Artificial intelligence training dataset construction method and device for avalanche disaster prediction

Abstract

The application provides an artificial intelligence training data set construction method and device for avalanche disaster prediction, and belongs to the technical field of avalanche disaster prediction. The method uses avalanche event position information, avalanche disaster information and data correlation to construct a training data set. The above-mentioned avalanche event position information describes the occurrence place of the avalanche event from three aspects of point, line and surface. The above-mentioned avalanche disaster information describes the occurrence process of the avalanche event from five stages of disaster gestation, disaster causing, disaster, disaster bearing and rescue. The above-mentioned data correlation associates the above-mentioned avalanche event position information with the above-mentioned avalanche disaster information, so that the training data set can fully and effectively describe the avalanche disaster, provide sufficient effective training samples for the artificial intelligence model to realize accurate prediction of the avalanche disaster, help to formulate corresponding control means, and thus reduce the damage caused by the avalanche disaster.

Description

Technical Field

[0001] This invention relates to the field of avalanche disaster prediction technology, and in particular to a method and apparatus for constructing an artificial intelligence training dataset for avalanche disaster prediction. Background Technology

[0002] Avalanches are large-scale snow collapses that occur in snow-covered mountainous areas, characterized by their suddenness and destructive power. In managing snow-covered mountainous areas, using artificial intelligence models to predict avalanches can help develop appropriate control measures, enabling early evacuation of people and reducing the damage caused by avalanches.

[0003] However, due to the relative scarcity of avalanche events and the lack of effective information related to avalanche events, artificial intelligence models for avalanche disaster prediction lack sufficient effective training samples, thus failing to predict avalanche disasters and consequently failing to reduce the damage caused by avalanche disasters. Summary of the Invention

[0004] This invention proposes a method and apparatus for constructing an artificial intelligence training dataset for avalanche disaster prediction. The training dataset is constructed using avalanche event location information, avalanche disaster information, and data correlations. The training samples in this dataset can describe the location of an avalanche event from three perspectives: point, line, and surface. It also describes the avalanche event process from five stages: gestation, disaster initiation, disaster, damage, and rescue. Furthermore, it provides the correlation between the avalanche event location information and the avalanche disaster information, thus fully and effectively describing avalanche disasters. This allows the training samples to provide sufficient and effective training samples for the artificial intelligence model to achieve accurate avalanche disaster prediction, which helps in formulating corresponding control measures and thereby reducing the damage caused by avalanche disasters.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for constructing an artificial intelligence training dataset for avalanche disaster prediction, comprising: acquiring avalanche event location information; the avalanche event location information includes location data of each avalanche event among multiple avalanche events; the location data includes the avalanche center, avalanche path, and avalanche slope. Acquiring avalanche disaster information; the avalanche disaster information includes disaster-related data of each avalanche event among multiple avalanche events; the disaster-related data includes disaster-inducing data, disaster-causing data, disaster data, disaster-bearing data, and rescue data. Using the avalanche event location information and avalanche disaster information, constructing a training dataset; the training dataset includes multiple training samples; the training samples include location data, disaster-related data, and data correlation relationships of one avalanche event among multiple avalanche events; the data correlation relationships indicate the correlation between the avalanche center, avalanche path, and avalanche slope and the disaster-related data, respectively.

[0006] In one implementation of the first aspect, obtaining avalanche event location information includes: for each of the plurality of avalanche events, obtaining event information for the avalanche event; the event information includes disaster data, rescue data, and fuzzy location information. Using a large language model, the fuzzy location information in the event information is converted into geodetic coordinates, and these geodetic coordinates are used as the avalanche center of the avalanche event. The geodetic coordinate range of the road where the avalanche center is located is used as the avalanche section of the avalanche event. Slope units of the slope where the avalanche center is located are extracted from the geographic information system as the avalanche slope surface of the avalanche event.

[0007] In one implementation of the first aspect, a large language model is used to convert fuzzy location information in event information into geodetic coordinates. This includes: inputting event information and coordinate conversion prompts into the large language model; and, guided by the coordinate conversion prompts, the large language model infers and calculates the geodetic coordinates of the fuzzy location information by combining disaster data and rescue data, and then outputs the results.

[0008] In one implementation of the first aspect, the slope unit is enclosed by a forward catchment area and a reverse catchment area; the boundary of the forward catchment area is defined by the ridgeline of the original topography of the slope; the boundary of the reverse catchment area is defined by the valleyline of the original topography of the slope.

[0009] In a second aspect, the present invention provides an application of the artificial intelligence training dataset constructed by the method described in the first aspect, wherein the application uses the training dataset to train an artificial intelligence model to predict avalanche disasters.

[0010] In one implementation of the second aspect, the aforementioned artificial intelligence model is a discriminative model, a generative model, or a multimodal model.

[0011] Thirdly, the present invention provides an artificial intelligence training dataset construction device for avalanche disaster prediction, comprising a first acquisition module, a second acquisition module, and a construction module. The first acquisition module is used to acquire avalanche event location information; the avalanche event location information includes location data for each of multiple avalanche events; the location data includes the avalanche center, avalanche path, and avalanche slope. The second acquisition module is used to acquire avalanche disaster information; the avalanche disaster information includes disaster-related data for each of multiple avalanche events; the disaster-related data includes disaster-inducing data, disaster-causing data, disaster data, disaster-bearing data, and rescue data. The construction module is used to construct a training dataset using the avalanche event location information and the avalanche disaster information; the training dataset includes multiple training samples; the training samples include location data, disaster-related data, and data correlation relationships for one avalanche event among multiple avalanche events; the data correlation relationships indicate the correlation between the avalanche center, avalanche path, and avalanche slope and the disaster-related data, respectively.

