Mountain torrent disaster air-space-ground collaborative real-time deduction and disaster avoidance path planning method

Abstract

The invention discloses a mountain torrent disaster air-space-ground collaborative real-time deduction and disaster avoidance path planning method, relates to the field of path planning, and decomposes a complex global path finding problem into two levels of high-level macroscopic strategy and bottom-level local execution. A high-level strategy network quickly determines a strategic sub-target on an abstract and simplified road network diagram in combination with future risk prediction, so that high-time-consumption global re-calculation on a complete road network is avoided, and the problem of slow planning response is solved. And then, the underlying strategy network makes a decision only in the current local area according to the fine real-time risk map, and calculates the specific next action, so that the sudden change of the local environment can be quickly coped with. In this way, the calculation complexity is greatly reduced, real-time dynamic planning becomes possible, and therefore the disaster avoidance path giving consideration to safety and efficiency is efficiently generated in an instantly changing disaster environment.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of path planning, and more particularly, to a mountain flood disaster space-air-ground collaborative real-time deduction and disaster avoidance path planning method. BACKGROUND

[0002] Mountain flood disasters pose a serious threat to the safety of life and property of people in mountainous areas due to their strong suddenness, rapid evolution and great destructive power. When a disaster occurs, quickly and safely transferring trapped personnel to a shelter is the core task of emergency response. This requires not only accurate perception and prediction of the real-time dynamic evolution of floods, but also planning a real-time risk-avoiding life channel for each individual in danger. Therefore, it is of great practical significance and application value to build an intelligent solution that can integrate multi-source perception data, perform real-time disaster deduction, and dynamically generate disaster avoidance paths.

[0003] However, in the current disaster prevention and mitigation technology system, disaster monitoring and early warning are often disconnected from individual-level path navigation planning. Traditional disaster monitoring systems, such as satellite remote sensing and ground sensors, are good at assessing and warning macro-risk areas, but they cannot directly translate into fine-grained dynamic navigation instructions for individuals. Existing path planning methods, such as the shortest path algorithm commonly used in navigation software, are mostly based on static road network data and cannot adapt to the rapidly changing risk scenarios in mountain flood disasters. Once a planned route is suddenly flooded by rising water, the original plan becomes invalid, and even worse, it may lead disaster avoidance personnel to a more dangerous situation. Behind this technical deficiency is the real-time-optimization trade-off dilemma faced by disaster avoidance path planning in dynamic environments. On the one hand, the state of the mountain flood risk environment has a rapid time-varying nature, and the calculation time of the path planning algorithm must be much less than the change period of the environmental risk, otherwise its planning results will lag behind the actual disaster situation and lose timeliness. On the other hand, in a large and complex road network, repeated recalculation of the globally optimal path to respond to each update of the risk will bring an unbearable computational burden, which contradicts the requirement of emergency response to seize every second. At the same time, the prediction of future water depth, flow rate and other risks also contains uncertainties, and how to make reliable decisions between ensuring short-term safety and pursuing long-term efficiency is a problem that traditional path planning algorithms cannot solve.

[0004] Therefore, how to efficiently generate a dynamic disaster avoidance path that is both safe and reliable in a rapidly changing and uncertain risk environment is a technical problem that needs to be solved in the current mountain flood disaster emergency response field. SUMMARY

[0005] In view of the above problems existing in the prior art, according to an aspect of the present application, a mountain flood disaster space-air-ground collaborative real-time deduction and disaster avoidance path planning method is provided, which comprises: Based on the flood deduction model, the original sensor data is dynamically risk field perceived and time series predicted to obtain a series of risk prediction maps of N future time steps; Based on the road network data and topological feature data, a high-level abstract graph and a bottom-level local graph set are constructed; The current macro node, the macro node of the shelter, the high-level abstract graph, and the series of risk prediction maps of N future time steps are input into the pre-trained high-level strategy network to obtain a sub-target macro node; The current macro node area risk prediction map and the current macro node area bottom-level local graph are extracted from the bottom-level local graph set and the series of risk prediction maps of N future time steps; The current precise position, the current macro node area risk prediction map, and the current macro node area bottom-level local graph are input into the pre-trained bottom-level strategy network to obtain a next step action instruction; Based on the next step action instruction and user state feedback, a continuous navigation instruction stream is generated.

[0006] Compared with the prior art, the mountain flood disaster space-air-ground coordinated real-time deduction and disaster avoidance path planning method provided by the present application proposes a hierarchical decision path planning to solve the real-time-optimization trade-off problem in a dynamic environment. Specifically, the complex global routing problem is divided into high-level macro strategy and bottom-level local execution. The high-level strategy network quickly determines a strategic sub-target on the abstract simplified road network graph, combining future risk prediction, avoiding high-time-consuming global recalculation on the complete road network, and solving the problem of slow planning response. Then, the bottom-level strategy network only makes decisions in the current local area according to the fine real-time risk map, and calculates the specific next step action, so as to quickly respond to the mutation of the local environment. This high-level direction determination and bottom-level detail determination architecture greatly reduces the computational complexity, making real-time dynamic planning possible, thereby efficiently generating a disaster avoidance path that takes into account safety and efficiency in the rapidly changing disaster environment. BRIEF DESCRIPTION OF DRAWINGS

[0007] The above and other objects, features and advantages of the present application will become more apparent from the following detailed description of embodiments of the present application, taken in conjunction with the accompanying drawings. The drawings provided in the specification and the embodiments of the present application together serve to provide a further understanding that enables others skilled in the art to make or use the present application. The drawings provided are for illustrative purposes and are not intended to limit the present application. In the drawings, the same reference numerals generally refer to the same components or steps throughout the drawings.

[0008] Figure 1 A flowchart of the mountain flood disaster space-air-ground coordinated real-time deduction and disaster avoidance path planning method according to the embodiments of the present application.

[0009] Figure 2A data flow diagram of the mountain flood disaster space-time-ground collaborative real-time deduction and disaster avoidance path planning method according to the embodiment of the present application.

[0010] Figure 3 A flow chart of step S3 in the mountain flood disaster space-time-ground collaborative real-time deduction and disaster avoidance path planning method according to the embodiment of the present application.

[0011] Figure 4 A flow chart of step S5 in the mountain flood disaster space-time-ground collaborative real-time deduction and disaster avoidance path planning method according to the embodiment of the present application. DETAILED DESCRIPTION

[0012] Embodiments of the present disclosure will be described in more detail with reference to the drawings. Although certain embodiments of the present disclosure are shown in the drawings, it is understood that the present disclosure can be implemented in various forms, and should not be interpreted as being limited to the embodiments set forth herein, but rather, these embodiments are provided to more thoroughly and completely understand the present disclosure. It should be understood that the drawings and embodiments of the present disclosure are for exemplary purposes only, and are not intended to limit the scope of protection of the present disclosure.

[0013] To solve the problems mentioned in the background art, the present application proposes a mountain flood disaster space-time-ground collaborative real-time deduction and disaster avoidance path planning method. Figure 1 A flow chart of the mountain flood disaster space-time-ground collaborative real-time deduction and disaster avoidance path planning method according to the embodiment of the present application. Figure 2 A data flow diagram of the mountain flood disaster space-time-ground collaborative real-time deduction and disaster avoidance path planning method according to the embodiment of the present application. As shown in Figure 1 and Figure 2 The mountain flood disaster space-time-ground collaborative real-time deduction and disaster avoidance path planning method according to the embodiment of the present application includes: S1, based on a flood deduction model, performing dynamic risk field perception and time series prediction on original sensor data to obtain a series of risk prediction maps of N future time steps; S2, based on road network data and topological feature data, constructing a high-level abstract graph and a bottom-level local graph set; S3, inputting the current macro node, the macro node where the shelter is located, the high-level abstract graph, and the series of risk prediction maps of N future time steps into a pre-trained high-level strategy network to obtain a sub-target macro node; S4, extracting a current macro node area risk prediction map and a current macro node area bottom-level local graph from the bottom-level local graph set and the series of risk prediction maps of N future time steps; S5, inputting the current precise position, the current macro node area risk prediction map, and the current macro node area bottom-level local graph into a pre-trained bottom-level strategy network to obtain a next step action instruction; and S6, based on the next step action instruction and user state feedback, generating a continuous navigation instruction stream.

