accuracy in estimating the occurrence of potential debris flow from a global perspective after capturing the changes in rainfall characteristics. The different rainfall intensity, duration and other characteristics of each region will also affect the triggering conditions and development process of debris flow, which may cause deviations or delays in the prediction results, resulting in the problem of insufficient timeliness of debris flow prediction results. Summary of the invention

[0004] The present invention provides a dynamic early warning method for debris flows in mountainous areas based on deep learning, which solves the problems of poor timeliness and limited effective early warning range of dynamic early warning for debris flows in mountainous areas when a target monitoring mountainous area with a large number of risk areas and large altitude differences encounters a complex rainfall source in the prior art.

[0005] The present invention is achieved through the following technical solutions: A dynamic early warning method for debris flow in mountainous areas based on deep learning, the method comprising: Step S1: Capture the geographic information of the target mountain area to construct a contour map of the target mountain area, mark the valley and saddle areas along the ridge line from each mountain top as regular rain collection areas on the contour map, collect precipitation information of the target mountain area and mark the regular rain collection areas covered by the precipitation area as rain collection risk areas on the contour map, and divide the whole day into several time periods; Step S2: setting an abnormal assessment strategy for predicting potential debris flow in the target mountain area, using the abnormal assessment strategy to divide the risk of potential debris flow in each rain-collecting risk area into multiple risk assessment levels, and recording the real-time influencing factor set of potential debris flow in each rain-collecting risk area in units of time periods according to the geographical information of the target mountain area and the precipitation information of the target mountain area; Step S3: construct a deep learning network, select one of the time periods, input the recorded data of the real-time impact factor set of each rainwater collection risk area as iterative input values ​​into the deep learning network for iteration to predict the effective change tendency of the real-time impact factor set of each rainwater collection risk area in the current time period, and set the iterative output value as the predicted impact factor set; Step S4: Use the abnormal assessment strategy to conduct risk assessment on the recorded data of the prediction impact factor set, screen out all the prediction impact factor sets that meet the risk assessment level, and mark all the corresponding rainwater risk areas as debris flow risk areas. After implementing the emergency plan for each debris flow risk area, continue to maintain real-time impact factor set records in the next time period.

[0006] In mountainous area debris flow disasters, heavy rainfall in the rainy season is an important factor in causing debris flow disasters. In the existing data processing technology for dynamic early warning of mountainous area debris flow, when dynamic early warning is carried out for mountainous area debris flow in target mountainous area with large mountain range and complex landform, when there are multiple rainfall sources that do not cover the entire area, such as local rainstorms, comprehensive rainfall in the basin, etc., the accuracy of debris flow prediction is still relatively low, or more monitoring and early warning resources are required. Based on this, the present invention provides a method for dynamic early warning of mountainous area debris flow based on deep learning, which solves the problems of poor timeliness and limited effective early warning range of dynamic early warning for mountainous area debris flow when the target monitoring mountainous area with more risk areas and large height difference changes encounters complex rainfall sources in the prior art.

[0007] Furthermore, the content of the abnormal assessment strategy includes: the recorded data of the real-time influencing factor set includes precipitation time value, precipitation intensity value and contour difference value representing the maximum height difference of contour lines in the current rain-collecting risk area, wherein the contour difference value is obtained by calculating the contour height value in each rain-collecting risk area on the contour map, and the precipitation time value and precipitation intensity value are obtained based on the precipitation information of the target mountain area in the rain-collecting risk area; the precipitation time value and precipitation intensity value in each rain-collecting risk area are recorded in each time period, and data thresholds are set for the precipitation time value, precipitation intensity value and contour difference value in each rain-collecting risk area to classify and set the risk assessment level of the corresponding rain-collecting risk area.

[0008] Furthermore, the data thresholds of the recorded data include: setting a first height difference threshold and a second height difference threshold for the contour line difference value, wherein the second height difference threshold is greater than the first height difference threshold; setting a precipitation duration threshold and a precipitation intensity threshold for the precipitation time value and the precipitation intensity value respectively; the risk assessment levels include: First assessment level: in the current rainwater risk area in the current time period, the precipitation time value is lower than the precipitation duration threshold, the precipitation intensity value is lower than the precipitation intensity threshold, and the contour line difference value is lower than the second height difference threshold; Second assessment level: in the current rainwater risk area in the current time period, only one recorded data between the precipitation time value and the precipitation intensity value is higher than the corresponding data threshold, and the contour line difference is between the first height difference threshold and the second height difference threshold; Third assessment level: in the current rainwater risk area in the current time period, only one recorded data between the precipitation time value and the precipitation intensity value is higher than the corresponding data threshold, and the contour line difference value is higher than the second height difference threshold; Fourth assessment level: in the current rainwater risk area in the current time period, the precipitation time value is higher than the precipitation duration threshold, the precipitation intensity value is higher than the precipitation intensity threshold, and the contour line difference value is higher than the first height difference threshold.