[0012] In one implementation of the first and third aspects, the disaster-prone data includes earthquake risk, fault distance, stratigraphic lithology, road distance, and topographic data. Disaster-causing data includes earthquake magnitude, air temperature, wind speed, surface temperature, land use type, snow depth, temperature gradient, unfrozen water content, location of weak layers, and ski resort markers. Disaster data includes avalanche occurrence time, highest temperature of the day, highest wind speed of the day, snow depth of the day, avalanche disaster coordinates, disaster area elevation, satellite image samples, avalanche perimeter, avalanche leading edge location, avalanche backwall location, avalanche throw distance, avalanche accumulation size, and avalanche casualties. Disaster-bearing data includes the resident population, number of buildings, transportation network, and geographical location of protective works within the first area affected by the avalanche event. Rescue data includes the number of hospitals, rescue teams, and material reserves within the second area affected by the avalanche event.

[0013] In one implementation of the first and third aspects, in the disaster-prone data: earthquake risk and stratigraphic lithology are both related to the avalanche center, avalanche path, and avalanche slope; fault distance is related to the avalanche center and avalanche path; and road distance is only related to the avalanche center. In the disaster-causing data: earthquake magnitude and land use type are both related to the avalanche center and avalanche path; air temperature, wind speed, surface temperature, snow depth, temperature gradient, and unfrozen water content are only related to the avalanche slope; the location of the weak layer is only related to the avalanche center; and ski resort signage is related to the avalanche center and avalanche slope. In the disaster data: avalanche occurrence time, highest temperature of the day, maximum wind speed of the day, snow depth of the day, satellite image samples, avalanche accumulation size, and avalanche casualties are all related to the avalanche center, avalanche path, and avalanche slope; avalanche disaster coordinates and disaster area elevation are both related to the avalanche center and avalanche path; and the location of the avalanche backwall is only related to the avalanche slope. In the disaster impact data: the resident population, the number of buildings, and the transportation network are all related to the avalanche center and the avalanche slope; the geographical location of protective works is only related to the avalanche slope. In the rescue data: the number of hospitals, the number of rescue teams, and the amount of supplies are all related to the avalanche center and the avalanche route.

[0014] Fourthly, the present invention provides an electronic device including a processor and a memory coupled to the processor; the memory is used to store computer instructions, and when the electronic device is running, the processor executes the computer instructions stored in the memory to cause the electronic device to perform the method as described in the first aspect above or any implementation thereof.

[0015] Fifthly, the present invention provides a computer-readable storage medium including computer program instructions that, when executed by a computer, cause the computer to perform the method described in the first aspect above or any implementation thereof.

[0016] In a sixth aspect, the present invention provides a computer program product comprising computer program instructions that, when executed on a computer, cause the computer to perform the method described in the first aspect above or any implementation thereof.

[0017] The technical effects of the third to sixth aspects and their possible implementations can be referred to the above description of the technical effects of the first aspect and its possible implementations, and will not be repeated here.

[0018] Compared with the prior art, the present invention has the following beneficial effects.

[0019] In the method for constructing an artificial intelligence training dataset for avalanche disaster prediction provided by the present invention, after obtaining avalanche event location information and avalanche disaster information, the above-mentioned avalanche event location information, avalanche disaster information and data correlation are used to construct a training dataset for training an artificial intelligence model to predict avalanche disasters. The aforementioned avalanche location information includes location data for each of multiple avalanche events. The avalanche center, from a point perspective, identifies the location of the avalanche; the avalanche road segment, from a line perspective, identifies the road segment affected by the avalanche; and the avalanche slope, from a surface perspective, identifies the slope affected by the avalanche. The aforementioned avalanche disaster information describes the avalanche disaster process from five stages: gestation, occurrence, damage, and rescue. The aforementioned data correlations indicate the associations between the avalanche center, avalanche road segment, and avalanche slope with disaster-related data. This ensures that the training dataset not only effectively describes avalanche disasters but also guides the data correlation between avalanche location information and avalanche disaster information. This provides sufficient and effective training samples for artificial intelligence models to achieve accurate avalanche disaster prediction, contributing to the development of corresponding control measures and ultimately reducing the damage caused by avalanches. Attached Figure Description

[0020] Figure 1 This is one of the schematic diagrams of the artificial intelligence training dataset construction method for avalanche disaster prediction provided in the embodiments of this application; Figure 2 This is the second schematic diagram of the artificial intelligence training dataset construction method for avalanche disaster prediction provided in the embodiments of this application; Figure 3 This is the third schematic diagram of the method for constructing an artificial intelligence training dataset for avalanche disaster prediction provided in the embodiments of this application; Figure 4 This is a schematic diagram of the structure of the artificial intelligence training dataset construction device for avalanche disaster prediction provided in the embodiments of this application. Detailed Implementation

[0021] In the specification and claims of this invention, the terms "first" and "second," etc., are used to distinguish different objects, rather than to describe a specific order of objects.

[0022] In the embodiments of this application, "and / or" indicates a relationship between objects. For example, A and / or B can represent the following three situations: A exists alone, B exists alone, and A and B exist simultaneously.

[0023] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0024] In the description of this invention, unless otherwise stated, "multiple" means two or more. For example, multiple avalanche events refer to two or more avalanche events.

[0025] The methods and apparatus provided in this application relate to avalanche disaster prediction, and can be used to construct training datasets for training artificial intelligence models to predict avalanche disasters.

[0026] To address the problem in the background art where the scarcity of avalanche events and the lack of effective information related to them result in insufficient training samples for artificial intelligence models predicting avalanche disasters, thus hindering their ability to predict avalanche disasters and reduce the damage caused by them, this application provides a method and apparatus for constructing an artificial intelligence training dataset for avalanche disaster prediction. The training dataset is constructed using avalanche event location information, avalanche disaster information, and data correlations. The training samples in this dataset can describe the location of avalanche events from three perspectives: points, lines, and surfaces. They can also describe the avalanche event occurrence process from five stages: gestation, occurrence, disaster, damage, and rescue. Furthermore, the correlation between the avalanche event location information and the avalanche disaster information is provided, enabling a sufficient and effective description of avalanche disasters. This allows the training samples to provide sufficient effective training samples for artificial intelligence models to accurately predict avalanche disasters, facilitating the development of corresponding control measures and ultimately reducing the damage caused by avalanche disasters.