[0014] In step S1, based on the flood evolution model, the original sensor data is dynamically perceived and time-series predicted to obtain a series of risk prediction maps of N future time steps. It should be understood that in the sudden event of mountain flood disaster, the environmental risk changes rapidly, and any disaster avoidance path based on current static information may fail in a short time. To plan a truly safe and effective life channel for trapped personnel, it is not only necessary to know the dangerous area at the moment, but also more importantly to predict the evolution trend of risk in the future period of time, so as to avoid the danger in advance when making decisions. The traditional path planning lacks time-series prediction of the future in the decision basis. Therefore, the application first performs dynamic risk field perception and time-series prediction on the original sensor data based on the flood evolution model to convert the dispersed and multi-source original observation data into a series of high-resolution future risk distribution maps with spatial and temporal continuity through deep fusion of physical models and data-driven models. This provides a dynamic and forward-looking decision basis for the subsequent path planning algorithm, thereby solving the technical problems of poor timeliness of path planning, disconnection between calculation and environmental change mentioned in the background art.

[0015] In a specific embodiment, step S1, based on the flood evolution model, the original sensor data is dynamically perceived and time-series predicted to obtain a series of risk prediction maps of N future time steps, comprising: S11, water body boundary extraction is performed on satellite remote sensing images and unmanned aerial vehicle video streams to obtain water body boundary vectors; S12, based on the water body boundary vectors, water level site data and digital elevation model data, submerged water depth inversion is performed to obtain a submerged water depth map at the current time; S13, based on the rainfall site data, rainfall field spatial interpolation is performed to obtain a rainfall field map at the current time; S14, the current time submerged water depth map, the current time rainfall field map, the digital elevation model data and the future rainfall forecast data are input into the pre-trained flood evolution model for recursive time-series prediction to obtain a series of water depth prediction maps of N future time steps; S15, based on the digital elevation model data, the series of water depth prediction maps of N future time steps are dynamically risk-quantified and time-space risk field generated to obtain a series of risk prediction maps of the N future time steps.

[0016] In the above embodiments, step S1 performs the following operations: first, the original sensor data required for this step is defined and collected. In a specific embodiment, the original sensor data includes satellite remote sensing images, unmanned aerial vehicle video streams, rainfall site data, and water level site data. Specifically, the satellite remote sensing images refer to image data obtained by optical sensors carried by earth observation satellites, for example, multispectral images from Gaofen-2 satellites, which have a spatial resolution of up to meters and can provide land cover conditions in a large area. Such data is received by ground stations or regularly obtained from data distribution service platforms. The unmanned aerial vehicle video stream refers to continuous video data that is real-time captured and transmitted by a high-definition camera carried by an emergency response unmanned aerial vehicle during low-altitude flight over the disaster area. Unmanned aerial vehicles can provide higher spatial resolution up to centimeters and higher timeliness of local area images, which are particularly suitable for key river sections, bridges, or residential areas for intensive monitoring. Rainfall site data refers to rainfall information collected by automatic weather stations or hydrological stations arranged in mountain flood disaster prone areas and their upstream catchment areas. These sites are equipped with precision sensors such as tipping bucket rain gauges, which can automatically measure rainfall in a unit of time and convert it into rainfall intensity (e.g., in millimeters per hour). The site automatically reports data packets containing the site's unique identification, geographic coordinates, and current rainfall intensity value to the central data server at a preset time frequency, such as every 5 or 10 minutes, through a built-in wireless communication module such as GPRS or 4G / 5G. In this way, a set of data points can be obtained, each of which accurately represents the rainfall intensity of a specific geographic location at the current time. Water level site data refers to data collected and uploaded by hydrological monitoring stations arranged at key cross-sections of rivers in the basin or low-lying areas prone to flooding. These sites are equipped with pressure, radar, or float water level gauges that can automatically measure the height of the current water surface relative to a certain fixed elevation reference at a high frequency, such as once per minute, and send data packets containing the site number, geographic coordinates, and real-time water level elevation value to the data processing center through a wireless communication network.

[0017] After obtaining the above data, the specific process of dynamic risk field perception and timing prediction is started. The first step is to perform S11 water body boundary extraction. Processing of satellite remote sensing images: first, the pre-processing of the obtained satellite remote sensing images is performed, including radiation calibration, converting the recorded dimensionless digital quantity, i.e. DN value, into a radiation brightness value with actual physical meaning; and atmospheric correction, eliminating the influence of atmospheric scattering, absorption and other factors on the reflectivity of ground objects, to obtain the true ground reflectivity image. After completing the pre-processing, the difference characteristics of water body reflectivity at different spectral bands are used to enhance the water body information. Specifically, the normalized difference water index (NDWI) is used for calculation, and the calculation formula is: (green light band reflectivity-near infrared band reflectivity) / (green light band reflectivity+ near infrared band reflectivity). Since water has strong absorption characteristics in the near-infrared band, and the reflectivity is higher in the green light band, the calculated NDWI value can well distinguish water from non-water. After calculation, an NDWI index map will be generated, in which the NDWI value of water pixels is a larger positive value, while vegetation, soil and other objects are negative or smaller positive values. Then, a suitable threshold needs to be set to segment the water body. The threshold can be automatically determined by the Otsu method, which finds an optimal segmentation threshold according to the gray histogram of the image, so that the inter-class variance is maximized. For example, the calculated threshold is 0.2, then all the regions in the NDWI map with a value greater than 0.2 are identified as water bodies and assigned a value of 1, and the remaining regions are assigned a value of 0, thereby obtaining a binary water distribution raster map. The processing of the unmanned aerial vehicle video stream uses a semantic segmentation method based on deep learning to cope with the challenges of complex scenes, rapid changes in lighting, shadows and reflections in the unmanned aerial vehicle images. First, the real-time video stream is frame-extracted, for example, one frame of image can be extracted per second to reduce data redundancy. Then, each extracted frame of image is input into a pre-trained semantic segmentation model. In this embodiment, the U-Net model architecture is used, which is a convolutional neural network with an encoder-decoder structure designed specifically for image segmentation. The encoder part is stacked with multiple convolutional layers and pooling layers, which are responsible for downsampling the input image to extract abstract features at different levels; the decoder part restores the high-level semantic feature map extracted by the encoder to the original image resolution through upsampling and convolution operations, and performs pixel-level classification. The U-Net architecture is based on the skip connection between the encoder and the decoder, which directly transmits the shallow, high-resolution feature map generated by the encoder during downsampling to the corresponding upsampling layer in the decoder, so that the decoder can utilize both deep semantic information and shallow detail information during precise positioning, thereby achieving accurate segmentation of the water body boundary. In particular, the weights and biases of the model are obtained through supervised learning on a large public water body segmentation dataset.The trained model is applied to the UAV images, and the model predicts the probability of each pixel in the image belonging to water body. Finally, a binary mask with the same size as the input image is output, where the pixels with a value of 1 represent the water body area. Finally, the water body distribution maps extracted from satellite images and UAV video streams are fused and converted to vector format. With the large-scale satellite extraction results as the base map, the high-precision UAV extraction results are used to update and correct the key areas. For example, if a certain area is covered by both types of data, the UAV segmentation result is preferred. After fusion, the final authoritative water body distribution raster map covering the entire study area is obtained. Then, the raster map is vectorized, that is, through the contour tracking algorithm, which is a prior art in the field of image processing, the algorithm automatically finds the starting point of the boundary of the water body area (pixel value 1) in the raster map, and according to the preset neighborhood search rule, it iterates all boundary pixel points one by one, and concatenates their coordinates until it returns to the starting point, thereby accurately converting the boundary of the pixel area into a closed polygon composed of a series of coordinate points. Finally, a water body boundary vector file is obtained, which contains a series of polygons with clear geographical coordinates, such as Shapefile or GeoJSON format, each of which accurately circumscribes a continuous water body on the ground at the current time.