[0009] Furthermore, the current rainwater risk area in the current time period records data about the set of prediction influencing factors output by the iteration, the precipitation time value represents the final predicted maximum precipitation time length, the precipitation intensity value represents the final predicted maximum precipitation intensity, and the contour line difference represents a set of contour lines in the final predicted highest risk interval area.

[0010] Furthermore, the construction content of the deep learning network includes: Construct a convolutional neural network as a deep learning network, use the precipitation time value, precipitation intensity value and contour difference value recorded in the current rain collection risk area as the input value of the convolutional neural network, and output the iterative value of the convolutional neural network as the prediction influencing factor set; Based on the historical data records of precipitation information and geographic information of the target mountain area, a training set and a test set of the convolutional neural network are constructed, and the input forms of the precipitation time value, precipitation intensity value and contour difference value are converted into a two-dimensional array for CNN input format. Multiple convolutional layers are used to extract the spatial features of the geographic information and precipitation information of the target mountain area as compensation features of the convolution kernel. Then, an iterative optimization method is used for iterative training, and the predicted influencing factor set is output through the output layer.

[0011] Furthermore, a genetic optimization method is used in a convolutional neural network to iteratively predict the real-time impact factor set, and the process includes: Step A1: Initialize a group of population individuals, each individual represents a set of convolution parameters for the convolutional neural network, and preset the fitness function and iteration number of the iteration individual. At each iteration, the fitness score of each individual is calculated based on the fitness function. The fitness neural network performs forward propagation with the input real-time impact factor set according to the convolution parameters of each iteration individual; Step A2: preset a fitness threshold, screen out individuals with a fitness score greater than or equal to the fitness threshold based on the calculated fitness score, reproduce the screened individuals into new individuals in a uniform crossover manner, add the new individuals after crossover reproduction into the population to replace the individuals with low fitness, and set each screening and reproduction process of the population individuals as an iteration; Step A3: Set the real-time impact factor set corresponding to the individual with the highest fitness score generated in each iteration as the real-time impact factor set generated in that iteration. When the number of iterations is reached or there is no individual with a fitness score greater than or equal to the fitness threshold, the previous iteration process is regarded as the final iteration, and the real-time impact factor set generated by the previous iteration is set as the predicted impact factor set of the deep learning network iteration output.

[0012] Furthermore, the fitness function is set to the negative value of the cross entropy loss to measure the difference between the probability distribution of the model output and the true label distribution.

[0013] Furthermore, the calculation process of the fitness score includes: The convolutional neural network is trained using the convolution parameters of the current target individual, the predicted probability distribution of each sample is calculated, the cross entropy loss is calculated based on the true label and the predicted probability, and the fitness score of each individual is obtained by taking the negative value of the cross entropy loss and normalizing it. The fitness score is positively correlated with the model performance.

[0014] Furthermore, uncertainty assessment is introduced using a Bayesian approach to obtain confidence intervals for the predicted probability distribution of each sample.

[0015] Furthermore, the compensation characteristics of the convolution kernel include: taking the vegetation coverage information and soil coverage information in the target mountain area's geographic information, and the wind change information and air pressure change information in the target mountain area's precipitation information as compensation characteristics into the convolution kernel processing, and the compensation characteristics of the convolution kernel are set as compensation parameters within the convolution parameters of all the individuals.

[0016] Compared with the existing technology, the present invention sets multiple risk assessment levels and uses deep learning networks to process complex data sets, so that the warning range is wider and covers more potential risk areas. It has the advantages of improving the accuracy of mountain debris flow warnings and helping to more accurately identify and classify mountain areas with debris flow risks. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The drawings described herein are used to provide a further understanding of the embodiments of the present invention, constitute a part of this application, and do not constitute a limitation of the embodiments of the present invention. In the drawings: Figure 1 It is a flowchart of the present invention. DETAILED DESCRIPTION

[0018] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with embodiments and drawings. The exemplary embodiments of the present invention and their description are only used to explain the present invention and are not intended to limit the present invention. Example