[0027] For example, the method for constructing an artificial intelligence training dataset for avalanche disaster prediction provided in this embodiment of the invention can be executed by an electronic device with processing capabilities, such as a computer or server. Taking a computer as an example, the hardware components of the computer may include: a processor, memory, a network interface, a user interface, a communication bus, etc.

[0028] The processor is used to control electronic devices to perform related processing and computing tasks, such as acquiring avalanche event location information, acquiring avalanche disaster information, and building training datasets. The processor may include a central processing unit (CPU), an AI processing unit (such as a GPU, TPU, NPU, LPU, etc.), or other processors. The processor can be single-core or multi-core; for example, the processor may include multiple CPUs or multiple GPUs.

[0029] Memory is used to store computer instructions and related data, such as avalanche event location information, avalanche disaster information, and training datasets. Memory can be random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical storage, disk storage media, or other magnetic storage devices, or any other medium capable of storing program code or data accessible by a computer. Optionally, memory can be integrated into the processor, or it can be independent of the processor.

[0030] A network interface is used for communication between a computer and other devices or communication networks. A network interface can be a transceiver with transmit and receive capabilities. Optionally, a network interface may include standard wired interfaces or wireless interfaces (such as Wi-Fi interfaces, Bluetooth interfaces, and 5G interfaces).

[0031] The communication bus is used to enable communication between different components. For example, the processor, memory, network interface and user interface mentioned above can be interconnected through the communication bus.

[0032] The user interface may include a display screen and an input unit (such as a keyboard). Optionally, the user interface may also include a standard wired interface or a wireless interface.

[0033] Those skilled in the art will understand that the computer described above may include more or fewer components, or combine certain components, or have different component arrangements; the embodiments of this application do not limit this.

[0034] like Figure 1 As shown, the method for constructing an artificial intelligence training dataset for avalanche disaster prediction provided in this application includes S1-S3.

[0035] S1. Obtain the location information of the avalanche event.

[0036] The aforementioned avalanche location information may include the location data of each of multiple avalanche events. This location data may include the avalanche center, avalanche path, and avalanche slope.

[0037] In one implementation, combined with Figure 1 ,like Figure 2 As shown, S1 above includes S101-S104.

[0038] S101. For each avalanche event among multiple avalanche events, obtain the event information of the avalanche event.

[0039] In this embodiment of the application, the aforementioned event information may include disaster data, rescue data, and fuzzy location information, etc.

[0040] Optionally, combined Figure 2 ,like Figure 3 As shown, S101 includes S1011-S1012.

[0041] S1011. Using natural language processing technology, extract multiple initial pieces of information from text, audio, and / or video related to the avalanche event.

[0042] Each of the aforementioned initial pieces of information may include the time of the avalanche, disaster losses, casualties, and rescue status. It is understood that the use of Natural Language Processing (NLP) technology to extract information is a common technique in this field, and will not be elaborated upon further in this application.

[0043] In one application scenario, a text message related to an avalanche event reads: "At approximately 11:00 AM Beijing time on January 26, an avalanche occurred at 'Dongwozi,' located 15 kilometers northeast of the intersection of National Highways 217 and 218 in Nalati Town. Houses and livestock sheds were completely buried, 8 people were trapped, and more than 40 large livestock and more than 200 small livestock were buried." The location extracted from this text message is "Dongwozi," located 15 kilometers northeast of the intersection of National Highways 217 and 218 in Nalati Town. The time of the event was approximately 11:00 AM Beijing time on January 26. The disaster losses include the complete burial of houses and livestock sheds, the burial of more than 40 large livestock and more than 200 small livestock, and the casualties including the 8 people trapped.

[0044] S1012. Cross-validate and unify multiple initial pieces of information to obtain the event information of the avalanche event.

[0045] Taking an avalanche event as an example, when the avalanche event contains the following two initial pieces of information, the event information of the avalanche event is obtained by cross-validating the two initial pieces of information in three dimensions: time, space, and event.

[0046] 1. Location: Construction site of the West-East Gas Pipeline II project in Dadonggou, Huocheng County, Xinjiang; Date of occurrence: March 13, 2008; Disaster loss: Avalanche buried approximately 8 kilometers of road; Casualties: Approximately 20 people were trapped or buried; Source: Department of Transportation of Xinjiang Uygur Autonomous Region.

[0047] 2. Location: Exit section of the Guozigou Tunnel of the West-East Gas Pipeline in Huocheng County, Yili Prefecture; Time of occurrence: Approximately 10:20 AM on March 13, 2008; Casualties: 8 dead, 6 injured, 8 missing; Source: Emergency Management Department of Xinjiang Uygur Autonomous Region.

[0048] The cross-validation process for the above times includes: I. Natural date verification: The occurrence time of both initial information entries is March 13, 2008, and the dates are completely consistent; II. Detailed logic verification: The Emergency Management Department clearly states that 10:20 is the precise time of the incident, while the Transportation Department only marks the day of the incident without specifying the exact time, so there is no date contradiction or cross-day conflict; III. Conclusion: The occurrence time of the two initial information entries is the same day, and they are determined to be the same avalanche.

[0049] The cross-verification process for the above-mentioned space includes: I. Administrative location verification: The location in the first initial information is Dadonggou in Huocheng County, and the location in the second initial information is the exit of Guozigou Tunnel in Huocheng County. Both belong to the Guozigou area of ​​Huocheng County, Yili, and their administrative regions are the same; II. Coordinate verification: The geodetic coordinates of Dadonggou in Huocheng County and the exit of Guozigou Tunnel in Huocheng County are the same, so it is determined that the locations in the two initial information are accurate.

[0050] The cross-verification process for the above events includes: I. Disaster type verification: Both the first and second initial information records the disaster type as an avalanche event; II. Casualty verification: The casualties in the first initial information are approximately 20 people trapped or buried, which matches the 22 people reported in the second initial information: 8 deaths, 6 injuries, and 8 missing. Therefore, the casualties in the first and second initial information are accurate.

[0051] After both initial pieces of information were cross-validated, the more accurate data was selected as the event information for the avalanche disaster. The fuzzy location information in the event information was the exit section of the Guozigou Tunnel of the West-East Gas Pipeline in Huocheng County, Yili, and the time of occurrence was around 10:20 on March 13, 2008. The disaster loss was that the avalanche buried about 8 kilometers of road, and the casualties were 8 deaths, 6 injuries, and 8 missing persons.