[0018] The second step S12 is to perform the submerged water depth inversion. The digital elevation model data is a pre-prepared regional surface elevation information file stored in a grid form. It is made by aerial photogrammetry combined with stereo pairs or airborne laser radar scanning and other technical means, which can accurately depict the ups and downs of the ground. In this embodiment, a digital elevation model with a spatial resolution of 5 meters by 5 meters can be used, that is, the region is divided into countless 5-meter squares, and each square records an accurate ground elevation value. First, a continuous water surface elevation field is generated using discrete water level site data. Since the water surface of a flood in a connected region is generally considered as a continuous and gently changing surface, the water surface elevation in the entire submerged area can be calculated by spatial interpolation method according to the water level elevation values of the limited sites. In specific implementation, the geographic coordinates and the corresponding real-time water level elevation values of all water level sites located inside or adjacent to the water body boundary vector polygon are taken as input data. A continuous water surface is generated by using the triangular irregular network (TIN) interpolation method, which is a mature surface modeling technology in the field of spatial information. It connects the nodes of discrete water level sites by following the Delaunay criterion to build a network composed of optimized and non-overlapping triangles. Each triangle defines a plane with a specific slope and orientation, and the water surface elevation value at any position can be obtained by linear interpolation of the elevation values of the three vertices of the triangle. Thus, the discrete point data is effectively converted into a continuous three-dimensional surface that reflects the trend of the water surface. Subsequently, the continuous water surface elevation surface composed of vector triangles needs to be rasterized and overlaid with the water body boundary vector. The rasterization process creates a blank grid that is completely aligned with the digital elevation model data in terms of spatial range, coordinate system, and resolution. For each cell in the blank grid, the elevation value of the center point on the TIN surface is calculated and assigned to the cell, thereby generating a preliminary water surface elevation raster map. Then, in order to ensure that the water surface elevation only exists in the actual submerged area, a spatial overlay analysis is performed. The analysis uses the water body boundary vector polygon as a mask to cut the preliminary water surface elevation raster map. Specifically, each pixel in the raster map is traversed to determine whether its geographic location falls within the range defined by the water body boundary polygon. Only the pixels within the range are retained; otherwise, they are assigned a null value. Through this accurate cutting, a water surface elevation raster map with valid values only in the submerged area is finally obtained. Then, the water surface elevation raster map obtained in the previous step is subtracted from the pre-prepared digital elevation model raster map on a pixel-by-pixel basis. Specifically, for each grid cell, the final output value is equal to the water surface elevation value minus the ground elevation value of the cell.For example, in a certain cell at a specific geographic location, the water surface elevation is interpolated to be 125.8 meters, and the ground surface elevation is obtained from the digital elevation model to be 124.3 meters, then the result of the calculation is 1.5 meters, which is the inundation depth at this point. If at another location, the water surface elevation is 126.0 meters, and the ground surface elevation is 126.2 meters, the result of the calculation is -0.2 meters, which indicates that this location is not inundated, and the water depth value should be corrected to 0. By performing this calculation on all grid cells in the inundated area, a new raster image is finally generated, which is the current inundation depth map. Each pixel value of the map quantitatively represents the real-time inundation depth at the corresponding geographic location. The pixel values of the non-inundated area are 0, which intuitively and accurately reflects the spatial distribution and severity of the current flood.

[0019] The third step S13 performs spatial interpolation of the rainfall field. Since rainfall has certain continuity and correlation in space, i.e. the rainfall intensity of nearby locations is more likely to be similar, a spatial interpolation algorithm can be used to scientifically estimate the rainfall intensity of unknown locations (i.e. areas without deployed stations) based on the rainfall data of known stations. In this embodiment, the Kriging interpolation method is used, which is an existing geostatistical method based on spatial autocorrelation theory. It not only considers the distance relationship between the point to be interpolated and the known stations, but also considers the spatial distribution structure of the known station data points, and can provide optimal and unbiased estimates. The implementation process of Kriging interpolation mainly includes the following steps. First, perform variogram analysis. The geographic coordinates and rainfall intensity values of all rainfall stations are input to calculate the experimental variogram. The experimental variogram describes the variation of the half variance of the rainfall intensity variable with the distance between sample points. Specifically, all stations are paired, the distance between each pair of stations and the square of the difference in their rainfall intensity are calculated, then the station pairs with similar distances are grouped, and the half of the average value of the square of the difference in each distance group is calculated, i.e. the half variance. Plot the half variance values and their corresponding average distances to obtain the experimental variogram graph. Then fit the theoretical variogram model. In order to perform subsequent interpolation calculations, a continuous mathematical function model is needed to fit the discrete experimental variogram graph obtained in the previous step. Common theoretical models include spherical, exponential and Gaussian models. By least squares method or other methods, select a theoretical model that best fits the trend of the experimental variogram scatter plot, and determine the three key parameters of the model: nugget value, representing the variability at the smallest scale or measurement error; sill value, representing the total variance of the variable, i.e. the value when the variogram curve tends to be stable; and range, representing the maximum distance of spatial autocorrelation, beyond which two points can be considered independent of each other. For example, by fitting, a spherical model with a range of 15 kilometers may be obtained, which means that the spatial correlation of rainfall intensity mainly exists within a range of 15 kilometers. Then perform interpolation calculation. After the theoretical variogram model is determined, the rainfall intensity value of any unknown point in the study area can be estimated. Specifically, to generate a complete rainfall field map, the center point of each pixel of the output grid needs to be estimated. For a pixel point to be interpolated, the Kriging algorithm assigns an optimal weight to each known rainfall station based on the spatial relationship between the point and the surrounding stations, i.e. defined by the variogram model. The determination of these weights follows the principles of unbiasedness and minimum variance, i.e. ensures that the sum of all weights is 1, and minimizes the variance of the estimated value, i.e. optimality. For example, for a pixel point P to be interpolated, there are three rainfall stations A, B and C around it, with measured rainfall intensities of 12 mm / h, 18 mm / h and 15 mm / h respectively.The Kriging algorithm will solve a set of weight equations to get the weights assigned to A, B, C, such as 0.6, 0.3 and 0.1. Then, the estimated rainfall intensity of the pixel P is: 12*0.6+18*0.3+15*0.1=14.1mm / h. By repeating this interpolation calculation for all pixels in the region, a complete, spatially continuous raster image, i.e. the rainfall field map at the current time, can be generated. The coordinate system, spatial range and resolution of this map are strictly consistent with the aforementioned digital elevation model and the flood depth map, and each pixel value in the map represents the best estimated rainfall intensity at the geographical location at the current time.

[0020] The fourth step S14 performs recursive temporal prediction. The future rainfall forecast data is from the meteorological service agency, which is expressed as a series of future time step rainfall field maps, such as a 10-minute rainfall intensity forecast map for the next 6 hours. In a specific embodiment, the architecture of the pre-trained flood routing model is a CNN-LSTM architecture. This architecture is specifically designed for spatio-temporal sequence prediction problems. Its specific composition includes a spatial feature encoder CNN part, a time series processor LSTM part, and a spatial state decoder composed of transposed convolution layers. The CNN part acts as an encoder, composed of multiple convolution layers, activation function layers (such as ReLU), and pooling layers. Its function is to receive the input two-dimensional grid map (current time submerged water depth map, current time rainfall field map, digital elevation model data) and automatically extract spatial features that have a significant impact on flood evolution from it. For example, the convolution kernel can identify key patterns such as gullies, depressions in the terrain, and the main flow line of the current flooded area, and compress them into a low-dimensional feature vector. The LSTM part is responsible for processing the temporal dependence. It receives the current time spatial feature vector encoded by CNN and combines it with the state information saved in its internal memory unit from the previous time step to predict the spatial feature at the next time step. LSTM controls the flow and retention of information in the time series through its unique gating mechanism (input gate, forget gate, and output gate), effectively capturing the inertia, cumulative effect, and other temporal dynamics of flood evolution. Finally, the decoder part receives the predicted feature vector output by LSTM and performs upsampling through a series of transposed convolution layers (or deconvolution layers) to gradually restore the low-dimensional feature vector to a two-dimensional grid map with the same resolution as the input image. This map is the predicted next time submerged water depth map. It is worth mentioning that the internal parameters of the model, i.e. the weights and biases of all convolution layers and fully connected layers, are learned through the pre-training phase. In the specific implementation of the prediction, a recursive approach is adopted. The prediction time step is set to 5 minutes, and the total number of steps N is 12, i.e. the flood evolution in the next 1 hour is predicted. In particular, the setting of the time step and the total prediction time is determined by balancing the characteristics of rapid mountain flood disaster evolution and the actual needs of the evacuation route planning, aiming to ensure that the prediction can meet the timeliness of real-time decision-making and provide sufficient effective prediction range for route planning. In the first prediction step, i.e. predicting t=1, which is 5 minutes later: the submerged water depth map at the current time t=0, the rainfall field map at the current time t=0, the static digital elevation model map, and the future first 5 minutes, i.e. t=1, rainfall forecast map, are input into the pre-trained CNN-LSTM model as a data stack. After the encoding-processing-decoding process of the model, a grid map is output, which is the submerged water depth prediction map at t=1.At the second prediction step, i.e. predicting the time t=2, which is 10 minutes later: the inundation depth prediction map at time t=1 generated in the last step is used as the new input, replacing the initial observed water depth map. Meanwhile, the forecasted rainfall map at the second 5 minutes, i.e. time t=2, is used as the new rainfall driver. This updated data stack, i.e. the predicted water depth map at time t=1, the forecasted rainfall map at time t=2, and the digital elevation model data, is input into the same model again. The output of the model is the inundation depth prediction map at time t=2. This recursive process is looped for N times, which is 12 in this example. In each loop, the predicted water depth map output from the last loop is used as the initial condition for the next calculation, together with the forecasted rainfall at the corresponding future time step. In this way, a single-step prediction model is extended to a multi-step prediction tool, which iteratively and relayingly evolves the flood state at a series of future time points. The final output is a time series file containing N raster maps, i.e. a series of water depth prediction maps at N future time steps, which completely depicts the spatio-temporal evolution of the flood inundation depth in the next one hour from the current time.