[0019] like Figure 1 As shown, this embodiment is a dynamic early warning method for debris flow in mountainous areas based on deep learning, and the method includes: Step S1: Capture the geographic information of the target mountain area to construct a contour map of the target mountain area, mark the valley and saddle areas along the ridge line from each mountain top as regular rain collection areas on the contour map, collect precipitation information of the target mountain area and mark the regular rain collection areas covered by the precipitation area as rain collection risk areas on the contour map, and divide the whole day into several time periods; Step S2: setting an abnormal assessment strategy for predicting potential debris flow in the target mountain area, using the abnormal assessment strategy to divide the risk of potential debris flow in each rain-collecting risk area into multiple risk assessment levels, and recording the real-time influencing factor set of potential debris flow in each rain-collecting risk area in units of time periods according to the geographical information of the target mountain area and the precipitation information of the target mountain area; Step S3: construct a deep learning network, select one of the time periods, input the recorded data of the real-time impact factor set of each rainwater collection risk area as iterative input values ​​into the deep learning network for iteration to predict the effective change tendency of the real-time impact factor set of each rainwater collection risk area in the current time period, and set the iterative output value as the predicted impact factor set; Step S4: Use the abnormal assessment strategy to conduct risk assessment on the recorded data of the prediction impact factor set, screen out all the prediction impact factor sets that meet the risk assessment level, and mark all the corresponding rainwater risk areas as debris flow risk areas. After implementing the emergency plan for each debris flow risk area, continue to maintain real-time impact factor set records in the next time period.

[0020] The target mountain area geographical information mainly includes contour maps or isotherm maps showing topographic features such as ridges, valleys, saddles, etc., water distribution of the location and flow direction of rivers, lakes and other water bodies, different types of vegetation and their distribution coverage, and information such as soil permeability and clay content in each region. In conventional technology, the target mountain area geographical information can be obtained through remote sensing technology or geographic information system GIS and existing geographic databases. The target mountain area precipitation information mainly includes data information such as precipitation intensity indicating precipitation per unit time, precipitation time indicating rainfall duration, and spatial distribution of precipitation in the region. In conventional technology, the target mountain area precipitation information can be obtained through meteorological station monitoring, meteorological radar monitoring, and remote sensing monitoring. On the contour map, the valley and saddle areas are all marked as regular rain collection areas along the ridge line from each mountain top because the ridge line is the dividing line of water flow, and water usually flows to low-lying areas along the direction of gravity. Therefore, marking the rain collection area along the ridge line can capture the trend of water flow convergence. By marking valleys and saddles and other sunken areas, it is possible to clearly identify where rainwater gathers, helping to identify potential rain collection risk areas. The rain-collecting risk area is a conventional rain-collecting area with precipitation. The conventional rain-collecting area covered by the precipitation area is marked as the rain-collecting risk area because the frequency of wind direction in the mountainous area is high, and cumulonimbus clouds are prone to move within a time period and within a certain mountainous area. It is the most appropriate range size to predict the potential area of ​​debris flow in units of conventional rain-collecting areas. The contour map provides information on the height change of the terrain, which helps to understand the impact of the terrain on precipitation, including the accumulation and distribution of water flow. By defining the conventional rain-collecting area, the relationship between rainfall and debris flow can be better analyzed, and high-risk areas can be identified. Dividing the whole day into several time periods is convenient for more detailed monitoring of precipitation changes and management in different time periods. The allocation method of the time period can be unspecified. It can be divided equally into 24 hours, or the segment size can be adjusted according to the time period when precipitation is prone to occur in the target mountainous area. For example, with the help of precipitation information in the target mountainous area, a larger monitoring time period is allocated for the prediction time range with a high probability of precipitation. More importantly, the real-time influencing factor set is used as input, and the deep learning network is used for training and prediction, and it is ensured that the model can effectively capture the dynamic changes of time series data. The anomaly assessment strategy is used as a comparison standard for measuring the set of predicted impact factors output by the deep learning network, and is used to take corresponding response measures for different risk levels. Its effectiveness depends on whether the selected impact factors are comprehensive and whether they can reflect the risk of potential debris flow. The implementation of emergency plans in debris flow risk areas requires ensuring the feasibility and timeliness of various measures to reduce possible losses. After each time period, keeping records and updates of real-time impact factors will help dynamically adjust the early warning strategy and enhance the response capability of the early warning system.

[0021] Furthermore, the content of the abnormal assessment strategy includes: the recorded data of the real-time influencing factor set includes precipitation time value, precipitation intensity value and contour difference value representing the maximum height difference of contour lines in the current rain-collecting risk area, wherein the contour difference value is obtained by calculating the contour height value in each rain-collecting risk area on the contour map, and the precipitation time value and precipitation intensity value are obtained based on the precipitation information of the target mountain area in the rain-collecting risk area; the precipitation time value and precipitation intensity value in each rain-collecting risk area are recorded in each time period, and data thresholds are set for the precipitation time value, precipitation intensity value and contour difference value in each rain-collecting risk area to classify and set the risk assessment level of the corresponding rain-collecting risk area.