[0052] S102. Using a large language model, the fuzzy location information in the event information is converted into geodetic coordinates, and the geodetic coordinates are used as the avalanche center of the avalanche event.

[0053] Specifically, the event information and coordinate transformation prompts are input into the large language model. Guided by the coordinate transformation prompts, the large language model combines disaster data and rescue data to infer and output the geodetic coordinates of the fuzzy location information. The above geodetic coordinates are based on the geodetic coordinate system and use latitude, precision, and elevation to accurately describe the avalanche center.

[0054] In some embodiments, the coordinate transformation prompts may include single-point whole-kilometer marker prompts (e.g., RA-KM), single-point markers with 100-meter increments (e.g., RA-KM+m), markers for segments within the same kilometer (e.g., RA-KM+m to +m), and markers for segments spanning multiple kilometer segments (e.g., RA-KM+m to KM+m). Here, R is the road identification code, which can be replaced with G (national highway / national expressway), S (provincial highway / provincial expressway), X (county road), Y (township road), C (village road), or Z (dedicated road). A is the route number segment, a 1-4 digit code. K is the kilometer marker identification code. M is the kilometer marker base number, an arbitrary length string of numbers with no limit on the number of digits. m is the 100-meter offset value, a fixed 3-digit number ranging from 000 to 999, representing the number of meters after the kilometer.

[0055] In one application scenario, the fuzzy location information is K130+800-900 on National Highway 219. Guided by the coordinate transformation prompts mentioned above, the large language model first inverts the longitude and latitude corresponding to the road station K130+800-900 on National Highway 219, and then obtains the corresponding elevation value based on the aforementioned longitude and latitude, thus obtaining the geodetic coordinates (latitude, longitude, elevation). The plane coordinates (latitude, longitude) of the above geodetic coordinates are obtained based on the 2000 National Geodetic Coordinate System (CGCS2000), and the above elevation values ​​are determined using the 1985 National Elevation Datum.

[0056] S103. The geodetic coordinate range of the road where the avalanche center is located shall be taken as the avalanche section of the avalanche event.

[0057] Taking one application scenario in S102 above as an example, the process of determining the avalanche section of an avalanche event includes steps 1-4.

[0058] Step 1: Extract the disaster-affected road section markers related to the avalanche event, namely, National Highway 219 from K130+800 to K130+900. After obtaining the disaster-affected road section markers, use existing map services (such as Gaode / Baidu Maps) to extract the longitude and latitude corresponding to the road station numbers (such as K130+800) in the disaster-affected road section markers.

[0059] It should be understood that since the aforementioned Geographic Information System (GIS) is existing technology, this application embodiment will not further elaborate on the longitude and latitude corresponding to the road station number (such as K130+800) extracted from the disaster road section identifier in the geographic information system.

[0060] Step 2: Extract the longitude and latitude of the starting node K130+800 and the ending node K130+900 of the road corresponding to the road station number, and use them as the basic points at both ends of the road segment. Step 3: Along the road direction, from the starting node to the ending node, select multiple consecutive nodes as the spatial node set of the closed avalanche section affected by the avalanche.

[0061] Step 4: Connect all nodes in the above spatial node set in spatial topological order to form a linear avalanche road segment coordinate range, and take the avalanche road segment coordinate range associated with the above avalanche center as the avalanche road segment of the avalanche event.

[0062] Specifically, the geometric type of the coordinate range of the aforementioned avalanche section is a polyline, and the common file carrier formats are SHP, GeoJSON, and GDB. The coordinate strings support two storage formats: text and floating-point arrays. The storage method is to store the coordinate strings sequentially according to the node order.

[0063] S104. Extract the slope unit of the slope where the avalanche center is located from the geographic information system, and use it as the avalanche slope surface of the avalanche event.

[0064] Understandably, the aforementioned slope unit is enclosed by a forward catchment area and a reverse catchment area; the boundary of the forward catchment area is defined by the ridgeline of the original topography where the slope is located; the boundary of the reverse catchment area is defined by the valleyline of the original topography where the slope is located. The ridgeline is the line connecting the highest points on the ridge surface; the valleyline is the line connecting the lowest points on the valley surface.

[0065] The construction process of the above-mentioned slope element is given below, including steps 1-3.

[0066] Step 1: Obtain the first and second topographic data of the slope.

[0067] The first topographic data of the aforementioned slope is the original elevation value raster of the slope (hereinafter referred to as the forward DEM raster). The forward DEM raster includes multiple first rasters, and each of the multiple first rasters corresponds to the original DEM of a certain geographical location within the slope. The size of this geographical location is consistent with the resolution of the raster. For example, when the resolution of the raster is 30m × 30m, the size of the geographical location corresponding to the first raster can also be 30m × 30m.

[0068] The second topographic data of the aforementioned slope is a reverse DEM raster of the slope. This reverse DEM raster comprises multiple second rasters, each of which corresponds to the reverse DEM of a specific geographic location within the slope. For each of these second rasters, the reverse DEM is obtained by subtracting the original DEM of the corresponding geographic location from the maximum elevation value of the slope.

[0069] Step 2: Perform watershed extraction on the forward topographic data of the above slope to obtain the forward catchment watershed; perform watershed extraction on the reverse topographic data of the above slope to obtain the reverse catchment watershed.

[0070] For example, the specific process of the above watershed extraction operation is as follows.

[0071] Step 2.1: Perform depression filling preprocessing on the forward and reverse topographic data of the slope respectively to obtain the preprocessed forward DEM raster and the preprocessed reverse DEM raster.

[0072] Since the above-mentioned depression filling pretreatment is a commonly used technical means in this field, the embodiments of this application will not elaborate on the above-mentioned depression filling pretreatment process.

[0073] Step 2.2: Calculate the cumulative flow of each first cell in the preprocessed forward DEM grid using the D8 unidirectional flow algorithm to obtain the first cumulative flow grid; calculate the cumulative flow of each second cell in the preprocessed reverse DEM grid using the unidirectional flow algorithm to obtain the second cumulative flow grid.