[0021] Finally, the dynamic risk quantification and the spatio-temporal risk field generation are performed in S15. Firstly, based on the input series of water depth prediction maps of N future time steps and the digital elevation model data, a series of water flow velocity prediction maps corresponding to the N future time steps are calculated. Water flow velocity is one of the key factors for risk assessment. In this embodiment, the classic Manning formula in hydraulics is used to estimate the water flow velocity of each grid cell. The Manning formula expresses the relationship between the flow velocity, the hydraulic radius, the bed slope and the roughness in the uniform flow of open channels. For each water depth prediction map of each time step, the following calculations are performed for all the grid cells that are submerged by water, i.e. the pixels with water depth greater than 0. Firstly, the topographic slope of the cell is calculated according to the digital elevation model. This can be achieved by a 3 by 3 moving window, which calculates the ratio of the maximum elevation difference and the distance between the center pixel and its eight adjacent pixels, thereby obtaining the slope value S of the point. Secondly, the Manning roughness coefficient n of the cell is determined. The roughness coefficient n represents the resistance of the ground surface to water flow, which is related to the land cover type. The value can be set in advance according to the land use classification map, for example, for a cement road, n can be set to 0.015; for grassland, n can be set to 0.035; for forest land, n can be set to 0.1. These values are stored in a roughness coefficient grid map aligned with the digital elevation model. Then, the hydraulic radius R is determined. For shallow water flow formed by flash floods, the hydraulic width is much larger than the water depth, so the hydraulic radius can be approximated as the water depth h, which is directly read from the current water depth prediction map. Finally, these values are substituted into the Manning formula v = (1 / n)*h^(2 / 3)*S^(1 / 2) to calculate the water flow velocity v of the cell. Repeat this operation for all the submerged pixels in the water depth prediction map to generate a complete water flow velocity prediction map. Perform this process for the series of water depth prediction maps of N time steps to obtain a series of water flow velocity prediction maps containing N grid maps. Then, a risk quantification function is established, and the two physical quantities of water depth and water flow velocity are integrated into a single risk index using the function. The risk index aims to reflect the danger of flood to personnel and vehicle traffic. In this embodiment, a risk level system based on threshold division is established, which divides the risk into four levels: level 1 low risk, level 2 medium risk, level 3 high risk and level 4 extremely high risk. The specific quantification rules are as follows: if the water depth h of a cell is less than 0.3 meters and the water flow velocity v is less than 0.5 meters per second, then the risk index R is assigned a value of 1; if the water depth h is between 0.3 meters and 1.0 meters, or the water flow velocity v is between 0.5 meters per second and 1.5 meters per second, then the risk index R is assigned a value of 2; if the water depth h is between 1.0 meters and 2.0 meters, or the water flow velocity v is between 1.5 meters per second and 2.5 meters per second, then the risk index R is assigned a value of 3; if the water depth h is greater than or equal to 2.0 meters, or the water flow velocity v is greater than or equal to 2.5 meters per second, then the risk index R is assigned a value of 4.The setting of these thresholds is based on the safety standards for people to remain stable in floods in the field of water conservancy and disaster management. Then, the above-mentioned risk quantification function is applied to generate a final series of risk prediction maps. For each of the N future time steps, the corresponding water depth prediction map and the water flow velocity prediction map are paired. Then, for each corresponding grid cell in the two maps, the values of the water depth h and the flow velocity v are read, and the risk index R of the cell is calculated according to the risk quantification rule defined in the second step. The calculated risk index R is taken as the pixel value of the new grid map at this position. After the calculation of all grid cells is completed, a risk prediction map for this time step is generated, and each pixel value of the map is no longer a physical unit (such as meters or meters / second), but a dimensionless integer representing the danger level, i.e. 1 to 4. After performing this operation on the prediction data of N time steps, a complete time series consisting of N risk prediction maps is obtained. This sequence file, i.e. the series of risk prediction maps for the next N time steps, constitutes a dynamic spatiotemporal risk field, which intuitively and quantitatively shows how the danger level at each location in the region will evolve over time in the future period.

[0022] In step S2, based on the road network data and the topological feature data, a high-level abstract graph and a set of bottom-level local graphs are constructed. Correspondingly, in extreme emergency scenarios such as flash floods, path planning faces a serious real-time-optimization trade-off dilemma. A complete road network covering a vast mountainous area may have a number of nodes and edges reaching tens of thousands, and if a global path search is directly performed on this fine road network, the computational overhead is huge, and it is difficult to meet the real-time re-planning requirements under the rapidly changing disaster conditions. In order not to sacrifice the necessary accuracy, and greatly improve the planning efficiency, the present application constructs a high-level abstract graph and a set of bottom-level local graphs based on the road network data and the topological feature data, which is to decompose a single and complex global path finding problem into two levels of macro-strategic decision and local tactical execution, by constructing a simplified high-level abstract graph for fast strategic path finding and a set of bottom-level local graphs retaining all details for fine navigation, so as to balance between computational efficiency and path quality, and solve the technical problems of heavy computational burden and slow response mentioned in the background art.

[0023] In a specific embodiment, step S2 performs the following operations: the road network data is a digital map file describing the geometry and connection relationship of the road network, stored in a graph data structure. This data can be obtained from public geographic information platforms such as OpenStreetMap or professional map data suppliers. Its basic constituent units are nodes and edges. Nodes represent road intersections or endpoints and contain attributes such as geographic coordinates; edges represent specific road segments connecting two nodes and contain attributes such as length, road grade, speed limit, etc. Topological feature data refers to information describing the structural characteristics and importance of the road network, which is usually not original data but derived data obtained by graph theory analysis of the road network data. For example, node clusters in the road network can be identified by community detection algorithms, and these cluster information is a kind of topological feature data.

[0024] After the above data is prepared, the specific operation of constructing a hierarchical graph set is performed.

[0025] The first step is to perform in-depth analysis of the topological structure of the original road network data to construct a high-level abstract graph. The high-level abstract graph is a simplification and generalization of the original road network, which only retains the connection relationship between regions and ignores the details within the region. Its construction process applies a community detection algorithm based on network topological features, such as the Louvain algorithm. This algorithm can automatically divide the entire road network into several node communities according to the key topological information of network connection tightness. Within each community, the connection between nodes is much denser than the connection between communities. After division, each community is logically regarded as a macro node, which constitutes the basic unit of the high-level abstract graph. For example, a complex mountainous road network containing 10,000 intersections may be divided into 50 communities, so the high-level abstract graph only contains 50 macro nodes. Then, macro edges are established in the high-level abstract graph. If there is at least one edge in the original road network connecting nodes belonging to two different communities, a macro edge is established between the macro nodes corresponding to these two communities. The original nodes connecting different communities are called boundary nodes. In this way, the complex connection relationship in the original road network is simplified to the connection relationship between macro regions. The final generated high-level abstract graph has a much smaller number of nodes and edges than the original road network, greatly reducing the search space and computational complexity when planning strategic and long-distance paths.

[0026] In the second link, based on the above topology division result, a bottom local area graph set is constructed. The bottom local area graph retains all the details of the road network, but its scope is limited to a single macro node, i.e. the interior of a single community. Its construction process is as follows: traverse each community divided in the first link. For each community, extract a subgraph from the original road network graph. The subgraph contains all original nodes and original edges belonging to the community. In addition, the subgraph also contains all boundary nodes of the community, because these boundary nodes are the portals to and from the local area. By repeating this operation for all communities, a set consisting of multiple independent detailed local road network graphs is obtained. For example, if the road network is divided into 50 communities, then a set of 50 bottom local area graphs will be obtained, each of which accurately describes the detailed road network structure inside a macro node. The final output is a unique high-level abstract graph and a set of bottom local area graphs containing multiple graphs. This data structure reconstructs a large, flat road network into a total-division hierarchical system.

[0027] In step S3, the current macro node, the macro node where the refuge is located, the high-level abstract graph and the series of risk prediction graphs of the next N time steps are input into the pre-trained high-level strategy network to obtain the sub-target macro node. It can be understood that directly planning a globally optimal path to the final refuge in the complete bottom road network will result in unacceptable computational delay due to its huge search space and dynamically changing risk cost. This delay makes the planning result often lag behind the rapidly changing disaster situation, losing the significance of guiding risk avoidance. In order to balance between safety and efficiency, the current macro node, the macro node where the refuge is located, the high-level abstract graph and the series of risk prediction graphs of the next N time steps are input into the pre-trained high-level strategy network to utilize an intelligent decision-making model with foresight to quickly conduct strategic reasoning on the simplified macro road network. It does not pursue to calculate the ultimate path in one step, but according to the comprehensive evaluation of future risks, it intelligently selects the next most favorable intermediate area, i.e. the sub-target macro node, as the direction of the current stage.