[0022] The setting of the contour difference, precipitation time value and precipitation intensity value is based on the theory of hydrometeorology and geographic information system, combining the time and intensity of precipitation with terrain factors, and providing comprehensive data support for the real-time influencing factor set. By comprehensively analyzing these three data, it is possible to more effectively predict the risk of debris flow and formulate corresponding response strategies. The larger the contour difference, that is, the denser the contour lines, the higher the steepness of the mountain, and the greater the probability of potential debris flow. The risk assessment of the recorded data of the prediction influencing factor set is performed using an abnormal assessment strategy, that is, the potential debris flow degree of the rain collection risk area is obtained by comparing the recorded data of the prediction influencing factor set output by iteration with the risk assessment level, and a corresponding emergency plan is made to deal with the potential debris flow risk. The contour difference can mainly directly evaluate the convergence of water flow and soil stability. A large height difference means that water flow may gather in areas with large terrain changes, thereby increasing the risk of debris flow. When water flows down a steep slope, it tends to form rapid flows, leading to debris flows. Contour line differences can help assess the stability of soil in an area. Larger height differences may mean that the soil is more affected by gravity and water flow, thereby increasing the potential risk of debris flows.

[0023] Furthermore, the data thresholds of the recorded data include: setting a first height difference threshold and a second height difference threshold for the contour line difference value, wherein the second height difference threshold is greater than the first height difference threshold; setting a precipitation duration threshold and a precipitation intensity threshold for the precipitation time value and the precipitation intensity value respectively; the risk assessment levels include: First assessment level: in the current rainwater risk area in the current time period, the precipitation time value is lower than the precipitation duration threshold, the precipitation intensity value is lower than the precipitation intensity threshold, and the contour line difference value is lower than the second height difference threshold; Second assessment level: in the current rainwater risk area in the current time period, only one recorded data between the precipitation time value and the precipitation intensity value is higher than the corresponding data threshold, and the contour line difference is between the first height difference threshold and the second height difference threshold; Third assessment level: in the current rainwater risk area in the current time period, only one recorded data between the precipitation time value and the precipitation intensity value is higher than the corresponding data threshold, and the contour line difference value is higher than the second height difference threshold; Fourth assessment level: in the current rainwater risk area in the current time period, the precipitation time value is higher than the precipitation duration threshold, the precipitation intensity value is higher than the precipitation intensity threshold, and the contour line difference value is higher than the first height difference threshold.

[0024] Among them, the current rainwater risk area in the current time period, the recorded data about the set of prediction influencing factors output by the iteration, the precipitation time value represents the final predicted maximum precipitation time length, the precipitation intensity value represents the final predicted maximum precipitation intensity, and the contour line difference represents a set of contour lines in the final predicted highest risk interval area.

[0025] The risk assessment level in this embodiment mainly represents the key potential debris flow identification index, not all measurement level options. The precipitation duration threshold refers to the minimum duration of precipitation. Exceeding this time period may cause soil saturation and increase the risk of debris flow; if the precipitation duration exceeds the set threshold, it means that precipitation may have a significant impact on the soil structure, which may cause water accumulation and trigger debris flow. The first height difference threshold and the second height difference threshold divide the contour height difference into three levels. Because the second height difference threshold is greater than the first height difference threshold, in the specific implementation, the recorded data is judged as a low-risk height difference when it is lower than the first height difference threshold (that is, the height difference is at a lower value), the recorded data is judged as a medium-risk height difference when it is between the first height difference threshold and the second height difference threshold, and the recorded data is judged as a high-risk height difference when it is higher than the second height difference threshold (that is, the height difference is at a higher value). This threshold helps to judge the potential dangerous period and provide a basis for early warning. The precipitation intensity threshold refers to the minimum amount of precipitation per unit time. Exceeding this intensity may lead to an increase in surface water flow and increase the risk of debris flow. When the precipitation intensity exceeds the set threshold, it indicates that the precipitation intensity is sufficient to cause soil erosion or flow, increasing the possibility of debris flow. The contour height difference threshold can divide areas with different terrain features into different risk levels. Areas above the threshold may be marked as high-risk areas; the first height difference threshold and the second height difference threshold reflect the undulation of the terrain, and a larger height difference usually means a steeper slope. This terrain feature will affect the speed and direction of the water flow, and thus affect the occurrence of debris flow.

[0026] In a specific implementation, the setting method of the contour height difference threshold, the precipitation duration threshold and the precipitation intensity threshold may not be specifically limited; the first height difference threshold and the second height difference threshold can be used to identify the frequency of debris flow at different height differences by analyzing the data of historical debris flow events. The critical value of a certain height difference when a debris flow occurs can be determined by statistical methods; the height difference threshold can also be set in combination with terrain characteristics and hydrological models to ensure that a risk warning is triggered when a large height difference occurs. The precipitation duration threshold can be obtained by analyzing the relationship between different precipitation durations and the occurrence of debris flow, and finding the critical point of a specific duration based on historical data, where longer precipitation usually increases soil saturation; it can also be combined with the precipitation duration predicted by the meteorological model, and the critical value within a certain period of time can be selected as the precipitation duration threshold to judge the impact of precipitation on debris flow. The precipitation intensity threshold can use the real-time precipitation data provided by the meteorological monitoring station to set the threshold at different intensities. Generally, the critical point of risk can be determined according to the standard of short-term heavy precipitation.