[0074] It should be understood that the aforementioned D8 unidirectional flow algorithm is existing technology. This algorithm generates a flow direction raster map by traversing each cell in the DEM grid, comparing the size difference between the DEM of that cell and the DEMs of its eight surrounding cells, and taking the direction of the steepest slope descent as the water outlet direction for that cell. Following the principle that water flows downwards from higher terrain levels, the algorithm calculates the total number of upstream catchment cells connected to each cell in the flow direction raster map, resulting in a flow accumulation raster. Each cell in the flow accumulation raster corresponds to its upstream catchment area, which is the product of the total number of upstream catchment cells and the resolution.

[0075] Step 2.3: Binarize the first confluence accumulation raster to obtain the first binarized river network raster; binarize the second confluence accumulation raster to obtain the second binarized river network raster.

[0076] The binarization process described above is as follows: for each cell in the runoff accumulation grid, if the upstream catchment area of ​​the cell is greater than or equal to a catchment area threshold (e.g., 1 square kilometer), the cell is assigned a value of 1; otherwise, it is assigned a value of 0. It is understood that the catchment area threshold can also be set according to the geomorphological characteristics of the study area and the scale of the avalanche slope. This embodiment of the application does not limit the value of the catchment area threshold.

[0077] Step 2.4: The first binary river network grid and the second binary river network grid are classified into levels using the Strahler classification system respectively. The continuous river network grid is then cut into several independent river segment connection lines using the river network linking tool. Each river segment connection line corresponds to a unique outlet point.

[0078] It should be noted that, since the Strahler classification system is existing technology, the specific details of the classification of the first binary river network grid and the first binary river network grid using the Strahler classification system in this application embodiment will not be elaborated.

[0079] Step 2.5: Input the outlet points of each river segment connecting line contained in the first binarized river network raster, the above-mentioned forward DEM raster, and the ridgeline of the original topography of the slope into the watershed extraction tool (ArcGIS). The watershed extraction tool outputs the forward catchment area. Input the outlet points of each river segment connecting line contained in the second binarized river network raster, the above-mentioned reverse DEM raster, and the valley line of the original topography of the slope into the watershed extraction tool. The watershed extraction tool outputs the reverse catchment area.

[0080] It is understood that the above-mentioned watershed extraction tool is existing technology, and the process of extracting the catchment area using the above-mentioned watershed extraction tool will not be described in detail in the embodiments of this application.

[0081] Furthermore, the raster-to-vector tool can be used to convert the above-mentioned catchment area into a vector polygon feature, and the simplified polygon option can be checked to eliminate jagged boundaries, generating a smooth closed polygon, thus obtaining a forward catchment area in vector polygon form and a reverse catchment area in vector polygon form. The boundary of the forward catchment area in vector polygon form is the ridgeline of the original terrain of the hillside, and the boundary of the reverse catchment area in vector polygon form is the valleyline of the original terrain of the hillside.

[0082] Step 3: Intersect the above-mentioned forward and reverse watersheds to obtain the slope unit.

[0083] It should be noted that the aforementioned geographic information system is existing technology, and the aforementioned extraction of slope units of the slope where the avalanche center is located from the geographic information system is a commonly used technical means in this field. The embodiments of this application will not elaborate on the above process.

[0084] Preferably, the avalanche slope of the aforementioned avalanche event can be the aforementioned slope unit that has been processed by a raster vectorization algorithm and is associated with slope topographic parameters.

[0085] Specifically, a raster vectorization algorithm is used to convert the aforementioned slope units into continuous planar geometric contours, and the slope boundary nodes are optimized based on contour lines and terrain abrupt change points to restore the natural mountain slope morphology. Then, the aforementioned slope units are associated with the avalanche center and avalanche path, and the slope topographic parameters of the aforementioned slope units are supplemented to obtain the avalanche slope of the aforementioned avalanche event.

[0086] S2. Obtain avalanche disaster information.

[0087] The avalanche disaster information mentioned above can be obtained by retrieving the location data of each avalanche event in the avalanche event location information, or by directly retrieving the accident name of the avalanche event (such as the 2010 Sarang Pass avalanche in Afghanistan). This application embodiment does not limit the method of obtaining the avalanche disaster information mentioned above.

[0088] The aforementioned avalanche disaster information may include disaster-related data for each of the multiple avalanche events mentioned above. This disaster-related data may include data on the avalanche's initiation, the factors causing the avalanche, the disaster data, the damage-bearing data, and the rescue data. The initiation data indicates the indirect environmental factors leading to the avalanche event; the factors causing the avalanche indicate the direct triggering factors; the disaster data shows the environmental factors and damage at the time of the avalanche; the damage-bearing data shows the surrounding conditions at the time of the avalanche; and the rescue data shows the rescue efforts following the avalanche.

[0089] In a real-world scenario, the aforementioned disaster-prone data refers to the set of basic natural environmental data that contributes to avalanche disasters. This data reflects long-term stable environmental characteristics such as topography, geology, vegetation, and meteorological baselines. Therefore, the aforementioned disaster-prone data may include earthquake risk, fault distance, stratigraphic lithology, road distance, and topographic data.

[0090] The aforementioned disaster-causing data refers to dynamic driving factors that directly trigger and exacerbate avalanches. These are short-term, variable inducing conditions and are the direct causes of avalanche events. This disaster-causing data may include earthquake magnitude, air temperature, wind speed, surface temperature, land use type, snow depth, temperature gradient, unfrozen water content, location of weak layers, and ski resort markers. The ski resort markers can have values ​​of 0 or 1, where 0 indicates the absence of a ski resort on the slope and 1 indicates the presence of a ski resort. It should be understood that the ski resort markers treat the ski resort as a disaster-causing location (for example, human skiing behavior can trigger avalanches) to describe the dynamic driving factors that directly trigger and exacerbate avalanche events.