[0028] In a specific embodiment, Figure 3 The flow chart for step S3 of the flash flood disaster space-air-ground coordinated real-time reasoning and refuge path planning method according to the embodiment of the application. As shown in Figure 3As shown, in step S3, the current macro node, the shelter macro node, the high-level abstract graph and the series of risk prediction graphs of the future N time steps are input into the pre-trained high-level strategy network to obtain a sub-target macro node, including: S31, performing macro path time sequence risk cost aggregation on the series of risk prediction graphs of the future N time steps and the high-level abstract graph to obtain a macro edge time sequence cost matrix; S32, constructing a state vector based on the macro edge time sequence cost matrix, the current macro node, the shelter macro node and the high-level abstract graph; S33, inputting the state vector into the pre-trained high-level strategy network to obtain the sub-target macro node.

[0029] In the above embodiment, step S3 performs the following operations: Firstly, macroscopic path timing risk cost aggregation is performed in S31. The first step is to traverse each macro edge in the high-level abstraction graph. A macro edge represents a macroscopic connection between two adjacent macro nodes, i.e. communities. For example, there is a macro edge connecting macro node 17 and macro node 18 in the high-level abstraction graph. The second step is to identify the specific physical paths corresponding to each macro edge in the real world. A macro edge can correspond to multiple actual roads connecting two communities. This process requires querying the original road network data. Specifically, for macro edge (17, 18), all road edges in the original road network graph that satisfy the following conditions need to be found: the two endpoints of the edge belong to macro node 17 and macro node 18, respectively. Through the search, it may be found that there are two actual road segments, denoted as segment A and segment B, which realize the span from macro node 17 to macro node 18. The third step is to calculate the risk cost of each physical path at each future time step. This is a nested calculation process that needs to traverse the future N time steps, where the prediction step N is 12 and the time step length is 5 minutes. Taking the calculation of the risk cost of segment A at the t-th time step in the future, e.g. t=3, 15 minutes in the future, as an example: first, obtain the sequence of geographic coordinates of segment A, i.e. its geometric line shape on the map. Then, call the 3rd picture in the series of risk prediction maps for the future N time steps, i.e. the risk prediction map at t=3. Then, perform spatial overlay analysis by overlaying the line shape of segment A onto this risk prediction map. Since segment A is a line and the risk prediction map is a grid, the segment will pass through a series of grid cells. Extract the risk index values of all grid cells passed by segment A. For example, the extracted risk value sequence may be [2, 2, 3, 2, 3, 3]. In order to obtain a single value that can represent the risk of the entire segment, the sequence needs to be aggregated. In this embodiment, the maximum value is taken, because for disaster avoidance, the safety of a path depends on its most dangerous point. Therefore, the risk cost of segment A at t=3 is 3. Similarly, the same operation is performed on segment B, e.g. its grid risk values passed at t=3 are [2, 4, 4, 3], and its risk cost is 4. For segment A and segment B, their risk costs at all 12 future time steps are calculated, thereby obtaining two cost time sequences with a length of 12. The fourth step is to aggregate the costs of multiple physical paths to obtain the final cost of the macro edge at each time step. When moving from one macro node to another adjacent macro node, the rational choice is to take the path with the lowest risk among all available physical paths. Therefore, the travel cost of macro edge (17, 18) at t=3 should be defined as the smaller value of the risk costs of segment A and segment B at t=3. According to the calculation in the previous step, min(3, 4)=3. Applying this aggregation rule to all 12 future time steps, a complete cost time sequence of macro edge (17, 18) is obtained.The fifth step is to build the final macro-edge time-series cost matrix. For all macro-edges in the high-level abstract graph, repeat the above first to fourth steps of calculation. Finally, all the calculation results are summarized into a two-dimensional matrix, each row of the matrix corresponds to a unique macro-edge in the high-level abstract graph, for example, sorted by the starting and ending node numbers of the edge, each column corresponds to a future time step, i.e. from t = 1 to t = N. The element C(i, t) in the matrix is the aggregated risk cost of the ith macro-edge at the future tth time step. This matrix completely and quantitatively describes the risk evolution trend of all optional paths in the future period of time when moving at the macro level.

[0030] Then, a fixed-dimension state vector is constructed in S32. In a specific embodiment, step S32, based on the macro-edge time-series cost matrix, the current macro-node, the shelter macro-node and the high-level abstract graph, constructs the state vector, including: S321, the current macro-node and the shelter macro-node are encoded with location information to obtain the current macro-node location information encoding vector and the shelter macro-node location information encoding vector; S322, based on the high-level abstract graph and the macro-edge time-series cost matrix, the current macro-node is encoded with action space cost to obtain the current macro-node action cost encoding vector; S323, the current macro-node location information encoding vector, the shelter macro-node location information encoding vector and the current macro-node action cost encoding vector are spliced to obtain the state vector.

[0031] Specifically, the current macro node refers to the macro area to which the user's current real-time location belongs. The user's current accurate geographic coordinates of longitude and latitude are obtained through the positioning function of the mobile terminal, and then the coordinates are matched with the geographic range of the community, i.e., the macro node, divided in step S2, to determine the community number in which the user is located. The number is the current macro node. For example, if the user's current coordinates fall within the community range with the number 17, the current macro node is 17. The macro node where the shelter is located refers to the macro area to which the final safe shelter belongs. Its acquisition method is the same as that of the determination of the current macro node. The community number where the shelter is located is obtained by matching the fixed geographic coordinates of the shelter, for example, the number is 42. The first link S321 converts the discrete node number into a numerical vector that is easy for the neural network to process. In this embodiment, one-hot encoding is used. First, the total number of macro nodes is obtained from the high-level abstract graph, for example, there are a total of 50 macro nodes. Then, the dimension of the location information encoding vector is determined as 50. For the current macro node 17, its location information encoding vector is a vector with a length of 50, and the element value of the 17th position is 1, and the element values of all the remaining 49 positions are 0. Similarly, for the macro node where the shelter is located 42, a vector with a length of 50 is also generated, in which the element value of the 42nd position is 1, and the element values of all the remaining positions are 0. In this way, two 50-dimensional numerical vectors that can clearly represent the starting and ending macro locations are obtained, i.e., the current macro node location information encoding vector and the macro node where the shelter is located location information encoding vector. The second link S322 encodes the future risk of all selectable macroscopic paths starting from the current location into a fixed-length vector. First, the maximum number of neighbors that the network can handle, i.e., the maximum action space, needs to be determined. This value needs to be preset, for example, set to 8. The preset value is determined by statistically analyzing the degrees (i.e., the number of neighbors) of all macro nodes in the high-level abstract graph and selecting an integer not less than the actual maximum degree in the graph. Then, all actual neighboring macro nodes of the current macro node 17 are found in the high-level abstract graph to obtain a list, for example, [16, 18, 25], which indicates that there are 3 possible moving directions starting from the macro node 17. Then, for each possible action, i.e., moving to a neighboring macro node, the corresponding cost time series is extracted from the macro edge time cost matrix. For example, for the action of moving to the macro node 18, the row representing the macro edge (17, 18) in the cost matrix is found, which contains the passing risk cost of the future N time steps, N=12, forming a 12-dimensional cost vector, such as [2, 2, 3, 3, 4, 4, 4, 3, 3, 2, 2, 1]. This operation is performed for all 3 actual neighbor nodes to obtain 3 12-dimensional cost vectors. Then, the 3 cost vectors are spliced in a predetermined order, such as in ascending order of neighbor node number, to form a 3*12=36-dimensional vector.Since the maximum number of neighbors is set to 8, and the actual number of neighbors is only 3, in order to ensure the final output vector dimension constant, the remaining 5 empty action bits need to be filled. In this embodiment, a preset extremely high risk value such as 9999 is used for filling. That is, a length of 5*12=60 filling vector is generated, and all elements of the filling vector are 9999. Finally, the 36-dimensional real cost vector obtained in the foregoing is spliced with the 60-dimensional filling vector to obtain a final length of 96-dimensional (8*12=96) current macro node action cost coding vector, which completely describes the future timing risk of all optional actions. The third link S323 is a direct vector merging operation. The two 50-dimensional current macro node position information coding vectors generated by the first link and the shelter macro node position information coding vector are spliced with the 96-dimensional current macro node action cost coding vector generated by the second link in a fixed order, such as current position, target position, and action cost, to form a longer single vector. Finally, a total dimension of 50+50+96=196 is obtained. The 196-dimensional vector is the state vector, which encapsulates all the key information required for current decision making in a unified structure, fixed dimension numerical form, including where to go, where to go, and how to go in different directions.