[0027] When the set of predicted impact factors iteratively output by the monitored rainwater risk area meets the first assessment level, it means that the rainwater risk area meets the low-risk potential debris flow standard in this time period. The precipitation time value is lower than the precipitation duration threshold, and the precipitation intensity value is lower than the precipitation intensity threshold, which means that the total precipitation in the rainwater risk area in this time period is at a low level; at the same time, as long as the height difference of the contour lines does not exceed the second height difference threshold, that is, the height difference of the contour lines has not reached a higher level, the rainwater risk area can be judged as a low-risk area, and the risk level decreases as the height difference of the contour lines decreases. The corresponding emergency plan can be set to not evacuate personnel in advance, but only continue to monitor the rainwater risk area.

[0028] When the set of predicted impact factors iteratively outputted by the monitored rainwater risk area meets the second assessment level, it means that the rainwater risk area meets the medium risk potential debris flow standard in this time period. There is only one recorded data between the precipitation time value and the precipitation intensity value that is higher than the corresponding data threshold, which means that when the precipitation time value is lower than the precipitation duration threshold, the precipitation intensity value is higher than the precipitation intensity threshold, and vice versa, that is, at least one of the precipitation time value and the precipitation intensity value is in a relatively high state, then the total precipitation in the rainwater risk area has exceeded the low state and is at a medium precipitation level. At this time, the contour line difference is between the first height difference threshold and the second height difference threshold, that is, the contour line difference is at a medium height difference level, then the potential debris flow risk assessment in the currently monitored rainwater risk area is judged to be medium risk, and the corresponding emergency plan can be implemented in the specific implementation to guide the residents near the rainwater risk area to transfer, and maintain regional monitoring, and guide the residents to return after several time periods or the precipitation process in the rainwater risk area stops for a period of time.

[0029] When the set of prediction influence factors iteratively outputted by the monitored rainwater risk area meets the third assessment level or the fourth assessment level, it means that the rainwater risk area meets the high-risk potential debris flow standard in this time period. When it meets the third assessment level, as with the second assessment level, at least one of the precipitation time value and the precipitation intensity value is in a relatively high state, then the total precipitation in the rainwater risk area has exceeded the low state and is at a medium precipitation level. At this time, the contour line difference is higher than the second height difference threshold, that is, the contour line difference is at a higher height difference level, that is, when the precipitation is not low, the height difference of the target rainwater risk area is large, which usually means that the terrain is undulating, that is, the slope of the mountain is steep or the height changes rapidly. Similarly, when the set of prediction influence factors meets the fourth assessment level, the precipitation time value is higher than the precipitation duration threshold, and the precipitation intensity value is higher than the precipitation intensity threshold, which means that the total precipitation in the target rainwater risk area in the current time period has reached a high level. At this time, even if the height difference of the contour line only exceeds the first height difference threshold, the mountain still has a high potential debris flow risk. At this time, the potential debris flow risk assessment result in the currently monitored rainwater risk area is determined to be high risk. The emergency plan can be further improved on the basis of the medium-risk plan, such as increasing the efforts to notify the management department to increase more human and material resources to help residents transfer and extend the time interval for residents to return.

[0030] Furthermore, as a feasible implementation method, the construction content of the deep learning network includes: Construct a convolutional neural network as a deep learning network, use the precipitation time value, precipitation intensity value and contour difference value recorded in the current rain collection risk area as the input value of the convolutional neural network, and output the iterative value of the convolutional neural network as the prediction influencing factor set; Based on the historical data records of precipitation information and geographic information of the target mountain area, a training set and a test set of a convolutional neural network are constructed, and the input forms of the precipitation time value, precipitation intensity value and contour line difference value are converted into a two-dimensional array for CNN input format, and multiple convolutional layers are used to extract the spatial features of the geographic information and precipitation information of the target mountain area as compensation features of the convolution kernel, and then iterative training is performed using an iterative optimization method, and a set of predicted influencing factors is output through an output layer; The compensation features of the convolution kernel include: taking vegetation coverage information and soil coverage information in the target mountainous area geographic information, and wind change information and air pressure change information in the target mountainous area precipitation information as compensation features and entering the convolution kernel for processing.