[0091] The aforementioned disaster data records thematic vector / raster data of the avalanche event itself, including its spatial location, morphology, intensity, and extent. It focuses on the intrinsic characteristics of the disaster, corresponding to avalanches that have occurred or are currently occurring. This disaster data may include avalanche occurrence time, highest temperature of the day, maximum wind speed of the day, snow depth of the day, avalanche disaster coordinates, elevation of the disaster area, satellite image samples, avalanche perimeter, avalanche leading edge location, avalanche back wall location, avalanche throw distance, avalanche accumulation size, and the number of avalanche casualties.

[0092] The aforementioned disaster-bearing data refers to data on various human activity carriers and infrastructure threatened or affected by avalanches, i.e., the objects affected by the disaster, used to assess the scope of disaster losses and impacts. This disaster-bearing data may include the population, number of buildings, transportation network, and geographical location of protective works within the first avalanche event radius (e.g., 5km). The aforementioned transportation network consists of multiple roads, and the geographical location of protective works includes basic geographic data related to the protective works.

[0093] The aforementioned rescue data serves as specialized operational data for emergency response and rescue operations following avalanches, and is structured around emergency actions, resource allocation, and on-site support. This data may include the number of hospitals, rescue teams, and material reserves within a second radius (e.g., 50km) of the avalanche event.

[0094] S3. Use avalanche event location information and avalanche disaster information to construct a training dataset.

[0095] In this embodiment of the application, the training dataset is used to train an artificial intelligence model to predict avalanche disasters.

[0096] Specifically, the aforementioned training dataset includes multiple training samples. These training samples may include location data, disaster-related data, and data correlations for one avalanche event from multiple avalanche events. The aforementioned data correlations indicate the association between the avalanche center, avalanche path, and avalanche slope with the disaster-related data, respectively.

[0097] Taking the actual scenario in S2 above as an example, the correlation between the avalanche center, avalanche section, and avalanche slope and the disaster-prone data in the actual scenario is shown in Table 1 below, where Y indicates correlation and N indicates no correlation.

[0098] Table 1. Correlation between avalanche center, avalanche route, avalanche slope, and avalanche risk data.

[0099] As shown in Table 1 above, among the disaster-prone data: earthquake risk and strata lithology are related to the avalanche center, avalanche route, and avalanche slope; fault distance is only related to the avalanche route; and road distance is only related to the avalanche center. Among the topographic data: elevation is only related to the avalanche center; slope, aspect, curvature, topographic location index, hydrodynamic index, topographic humidity index, drainage density, ground roughness, surface dissection, and vegetation cover are only related to the avalanche slope; stream distance is related to the avalanche center and avalanche route; shady slopes and earthquake hazard levels are related to the avalanche center, avalanche route, and avalanche slope; vegetation type is related to the avalanche center and avalanche slope; and tree age is related to the avalanche center and avalanche slope.

[0100] The correlations between the avalanche center, avalanche section, and avalanche slope and the disaster data in the above-mentioned actual scenarios are shown in Table 2 below.

[0101] Table 2. Correlation between avalanche center, avalanche route, avalanche slope, and disaster data.

[0102] As shown in Table 2 above, among the disaster data: earthquake magnitude is related to the avalanche center and avalanche section; air temperature, wind speed, surface temperature, snow depth, temperature gradient, and unfrozen water content are only related to the avalanche slope; land use type, location of weak layer, and ski resort signage are related to the avalanche center and avalanche slope.

[0103] The correlations between the avalanche center, avalanche section, and avalanche slope and the disaster data in the above-mentioned actual scenarios are shown in Table 3 below.

[0104] Table 3. Correlation between avalanche center, avalanche route, avalanche slope, and disaster data.

[0105] As shown in Table 3 above, in the disaster data: the time of avalanche occurrence, the highest temperature of the day, the maximum wind force of the day, the snow depth of the day, satellite image samples, the size of the avalanche accumulation, and the number of avalanche casualties are all related to the avalanche center, the avalanche route, and the avalanche slope. The coordinates of the avalanche disaster and the altitude of the disaster area are both related to the avalanche center and the avalanche route. The location of the avalanche back wall is only related to the avalanche slope.

[0106] The correlations between the avalanche center, avalanche section, and avalanche slope and the disaster-bearing data in the above-mentioned actual scenarios are shown in Table 4 below.

[0107] Table 4. Correlation between avalanche center, avalanche route, avalanche slope, and disaster bearing data.

[0108] As shown in Table 4 above, among the disaster-bearing data: the resident population, the number of buildings, and the transportation network are all related to the avalanche center and the avalanche slope; the geographical location of the protective works is only related to the avalanche slope.

[0109] The correlations between the avalanche center, avalanche section, and avalanche slope and the rescue data in the above-mentioned actual scenarios are shown in Table 5 below.

[0110] Table 5. Correlation between avalanche center, avalanche route, avalanche slope, and rescue data.

[0111] As shown in Table 5 above, the number of hospitals, rescue teams, and supplies in the rescue data are all related to the avalanche center and avalanche sections.

[0112] In some embodiments, the training dataset may also include multiple negative training samples. The negative training samples may be location information and avalanche disaster information of other locations in the same area as the location where the avalanche event occurred (i.e., geographical locations where no avalanche event occurred). The process of obtaining the negative training samples can refer to the relevant content in S1 and S2 above, and will not be repeated here in the embodiments of this application.

[0113] This application also provides an application of the artificial intelligence training dataset constructed by the above method. The application uses the training dataset to train an artificial intelligence model to predict avalanche disasters.

[0114] Furthermore, the aforementioned artificial intelligence models can be categorized according to their modeling methods into discriminative models, generative models, etc., or according to their input and output data types into unimodal models, multimodal models, etc. The discriminative models can be Support Vector Machines (SVM), Decision Trees, or Random Forests, etc.; the generative models can be Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), Diffusion Models, or Large Language Models (LLMs), etc.; and the multimodal models can include models capable of processing multiple modalities of data such as text, images, remote sensing images, structured numerical data, audio, or video.

[0115] When the above artificial intelligence model is a discriminative model, the process of training the above artificial intelligence model with the training dataset obtained in S3 to predict avalanche disasters is as follows.