[0032] Finally, a pre-trained high-level policy network is executed S33. In a specific embodiment, step S33, the state vector is input into the pre-trained high-level policy network to obtain the sub-target macro node, comprising: S331, inputting the state vector into the pre-trained high-level policy network to obtain an action Q value vector; S332, performing optimal action selection based on the action Q value vector to obtain an optimal action index; S333, based on the optimal action index, performing sub-target mapping on the actual neighbor list of the current macro node to obtain the sub-target macro node.

[0033] In particular, the pre-trained high-level policy network is a decision model based on deep reinforcement learning, which is essentially a multi-layer perceptron (MLP). Its specific network architecture is set as follows in this embodiment: an input layer whose number of neurons completely matches the dimension of the state vector generated in step S32, i.e. 196 neurons. Two hidden layers, each containing 256 neurons and using a rectified linear unit (ReLU) as the activation function to introduce nonlinear computing ability, so that the network can learn more complex decision boundaries. An output layer whose number of neurons matches the maximum number of neighbors (i.e. the maximum action space dimension) set in step S32, i.e. 8 neurons. All internal parameters of this network, including the connection weights between layers and the bias terms of neurons, are obtained by pre-training in an offline environment using the deep Q network (DQN) algorithm. The training process is carried out in a simulation environment that can generate a large number of dynamically evolving spatio-temporal risk fields based on historical flood disaster data or hydrodynamic simulation results. In training, a policy network is randomly placed at a starting macro node and given a target macro node. The agent selects an action (move to an adjacent macro node) based on its current state represented by the state vector, and the environment gives it a reward value (usually a negative risk cost) based on the macro edge time series cost corresponding to the action. Through thousands of repeated trials, the agent uses the DQN algorithm to continuously update its network parameters and learn an optimal action value function (Q function) that can accurately predict the expected value of future cumulative rewards that can be brought by taking any available action in any state. After training is complete, this fixed-parameter network has the ability to make long-term optimal decisions in complex dynamic environments. After preparing this pre-trained network model, the first step S331 is performed. The state vector is fed into the input layer of the high-level policy network, and the data is then forward propagated in the network: the input value is multiplied by the first layer weight matrix and added to the bias, and then passed through the ReLU activation function, and its output serves as the input of the second hidden layer; this process is repeated between hidden layers; finally, the output of the second hidden layer is multiplied by the output layer weight matrix and added to the bias to obtain the final output. Since the output layer has no activation function, its output value directly represents the Q value. The output of this process is an 8-dimensional numerical vector, i.e. the action Q value vector, each element of which corresponds to a potential action of moving to a certain neighbor node, and its numerical value represents the network's evaluation of the long-term value of the action. The second step S332. Traverse all elements of the Q value vector, find the maximum value, and return the index of the position where the maximum value is located, with the index starting from 0. If the actual neighbors of the current macro node 17 are 3, they correspond to the first 3 positions of the action space.The 8-dimensional action Q-value vector outputted by the network can be [-125.7, -88.2, -210.4, -9999, -9999, -9999, -9999, -9999]. Among them, the second element -88.2 is the maximum value among all values, i.e. the optimal action index is 1. The third step S333. The optimal action index obtained in the previous step is only an internal number relative to the standardized action space, which needs to be converted into the actual macro node ID. The optimal action index value is 1, and the actual adjacent node list of the current macro node 17. This list needs to be pre-ordered according to a fixed rule, for example, in ascending order of macro node ID, to obtain the list [16, 18, 25]. Then, using the optimal action index 1 as the subscript, searching in this list. The 0th element of the list is 16, and the 1st element is 18. Therefore, the result of the search is macro node 18, which is the sub-target macro node, which represents the most strategically valuable next macro movement target made by the high-level strategy network after comprehensively evaluating the current position, the final target and the timing risk of all future paths.

[0034] In a preferred specific embodiment, the step S33 of inputting the state vector into the pre-trained high-level strategy network to obtain the sub-target macro node comprises: S33-1, inputting the state vector into the pre-trained high-level strategy network based on a dynamic interactive attention and value decomposition mechanism to obtain an action Q-value vector; S33-2, performing optimal action selection based on the action Q-value vector to obtain an optimal action index; S33-3, performing sub-target mapping on the actual adjacent node list of the current macro node based on the optimal action index to obtain the sub-target macro node. In particular, steps S33-2 and S33-3 are the same as the embodiments of S332 and S333 in the above-mentioned embodiments, and therefore will not be described again. Here, the detailed implementation process of step S33-1 is described.

[0035] It can be understood that in a standard high-level strategy network, the spliced state vector is processed through a series of affine transformations, i.e. weight matrix and bias, which will cause the position features and cost features to be linearly combined by the same weight matrix at the first layer of the network. This processing method lacks dynamic correlation, i.e. the network cannot interactively and differently pay attention to and evaluate the risks (costs) of each selectable action according to the current position and target (state). When this dynamic interactive process of state and cost cannot be explicitly modeled, what the network learns is a static, global mapping relationship, rather than a dynamic weighting strategy for the current specific situation. For example, when a shelter is within arm's reach, but there is an imminent danger in a certain direction, an ideal strategy network should dynamically and greatly increase the attention to the cost features of the dangerous direction.

[0036] Therefore, the application proposes to input the state vector into a pre-trained high-level strategy network based on a dynamic interactive attention mechanism and a value decomposition mechanism to build a more refined decision-making model to solve the above-mentioned static mapping problem. This step introduces a dynamic interactive attention mechanism, which enables the network to explicitly learn and establish the dynamic association between state features and action cost features, thereby focusing on key risk information in the current situation when making decisions. At the same time, this step further adopts a value decomposition mechanism to decompose the Q value estimation into a composite state value function V(S) and an advantage function A(S, a). Among them, V(S) is used to evaluate the goodness of the current state S itself, regardless of which action is taken; while the advantage function A(S, a) is used to evaluate the relative goodness of taking action a in state S compared to other actions. This decomposition can make the network learn more efficiently and stably. For example, in some states (such as being completely surrounded by flood), no matter which action is taken, the long-term return is very low, at this time the network only needs to learn an accurate low V(S), and does not need to learn an accurate low Q value for each equally bad action, thereby speeding up the convergence of the model and improving the robustness of the decision.

[0037] Based on this, in a specific embodiment of the application, step S33-1, inputting the state vector into a pre-trained high-level strategy network based on a dynamic interactive attention mechanism and a value decomposition mechanism to obtain an action Q value vector, comprises: The state vector is split in a concatenation manner to obtain a current macro node position information encoding vector, a shelter macro node position information encoding vector, and a current macro node action cost encoding vector. It should be understood that the input original state vector is a single vector that simply concatenates position information and cost information, and its internal structure is not explicitly distinguished. In order to be able to apply different and targeted calculations to the position information representing the state and the cost information representing the action risk, the vector needs to be first split according to its original semantic structure to obtain three independent sub-vectors with dimensions of 50, 50, and 96, respectively.

[0038] The current location macro node position information encoding vector and the shelter location macro node position information encoding vector are spliced and input into the input layer and the hidden layer of the first multi-layer perception to obtain a position state vector, wherein the position state vector has the same length as the current location macro node action cost encoding vector. Correspondingly, the original position information encoding is highly sparse and can only represent isolated positions and cannot reflect the spatial relationship between the current position and the target position or the intrinsic state of the navigation task. Therefore, the first multi-layer perception is used to deeply encode the spliced position information, aiming to learn a dense, low-dimensional feature representation containing the complex relationship between the start point and the end point, so as to generate a position state vector with more information content, which represents not only the position but also the state of the current navigation task. At the same time, the output dimension is set to be the same as the action cost vector, which is for the next step of feature alignment and advantage correlation calculation. Specifically, first, the two 50-dimensional position information encoding vectors obtained in the previous step are spliced to form a 100-dimensional input vector. Then, the 100-dimensional vector is input into a pre-trained first multi-layer perception. The network contains one or more hidden layers (such as 128 neurons per layer using ReLU activation function), and the number of neurons in the output layer is accurately set to be the same as the length of the current location macro node action cost encoding vector, i.e. 96. After forward propagation calculation of the network, a 96-dimensional position state vector is finally output.