[0031] Collect historical data records of precipitation information and geographic information of target mountain areas, including precipitation data of target mountain areas, including precipitation amount, precipitation intensity, precipitation duration, etc., and geographic information of target mountain areas, including contour maps, terrain undulations, soil types, vegetation coverage and other data, for constructing training sets and test sets. Standardize or normalize precipitation and geographic information data so that different features can be analyzed at the same scale to help improve the training effect of the model. When extracting key features, select features related to the risk of debris flow, including precipitation time value, precipitation intensity value, contour line difference, etc., and consider other environmental features that may affect the risk. Divide the processed data into training sets and test sets. In specific implementation, 70%-80% of the data is usually used as the training set, and 20%-30% of the data is used as the test set to facilitate subsequent model training and verification. Regarding the vegetation coverage information and soil coverage information in the target mountainous area geographic information, as well as the wind change information and air pressure change information in the target mountainous area precipitation information, the vegetation coverage has a significant impact on the absorption of water and the stability of the soil. High vegetation coverage can reduce surface runoff and reduce the risk of debris flow. Therefore, vegetation information as a feature can help the model understand the hydrological characteristics of different regions; the soil type (such as sand, clay, stone, etc.) has a direct impact on the permeability and retention capacity of water. Different soil properties will affect the distribution and flow of water after precipitation, thereby affecting the occurrence of debris flow. Therefore, soil information can provide the necessary geological background for the convolution kernel; the wind change can affect the distribution and intensity of precipitation. For example, changes in wind direction may cause precipitation to accumulate in certain areas, thereby affecting the risk of debris flow. Incorporating wind change information into features can improve the model's understanding of precipitation patterns; the air pressure change is usually related to changes in weather systems and may affect the intensity and duration of precipitation. Low air pressure is usually accompanied by heavy precipitation. Therefore, air pressure information can provide an important meteorological background for debris flow risk.

[0032] Furthermore, as a feasible implementation method, a genetic optimization method is used in a convolutional neural network to iteratively predict the real-time impact factor set, and the process includes: Step A1: Initialize a group of population individuals, each individual represents a set of convolution parameters for the convolutional neural network, and preset the fitness function and iteration number of the iteration individual. At each iteration, the fitness score of each individual is calculated based on the fitness function. The fitness neural network performs forward propagation with the input real-time impact factor set according to the convolution parameters of each iteration individual; Step A2: preset a fitness threshold, screen out individuals with a fitness score greater than or equal to the fitness threshold based on the calculated fitness score, reproduce the screened individuals into new individuals in a uniform crossover manner, add the new individuals after crossover reproduction into the population to replace the individuals with low fitness, and set each screening and reproduction process of the population individuals as an iteration; Step A3: The real-time impact factor set corresponding to the individual with the highest fitness score generated in each iteration is set as the real-time impact factor set generated in the iteration. When the number of iterations is reached or there is no individual with a fitness score greater than or equal to the fitness threshold, the previous iteration process is regarded as the final iteration, and the real-time impact factor set generated in the previous iteration is set as the predicted impact factor set output by the deep learning network iteration. Among them, the compensation characteristics of the convolution kernel are set as compensation parameters within the convolution parameters of all the individuals; the fitness function is set to the negative value of the cross entropy loss to measure the difference between the probability distribution of the model output and the true label distribution.

[0033] Each individual represents a set of convolution parameters, such as the size, number, and step length of the convolution kernel. At the same time, each individual can contain the size of multiple convolution kernels, and each individual should contain the number of convolution kernels used in the convolution layer, which affects the complexity and depth of feature extraction. The padding method used in the convolution operation, such as "same" padding or "no" padding, will affect the size and information retention of the output feature map. The convolution parameters are used to affect feature extraction. The performance of the convolutional neural network depends largely on the setting of the convolution parameters; different parameter combinations will affect the model's ability to extract features from the input data, thereby affecting the final prediction results. By initializing different convolution parameter combinations, the genetic algorithm can explore a wider parameter space to find the optimal convolution setting, thereby improving the prediction accuracy of the real-time influencing factor set. The fitness neural network is used to calculate the performance of the convolution parameters of the current individual in the actual prediction task. Its purpose is to evaluate the effectiveness of the convolution parameter combination. During the forward propagation process, the convolution parameters will be applied to the input real-time influencing factor set to help the model extract features and generate output. The input real-time influencing factor set is transmitted through the neural network layer, and after processing such as the convolution layer and activation function, the final output is gradually formed. This process reflects the layer-by-layer transformation and feature extraction of the input data by the neural network. In each iteration, the fitness neural network performs forward propagation based on the latest convolution parameters and real-time influencing factor set, and dynamically adjusts the convolution parameters to optimize the model performance. In specific applications, the setting method of the fitness threshold can be based on the fitness distribution of individuals in the current population, setting a specific percentile (such as 75% or 90%), and screening individuals with fitness higher than this percentile as excellent individuals; or in the process of model training, you can use cross-validation and combine experience in this field to set the fitness threshold through the performance on the validation set to ensure that the selected threshold can maintain good performance on unseen data.