[0116] This training dataset utilizes supervised learning algorithms (such as random forests, support vector machines, and decision trees) to perform supervised learning on the location data, disaster-related data, and data correlations of avalanche events for each training sample, as well as on the negative training samples corresponding to events where no avalanches occurred. The avalanche occurrence status (occurred or not) and disaster risk level are used as training labels to iteratively optimize hyperparameters, resulting in multiple decision trees. After training, these decision trees can output binary classification results to determine whether an avalanche has occurred in the target area, or multi-class classification results to determine the avalanche risk level and impact range level, incorporating disaster impact and rescue data to supplement the classification of disaster severity and rescue difficulty. Therefore, in actual prediction, this AI model only needs to input real-time location data and environmental monitoring data of the target area to quickly obtain the probability of avalanche occurrence and the overall risk level, achieving intelligent prediction of avalanche disasters.

[0117] When the above artificial intelligence model is a generative model, the process of training the above artificial intelligence model with the training dataset obtained in S3 to predict avalanche disasters is as follows.

[0118] For generative models, based on the entire process from disaster initiation, disaster, disaster outbreak, damage to post-disaster relief, the location data, disaster-related data, and data correlation of each training sample in the training dataset can be integrated with the sample information of multiple historical avalanche events (such as related news reports) to construct an avalanche sample distribution library.

[0119] The generative model is then trained using the aforementioned avalanche sample distribution database. Specifically, the generative model uses real-time spatial data and dynamic disaster-causing environmental data of the target area as basic inputs, supplemented by historical samples with similar terrain and climate conditions and data correlation logic as auxiliary inputs. It learns the spatiotemporal evolution patterns of avalanche disasters, data distribution characteristics, and the intrinsic correlation between spatial and disaster data. During training, real avalanche process information is used as the ground truth. By minimizing the difference between the generated results and the actual disaster process data, the model parameters are optimized, and the model structure or hyperparameters are adjusted based on the validation results. After training, the model can output a complete avalanche evolution scenario over a future period, specifically covering core information such as avalanche occurrence time, movement path, spatial impact range, and accumulation scale. It can also simulate changes in disaster-causing factors, evolution of affected areas, damage to facilities and personnel, and identify potential disaster hazard points and disaster spread trends within the area. In actual prediction scenarios, by inputting real-time monitoring data and spatial information of the target area, the model can autonomously generate panoramic avalanche evolution simulation results, improving the accuracy and completeness of avalanche disaster key indicator predictions.

[0120] When the above artificial intelligence model is a multimodal model, the process of training the above artificial intelligence model with the training dataset obtained in S3 to predict avalanche disasters is as follows.

[0121] For multimodal models, such as multimodal base models, visual-language models, or fine-tuned multimodal disaster prediction models, information can be extracted from multi-modal data. Therefore, based on each training sample in the training dataset, we can associate image remote sensing data (such as satellite imagery, slope real-scene maps, and topographic raster maps), structured numerical data (such as location coordinates, meteorological and geological data, and population resources), and spatial elements with various disaster data. Before training, different modal data are preprocessed separately: image remote sensing data is cropped, denoised, and feature-extracted; text data is segmented, semantically encoded, and vectorized; and structured numerical data is standardized and dimensionally aligned. Then, cross-modal feature alignment and fusion are completed by combining data association rules to establish the mapping relationship between image, text, and numerical data.

[0122] Then, using a multi-channel parallel input approach, the aforementioned structured numerical data, image data, and regional text data are fed into the corresponding channels of the multimodal model to complete overall feature fusion. End-to-end training is performed using avalanche prediction results, disaster descriptions, and rescue plans as comprehensive labels, allowing the model to grasp the linkage logic between different modalities of data. After iterative optimization on the test set, the model possesses multi-dimensional output capabilities. On one hand, it outputs basic prediction results such as avalanche occurrence probability, risk level, occurrence time, and impact range; on the other hand, it outputs visualized images annotating potentially affected areas. Simultaneously, it combines data analysis to generate textual decision-making content such as risk interpretation, personnel evacuation suggestions, rescue route planning, and emergency material allocation. Therefore, when conducting comprehensive early warning and prediction work, only remote sensing images of the target area, real-time monitoring values, and regional archive text need to be input; the model can then integrate multi-source information to simultaneously complete avalanche risk prediction, disaster visualization, and emergency rescue decision output.

[0123] Optionally, the aforementioned training dataset can also be used to train artificial intelligence models to provide early warnings of avalanche events or to make avalanche rescue decisions.

[0124] In summary, the method for constructing an artificial intelligence training dataset for avalanche disaster prediction provided in this application embodiment involves, after obtaining avalanche event location information and avalanche disaster information, constructing a training dataset for training an artificial intelligence model to predict avalanche disasters using the aforementioned avalanche event location information, avalanche disaster information, and data correlation relationships. The aforementioned avalanche location information includes location data for each of multiple avalanche events. The avalanche center, from a point perspective, identifies the location of the avalanche; the avalanche road segment, from a line perspective, identifies the road segment affected by the avalanche; and the avalanche slope, from a surface perspective, identifies the slope affected by the avalanche. The aforementioned avalanche disaster information describes the avalanche disaster process from five stages: gestation, occurrence, damage, and rescue. The aforementioned data correlations indicate the associations between the avalanche center, avalanche road segment, and avalanche slope with disaster-related data. This ensures that the training dataset not only effectively describes avalanche disasters but also guides the data correlation between avalanche location information and avalanche disaster information. This provides sufficient and effective training samples for artificial intelligence models to achieve accurate avalanche disaster prediction, contributing to the development of corresponding control measures and ultimately reducing the damage caused by avalanches.

[0125] Accordingly, embodiments of this application provide an apparatus for constructing an artificial intelligence training dataset for avalanche disaster prediction, such as... Figure 4 As shown, it includes a first acquisition module 501, a second acquisition module 502, and a construction module 503.

[0126] The first acquisition module 501 is used to acquire avalanche disaster information; the avalanche disaster information includes disaster-related data for each of multiple avalanche events; the disaster-related data includes disaster-prone data, disaster-causing data, disaster data, disaster-bearing data, and rescue data. For example, the first acquisition module 501 is used to implement S1 of the above method.