[0039] The position state vector and the current location macro node action cost encoding vector are advantageously associated based on the feature value granularity to obtain an advantage association representation matrix, i.e. ; wherein, and are the feature values of the th position and the th position in the position state vector and the current location macro node action cost encoding vector, is the function value with the natural constant e as the base, is the potential correlation value between and in the advantage association representation matrix. That is, this step is the core of realizing dynamic interactive attention, which explicitly models the mutual dependence relationship between each feature of the position state and each feature of the action cost. By calculating an association matrix, the network is no longer blindly processing all information, but can judge which action cost features are more important according to the current state features, and vice versa. The formula measures the correlation through the negative reciprocal of the feature value, and when both feature values are large, tends to 1, indicating strong association; otherwise, it tends to 0. Calculate for all i and j (from 1 to 96), and finally generate a 96x96 dimensional advantage association representation matrix, whose element values quantify the association strength between states and cost features.

[0040] Based on the advantage association representation matrix, the position state vector and the current macro node action cost coding vector are mapped and fused to obtain an action-state fusion representation vector, i.e. ; wherein, is the advantage association representation matrix, and are the position state vector and the current macro node action cost coding vector, respectively, is matrix multiplication, is position point addition, is the action-state fusion representation vector. It should be understood that after obtaining the attention weight (advantage association matrix), the weights need to be applied to the original feature vector to generate an attention-weighted, deeply fused feature representation. To this end, through matrix mapping (weighting) operation, the final fusion vector is no longer a simple information splicing, but contains which action cost is more worthy of attention under the current state and which state feature is the key context information for these action costs. Specifically, first, weight and and to obtain a 96-dimensional weighted state vector and a 96-dimensional weighted cost vector Then, add the two 96-dimensional vectors by position point to generate a 96-dimensional action-state fusion representation vector , which has higher information density and can better reflect the core contradictions of current decision.

[0041] Input the action-state fusion representation vector into the second multilayer perceptron to obtain the state value function. Accordingly, this step is the value flow branch of the value decomposition mechanism. It extracts the evaluation of the current state S itself from , i.e. V(S), which is the fusion of all information. Learning the state value independently can enable the network to give an accurate global evaluation when all available actions are equally bad or equally good, thereby simplifying the learning task and finally outputting a single scalar value representing the intrinsic value of the current macro state, unaffected by any specific action selection. Specifically, the 96-dimensional vector The input is a separate, pre-trained second-layer perceptron. This network architecture may contain one or more hidden layers, but its output layer must have only one neuron and no non-linear activation function. After forward propagation, the network outputs a scalar value, namely the state value function.

[0042] The action-state fusion representation vector is input into a third-layer perceptron to obtain the action-number advantage function vector. This step can be understood as the advantage stream branch of the value decomposition mechanism. Running parallel to the value stream, its purpose is to obtain the advantage function vector from the same fusion vector. In this process, the relative merit of each available action relative to the average level in that state is evaluated, i.e., the advantage A(S,a). The introduction of the advantage function allows the network to more finely distinguish the differences in merit between different actions. Specifically, the 96-dimensional V3 is input into another independent, pre-trained third multilayer perceptron. The input layer of this network is the same as that of the second multilayer perceptron, but the number of neurons in its output layer is set to a predefined maximum number of actions, such as 8. After forward propagation, the network finally outputs an 8-dimensional vector, i.e., the advantage function vector A(S,a).

[0043] The final output is a vector, where each element corresponds to the advantage value of a potential action, providing differential information between actions for the final Q-value calculation. The state value function and action number advantage function vectors are aggregated in a identifiable manner to obtain the action Q-value vector, i.e.: ;in, It is a state-value function. It is the action number advantage function vector. It is subtracted based on position. Indicates the current state All optional actions Sum of the dominance values, It is the number of eigenvalues ​​in the action-number advantage function vector. This is the action Q-value vector. In other words, this step is the final stage of the value decomposition mechanism, aiming to recombine the separate state values ​​and action advantages into the final decision-making basis, the Q-value. A recognizable aggregation method is used, that is, the advantage function is centered. This ensures the uniqueness of V(S) and A(S,a), preventing indistinguishable biases between them, thus guaranteeing the stability of training and the accuracy of V(S) as the state value benchmark. Finally, a Q-value vector for each selectable action is generated. This vector integrates the global value of the state and the relative advantage of each action, providing a direct and reliable quantitative indicator for selecting the optimal action.

[0044] In step S4, the current macro node area risk prediction map and the current macro node area bottom layer local graph are extracted from the bottom layer local graph set and the series of risk prediction maps of future N time steps. It should be understood that the high-level decision itself does not care how to make specific, one-meter-by-one-meter movements within the region. In order to generate navigation instructions that can be directly executed by the user, which are fine to each intersection, the decision scale needs to be transferred from the macro region to the micro road level. Directly performing such a fine search on the global bottom layer road network and risk map containing all details has huge computational overhead and cannot meet the real-time response requirement. Therefore, the application focuses the calculation range on the local environment where the current user is located by extracting the current macro node area risk prediction map and the current macro node area bottom layer local graph from the bottom layer local graph set and the series of risk prediction maps of future N time steps, and ensures that the generation of subsequent navigation instructions can be truly real-time and efficient by extracting the local road network map and the local risk map corresponding to the current macro node.

[0045] In a specific embodiment, step S4 performs the following operations: the first step is to extract the corresponding local road network map from the bottom layer local graph set according to the current macro node number. In step S2, an independent and detailed bottom layer local graph has been generated and stored for each macro node. This bottom layer local graph set can be regarded as a dictionary or hash table with the macro node number as the key and the corresponding road network map data as the value in the data structure. For example, it is determined through positioning that the user is currently located in the macro node numbered 17. Then, 17 is taken as the key to directly search in this set. The result of the search, that is, a detailed road network map data containing only all road intersections (nodes) and road segments (edges) within the macro node 17 and the boundary nodes as entrances and exits. This extracted independent graph data is defined as the current macro node area bottom layer local graph, which contains all static road network information required for navigation within the local area.

[0046] The second step is to spatially crop the global series of risk prediction maps of future N time steps according to the geographical range of the current macro node to extract the local risk map sequence. First, the accurate geographical boundary of the current macro node 17 needs to be obtained. This boundary is a polygon, which has been determined and stored together with the macro node number when the community division is performed in step S2. Then, the series of risk prediction maps from future N time steps are traversed. In the present application, N=12, so 12 cropping operations are required.

[0047] For the future t-th time step, i.e. t from 1 to 12, a global risk prediction map, a spatial clipping or masking operation is performed. Specifically, the bounding polygon of the macro node 17 is applied as a clipping window to the global risk map. This operation will keep all the grid cells that fall inside the polygon, while discarding or setting to null all the cells outside the polygon. In this way, the large-scale risk map that originally covers the whole study area is clipped into a small-scale risk map that only covers the area of the macro node 17. This new, smaller grid map is the current macro node area risk prediction map at time t. This clipping operation is repeated for all the 12 global risk maps, and finally a new time series consisting of 12 small-scale risk maps is obtained, which completely describes the risk dynamic evolution of the local area where the user is located in the next hour. The final current macro node area bottom-up local map is a graph data structure that accurately describes the road network of the current area and the current macro node area risk prediction map sequence, which is a set of grid data that accurately describes the dynamic risk of the area in the future.

[0048] In step S5, the current precise position, the current macro node area risk prediction map and the current macro node area bottom-up local map are input into the pre-trained bottom-up strategy network to obtain the next action instruction. Accordingly, in the changing local environment, how to safely and efficiently move from the current position to the next intersection is the key tactical link to achieve the final disaster avoidance goal. Traditional path search algorithms based on fixed cost are difficult to adapt to dynamically updated risks, and there may still be calculation delays in complex local road networks. Therefore, the current precise position, the current macro node area risk prediction map and the current macro node area bottom-up local map are input into the pre-trained bottom-up strategy network to utilize an intelligent network that has been specially trained to deeply understand the local spatio-temporal risk, which directly converts complex local environmental information into an explicit and immediate action instruction, thereby bypassing the time-consuming local path re-planning process, providing the user with the optimal tactical action at the next decision point (such as an intersection), thereby ensuring that the generation of navigation instructions has the highest level of real-time and safety.

[0049] In a specific embodiment, Figure 4 The flowchart for step S5 of the flash flood disaster space-time-ground cooperative real-time deduction and disaster avoidance path planning method according to the embodiment of the present application. As shown in FIG. 6, the current precise position, the current macro node area risk prediction map and the current macro node area bottom-up local map are input into the pre-trained bottom-up strategy network to obtain the next action instruction. Figure 4As shown, in step S5, the current accurate position, the current macro node area risk prediction map and the current macro node area bottom layer local area map are input into the pre-trained bottom layer strategy network to obtain the next step action instruction, including: S51, inputting the current accurate position, the current macro node area risk prediction map and the current macro node area bottom layer local area map into the pre-trained bottom layer strategy network to obtain an action probability distribution; S52, selecting an action with the highest probability value from the action probability distribution as a selected action; and S53, performing instruction conversion on the selected action to obtain the next step action instruction.