[0034] More, the uniform crossover is a crossover operation in a genetic algorithm, which aims to combine the characteristics of two parent individuals to produce new offspring individuals. In uniform crossover, each gene (i.e., the convolution parameter in this embodiment) has the same probability of being selected from two parents, and this random selection ensures genetic diversity. The cross entropy is a measure of the difference between two probability distributions, which is usually used for classification problems. It calculates the difference between the probability distribution predicted by the model and the actual label (true distribution). In genetic algorithms, the fitness function is used to evaluate the performance of each individual (parameter combination). By calculating the cross entropy loss, the degree of match between the output of the model and the true label can be quantified. After the cross entropy loss is negative, the fitness function becomes a "bigger the better" form. In genetic algorithms, the optimization goal is to maximize the fitness function, so by taking a negative value, the optimization process tends to minimize the cross entropy loss, indirectly achieving the closeness of the model output probability distribution to the true label.

[0035] Furthermore, as a feasible implementation, the fitness score calculation process includes: The convolutional neural network is trained using the convolution parameters of the current target individual, the predicted probability distribution of each sample is calculated, the cross entropy loss is calculated based on the true label and the predicted probability, and the fitness score of each individual is obtained by taking the negative value of the cross entropy loss and normalizing it. The fitness score is positively correlated with the model performance; the Bayesian method is used to introduce uncertainty assessment to obtain the confidence interval of the predicted probability distribution of each sample.

[0036] When training a convolutional neural network, the model generates a predicted probability distribution for each sample based on the input data (such as a set of real-time impact factors). This distribution represents the model's confidence in each category, and the output is usually converted into a probability value using a softmax function. By calculating the cross entropy loss and performing negative and normalization processing, it is ensured that the fitness score can accurately reflect the predictive performance of the convolutional neural network on a specific sample, providing an effective basis for the optimization of the genetic algorithm. The Bayesian method is based on the Bayesian theorem, and by combining prior knowledge (prior distribution) with observed data (likelihood function), the belief in the model parameters is updated to form a posterior distribution. The introduction of uncertainty assessment not only improves the adaptability of the model to the data, but also adds a dimension to the calculation of the fitness score, which enables the genetic algorithm to consider the reliability of the model prediction when selecting individuals.

[0037] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A dynamic early warning method for debris flow in mountainous areas based on deep learning, characterized in that: The method includes: Step S1: Capture the geographic information of the target mountain area to construct a contour map of the target mountain area, mark the valley and saddle areas along the ridge line from each mountain top as regular rain collection areas on the contour map, collect precipitation information of the target mountain area and mark the regular rain collection areas covered by the precipitation area as rain collection risk areas on the contour map, and divide the whole day into several time periods; Step S2: setting an abnormal assessment strategy for predicting potential debris flow in the target mountain area, using the abnormal assessment strategy to divide the risk of potential debris flow in each rain-collecting risk area into multiple risk assessment levels, and recording the real-time influencing factor set of potential debris flow in each rain-collecting risk area in units of time periods according to the geographical information of the target mountain area and the precipitation information of the target mountain area; Step S3: construct a deep learning network, select one of the time periods, input the recorded data of the real-time impact factor set of each rainwater collection risk area as iterative input values ​​into the deep learning network for iteration to predict the effective change tendency of the real-time impact factor set of each rainwater collection risk area in the current time period, and set the iterative output value as the predicted impact factor set; Step S4: Use the abnormal assessment strategy to conduct risk assessment on the recorded data of the prediction impact factor set, screen out all the prediction impact factor sets that meet the risk assessment level, and mark all the corresponding rainwater risk areas as debris flow risk areas. After implementing the emergency plan for each debris flow risk area, continue to maintain real-time impact factor set records in the next time period.

2. A deep learning-based dynamic early warning method for debris flow in mountainous areas according to claim 1, characterized in that: The content of the abnormal assessment strategy includes: the recorded data of the real-time influencing factor set includes precipitation time value, precipitation intensity value and contour difference value representing the maximum height difference of contour lines in the current rain-collecting risk area, wherein the contour difference value is obtained by calculating the contour height value in each rain-collecting risk area on the contour map, and the precipitation time value and precipitation intensity value are obtained based on the precipitation information of the target mountain area in the rain-collecting risk area; the precipitation time value and precipitation intensity value in each rain-collecting risk area are recorded in each time period, and data thresholds are set for the precipitation time value, precipitation intensity value and contour difference value in each rain-collecting risk area to classify and set the risk assessment level of the corresponding rain-collecting risk area.