[0127] The second acquisition module 502 is used to acquire avalanche disaster information; the avalanche disaster information includes disaster-related data for each of multiple avalanche events; the disaster-related data includes disaster-prone data, disaster-causing data, disaster data, disaster-bearing data, and rescue data. For example, the second acquisition module 502 is used to implement S2 of the above method.

[0128] The construction module 503 is used to construct a training dataset using avalanche event location information and avalanche disaster information. The training dataset is used to train an artificial intelligence model to predict avalanche disasters. The training dataset includes multiple training samples. Each training sample includes location data, disaster-related data, and data correlations for one avalanche event from multiple avalanche events. The data correlations indicate the relationship between the avalanche center, avalanche path, and avalanche slope and the disaster-related data, respectively. For example, the construction module 503 is used to implement step S3 of the above method.

[0129] Optionally, the first acquisition module 501 is specifically used for: acquiring event information for each of the multiple avalanche events; the event information includes disaster data, rescue data, and fuzzy location information. Using a large language model, the fuzzy location information in the event information is converted into geodetic coordinates, and these geodetic coordinates are used as the avalanche center of the avalanche event. The geodetic coordinate range of the road where the avalanche center is located is used as the avalanche road segment of the avalanche event. The slope units of the slope where the avalanche center is located are extracted from the geographic information system as the avalanche slope surface of the avalanche event. For example, the first acquisition module 501 is specifically used to implement steps S101-S104 of the above method.

[0130] The modules of the artificial intelligence training dataset construction device for avalanche disaster prediction described above can also be used to perform other steps in the above method embodiments. All relevant content involved in the above method embodiments can be referred to in the functional description of the corresponding functional module, and will not be repeated here.

[0131] This application also provides an electronic device, including: a processor and a memory coupled to the processor; the memory is used to store computer instructions, and when the electronic device is running, the processor executes the computer instructions stored in the memory to cause the electronic device to perform the methods described in the above embodiments. The processor can implement the first acquisition module 501, the second acquisition module 502, and the construction module 503 described above; the memory can also be used to store avalanche event location information, avalanche disaster information, and training datasets, etc.

[0132] This application also provides a computer-readable storage medium including a computer program that, when run on a computer, performs the methods described in the above embodiments.

[0133] This application also provides a computer program product, which includes computer program instructions that, when run on a computer, execute the methods described in the above embodiments.

[0134] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0135] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for constructing an artificial intelligence training dataset for avalanche disaster prediction, characterized in that, include: Obtain the location information of the avalanche event; The avalanche location information includes the location data of each avalanche event in multiple avalanche events; The location data includes the avalanche center, avalanche path, and avalanche slope. Obtain avalanche disaster information; the avalanche disaster information includes disaster-related data for each of the multiple avalanche events; the disaster-related data includes disaster-prone data, disaster-causing data, disaster data, disaster-bearing data, and rescue data; A training dataset is constructed using the avalanche event location information and avalanche disaster information. The training dataset includes multiple training samples; the training samples include the location data, disaster-related data, and data correlations of one of the multiple avalanche events; the data correlations indicate the association between the avalanche center, avalanche section, and avalanche slope and the disaster-related data, respectively.

2. The method as described in claim 1, characterized in that, The acquisition of avalanche event location information includes: For each of the multiple avalanche events, obtain the event information for that avalanche event; the event information includes disaster data, rescue data, and fuzzy location information; Using a large language model, the fuzzy location information in the event information is converted into geodetic coordinates, and the geodetic coordinates are used as the avalanche center of the avalanche event; The geodetic coordinate range of the road where the avalanche center is located is taken as the avalanche section of the avalanche event; The slope units of the slope where the avalanche center is located are extracted from the geographic information system and used as the avalanche slope surface of the avalanche event.

3. The method as described in claim 2, characterized in that, The step of using a large language model to convert the fuzzy location information in the event information into geodetic coordinates includes: The event information and coordinate transformation prompts are input into a large language model. Guided by the coordinate transformation prompts, the large language model combines the disaster data and the rescue data to deduce and output the geodetic coordinates of the fuzzy location information.

4. The method as described in claim 2, characterized in that, The slope unit is enclosed by a forward catchment area and a reverse catchment area; the boundary of the forward catchment area is defined by the ridgeline of the original topography of the slope; the boundary of the reverse catchment area is defined by the valleyline of the original topography of the slope.

5. An application of an artificial intelligence training dataset constructed based on the method described in any one of claims 1 to 4, characterized in that, The training dataset was used to train an artificial intelligence model to predict avalanche disasters.

6. The application as described in claim 5, characterized in that, The artificial intelligence model can be a discriminative model, a generative model, or a multimodal model.

7. An apparatus for constructing an artificial intelligence training dataset for avalanche disaster prediction, characterized in that, It includes a first acquisition module, a second acquisition module, and a construction module; The first acquisition module is used to acquire avalanche event location information; the avalanche event location information includes the location data of each avalanche event in multiple avalanche events; the location data includes the avalanche center, avalanche section, and avalanche slope; The second acquisition module is used to acquire avalanche disaster information; the avalanche disaster information includes disaster-related data for each of the multiple avalanche events; the disaster-related data includes disaster-prone data, disaster-causing data, disaster data, disaster-bearing data, and rescue data; The construction module is used to construct a training dataset using the avalanche event location information and avalanche disaster information; The training dataset includes multiple training samples; the training samples include the location data, disaster-related data, and data correlations of one of the multiple avalanche events; the data correlations indicate the association between the avalanche center, avalanche section, and avalanche slope and the disaster-related data, respectively.

8. An electronic device, characterized in that, The device includes a processor and a memory coupled to the processor; the memory is used to store computer instructions, and when the electronic device is running, the processor executes the computer instructions stored in the memory to cause the electronic device to perform the method as described in any one of claims 1 to 4.

9. A computer-readable storage medium, characterized in that, It includes computer program instructions that, when executed by a computer, cause the computer to perform the method as described in any one of claims 1 to 4.

Images