3. A deep learning-based dynamic early warning method for debris flow in mountainous areas according to claim 2, characterized in that: The data thresholds of the recorded data include: setting a first height difference threshold and a second height difference threshold for the contour line difference value, wherein the second height difference threshold is greater than the first height difference threshold; setting a precipitation duration threshold and a precipitation intensity threshold for the precipitation time value and the precipitation intensity value respectively; the risk assessment levels include: First assessment level: in the current rainwater risk area in the current time period, the precipitation time value is lower than the precipitation duration threshold, the precipitation intensity value is lower than the precipitation intensity threshold, and the contour line difference value is lower than the second height difference threshold; Second assessment level: in the current rainwater risk area in the current time period, only one recorded data between the precipitation time value and the precipitation intensity value is higher than the corresponding data threshold, and the contour line difference is between the first height difference threshold and the second height difference threshold; Third assessment level: in the current rainwater risk area in the current time period, only one recorded data between the precipitation time value and the precipitation intensity value is higher than the corresponding data threshold, and the contour line difference value is higher than the second height difference threshold; Fourth assessment level: in the current rainwater risk area in the current time period, the precipitation time value is higher than the precipitation duration threshold, the precipitation intensity value is higher than the precipitation intensity threshold, and the contour line difference value is higher than the first height difference threshold.

4. A deep learning-based dynamic early warning method for debris flow in mountainous areas according to any one of claims 1 to 3, characterized in that: The current rainwater risk area in the current time period, the recorded data about the set of prediction influencing factors output by the iteration, the precipitation time value represents the maximum precipitation time length of the final prediction, the precipitation intensity value represents the maximum precipitation intensity of the final prediction, and the contour line difference represents a set of contour lines of the highest risk interval area of ​​the final prediction.

5. The method for dynamic early warning of debris flow in mountainous areas based on deep learning according to claim 2, characterized in that: The construction content of the deep learning network includes: Construct a convolutional neural network as a deep learning network, use the precipitation time value, precipitation intensity value and contour difference value recorded in the current rain collection risk area as the input value of the convolutional neural network, and output the iterative value of the convolutional neural network as the prediction influencing factor set; Based on the historical data records of precipitation information and geographic information of the target mountain area, a training set and a test set of the convolutional neural network are constructed, and the input forms of the precipitation time value, precipitation intensity value and contour difference value are converted into a two-dimensional array for CNN input format. Multiple convolutional layers are used to extract the spatial features of the geographic information and precipitation information of the target mountain area as compensation features of the convolution kernel. Then, an iterative optimization method is used for iterative training, and the predicted influencing factor set is output through the output layer.

6. The method for dynamic early warning of debris flow in mountainous areas based on deep learning according to claim 2, characterized in that: The genetic optimization method is used in the convolutional neural network to iteratively predict the real-time impact factor set. The process includes: Step A1: Initialize a group of population individuals, each individual represents a set of convolution parameters for the convolutional neural network, and preset the fitness function and iteration number of the iteration individual. At each iteration, the fitness score of each individual is calculated based on the fitness function. The fitness neural network performs forward propagation with the input real-time impact factor set according to the convolution parameters of each iteration individual; Step A2: preset a fitness threshold, screen out individuals with a fitness score greater than or equal to the fitness threshold based on the calculated fitness score, reproduce the screened individuals into new individuals in a uniform crossover manner, add the new individuals after crossover reproduction into the population to replace the individuals with low fitness, and set each screening and reproduction process of the population individuals as an iteration; Step A3: Set the real-time impact factor set corresponding to the individual with the highest fitness score generated in each iteration as the real-time impact factor set generated in that iteration. When the number of iterations is reached or there is no individual with a fitness score greater than or equal to the fitness threshold, the previous iteration process is regarded as the final iteration, and the real-time impact factor set generated by the previous iteration is set as the predicted impact factor set of the deep learning network iteration output.

7. A deep learning-based dynamic early warning method for debris flow in mountainous areas according to claim 6, characterized in that: The fitness function is set to the negative value of the cross entropy loss to measure the difference between the probability distribution of the model output and the true label distribution.

8. The method for dynamic early warning of debris flow in mountainous areas based on deep learning according to claim 6, characterized in that: The fitness score calculation process includes: The convolutional neural network is trained using the convolution parameters of the current target individual, the predicted probability distribution of each sample is calculated, the cross entropy loss is calculated based on the true label and the predicted probability, and the fitness score of each individual is obtained by taking the negative value of the cross entropy loss and normalizing it. The fitness score is positively correlated with the model performance.

9. The method for dynamic early warning of debris flow in mountainous areas based on deep learning according to claim 8, characterized in that: Uncertainty assessment is introduced using a Bayesian approach to obtain confidence intervals for the predicted probability distribution of each sample.

10. A deep learning-based dynamic early warning method for debris flow in mountainous areas according to any one of claims 5 to 6, characterized in that: The compensation features of the convolution kernel include: taking the vegetation coverage information and soil coverage information in the target mountain area's geographic information, and the wind change information and air pressure change information in the target mountain area's precipitation information as compensation features into the convolution kernel processing, and the compensation features of the convolution kernel are set as compensation parameters within the convolution parameters of all the individuals.

Images