[0050] In the above embodiment, step S5 performs the following operation: first, S51 action probability distribution generation is performed. The pre-trained underlying policy network is a deep learning model that simultaneously processes rasterized spatiotemporal risk data and graphed road network topology data. In this embodiment, the network adopts a dual-branch fusion architecture, mainly composed of a three-dimensional convolutional neural network (3D-CNN) branch, a graph convolutional network (GCN) branch, and a subsequent multi-layer perceptron (MLP) decision head. The 3D-CNN branch is used to process the time series of risk prediction maps from the current macro node area. This sequence can be regarded as a three-dimensional data block (width, height, time), and the 3D-CNN can extract features in both spatial and temporal dimensions through its three-dimensional convolution kernel, effectively capturing dynamic evolution patterns such as the movement, expansion, or recession of risk areas, and finally outputting a low-dimensional vector encoding local spatiotemporal risk features. The GCN branch is used to process the underlying local graph of the current macro node area. GCN can operate directly on graph structure data, updating the feature representation of the center node by aggregating the information of neighboring nodes. After stacking multiple GCN layers, the network can learn the topological importance, connectivity, and other deep structural features of each intersection (node) in the local road network, and generate a feature vector for each node in the graph. The MLP decision head is responsible for fusing information from the two branches and making the final decision. It receives the spatiotemporal risk vector output by the 3D-CNN, the node feature vector related to the current location output by the GCN, and the encoding vectors of the current precise location, boundary nodes of the sub-target macro node, etc., concatenates them into a comprehensive state vector, and performs non-linear transformation through several fully connected layers. Finally, through a Softmax activation function, a vector representing the action probability distribution is output. In particular, the internal parameters (weights and biases of all convolutional layers, graph convolutional layers, and fully connected layers) of this underlying policy network are pre-trained in an offline environment through imitation learning or reinforcement learning. After preparing this pre-trained network model, the current precise location, the current macro node area risk prediction map, and the current macro node area underlying local graph are input into the pre-trained underlying policy network to obtain the action probability distribution. In specific processing, first, the input data is preprocessed. The current precise location is mapped to the nearest node or edge on the current macro node area underlying local graph through a map matching algorithm to determine the user's logical position on the graph, denoted as the current node. Then, according to the number of the sub-target macro node, all boundary nodes belonging to the sub-target macro node in the current local graph are found, which collectively constitute the current local navigation target. Next, the sequence of the current macro node area risk prediction map is input into the 3D-CNN branch of the network, and the current macro node area underlying local graph, as well as the current node, target boundary nodes, and other information are input into the GCN branch. After the two branches extract features, they are fused in the MLP decision head part.The MLP decision head finally outputs an action probability distribution vector. The dimension of the vector is equal to the number of neighbor nodes of the current node in the local graph, i.e. how many optional intersections there are at the current intersection. Each element in the vector has a value between 0 and 1, and the sum is 1, representing the probability of choosing to go to the corresponding neighbor node. For example, if there are three optional directions at the current intersection corresponding to neighbor nodes A, B, and C, the output action probability distribution may be [0.15, 0.80, 0.05], indicating that the network recommends going in the direction represented by neighbor node B with 80% confidence.

[0051] Then the selected action is executed 52. The probability vector of the action probability distribution is traversed to find the element with the largest value. The action corresponding to this element (i.e. which neighbor node to go to) is determined as the currently optimal selected action. In the above example, since 0.80 is the highest probability, going to neighbor node B is selected.

[0052] Finally, the instruction conversion is performed S53. First, the geographic coordinates of the current node and the target neighbor node B are queried from the underlying local graph. Then, a geographic direction vector pointing from the current node to the target neighbor node B is calculated. The calculation of the geographic direction vector is based on the accurate geographic coordinates of the current node and the target neighbor node B. Specifically, the latitude and longitude coordinates of the current node (e.g. longitude E1, latitude N1) and the latitude and longitude coordinates of the selected target neighbor node B (e.g. longitude E2, latitude N2) are read from the attribute data of the underlying local graph. Subsequently, the two components of the vector are determined by coordinate difference operations: the longitude value of the target neighbor node B is subtracted from the longitude value of the current node to obtain a longitude difference (E2-E1); at the same time, the latitude value of the target neighbor node B is subtracted from the latitude value of the current node to obtain a latitude difference (N2-N1). These two differences together form a two-dimensional geographic direction vector that accurately represents the straight line direction and relative distance from the current node to the target neighbor node B in the local plane coordinate system. Then, the direction vector is compared with the user's current heading vector. By calculating the angle between the two vectors, the steering action that the user needs to perform can be determined. For example, if the angle is between -45 degrees and +45 degrees, the instruction is to continue straight ahead at the next intersection; if the angle is between +45 degrees and +135 degrees, the instruction is to turn right at the next intersection; if the angle is between -45 degrees and -135 degrees, the instruction is to turn left at the next intersection. In addition, more specific instructions can be generated by combining the road name information in the road network data, such as continuing straight ahead along Zhongshan Road for 100 meters and then turning left at the next intersection into Jiefang Road. The final result is a clear, specific, and immediately executable next action instruction, such as turning right into Ping'an Road at the intersection 50 meters ahead, which is presented to the user through the user interface of the mobile terminal (such as voice or map arrow).

[0053] In step S6, a continuous navigation instruction stream is generated based on the next action instruction and the user state feedback. That is, the output of the previous step S5 is a single, tactical next action instruction generated based on a snapshot of the current moment. This solves the problem of how to act at a certain decision point (e.g. an intersection). However, a complete disaster avoidance process is not composed of a single decision, but a dynamic process composed of countless continuous decisions. The user's position is constantly moving, and the disaster risk is evolving in real time. A static, one-time instruction will soon become invalid. In order to connect a series of discrete, independent decision points into a complete, coherent and continuously adaptive guidance path, in this application, a continuous navigation instruction stream is generated based on the next action instruction and the user state feedback to establish a closed-loop, event-driven control process, taking the user's real-time state as feedback, continuously and automatically triggering and updating navigation decisions, thereby converting a single instruction into an uninterrupted, dynamically adaptive navigation instruction stream, ensuring the continuous effectiveness of the guidance.

[0054] In a specific embodiment, step S6 performs the following operation: the user state feedback refers to a high-frequency sequence of user real-time geographic coordinates continuously provided by the positioning module (e.g. GPS) of the mobile terminal, which may also include heading, speed, etc.

[0055] After the first next action instruction is generated and issued to the user, this step starts to initiate, continuously obtaining the user's real-time geographic coordinates from the mobile terminal, i.e. the user state feedback.

[0056] To determine when the next instruction needs to be generated, the process continuously checks a set of preset triggering conditions in parallel. These conditions include: 1. The last instruction contains an explicit geographic decision point, such as a specific intersection, whose geographic coordinates can be accurately obtained from the underlying road network data. The process calculates the distance between the user's current location and this decision point in real time. When this distance is less than a preset threshold, for example 15 meters, it is considered that the user has reached the decision point and the next instruction needs to be generated, at which time the triggering condition is satisfied. 2. The road segment corresponding to the last instruction also has an explicit geometric line in the road network data. The process calculates the perpendicular distance between the user's current location and this predetermined path. If this distance exceeds a preset threshold, for example 20 meters, it is determined that the user has deviated from the course and needs to be immediately re-planned, at which time the triggering condition is satisfied. 3. Since the risk prediction map generated in step S1 is dynamically updated over time, even if the user strictly follows the instructions, the original path may no longer be the optimal choice. Therefore, a fixed time period is set, for example 30 seconds, and every time the timer reaches, regardless of the user's state, a decision refresh is triggered to ensure that the path is always based on the latest risk information. 4. According to the user's real-time coordinates, it is continuously determined which macro node the user belongs to. When it is detected that the user's location moves from the current macro node, for example macro node 17, to a new macro node, for example sub-target macro node 18 determined by step S3, it means that a strategic phase has been completed and the planning of the next strategic phase needs to be started, at which time the triggering condition is satisfied. In particular, these preset values are determined through a large number of simulation tests and experience optimization, taking into account the positioning service accuracy error, the time-varying rate of disaster environment, the calculation and energy consumption limitations of terminal devices, and the actual navigation experience of users.

[0057] Once any of the above trigger conditions is met, a new decision-making process is immediately started. The specific calling logic is as follows: if the trigger condition is macro node crossing, it means that a new high-level strategic decision needs to be made. At this time, the process takes the new macro node where the user is currently located as input, returns and re-executes step S3 to determine the next sub-target macro node. Then, steps S4 and S5 are sequentially executed to generate and issue a new next-step action instruction. If the trigger condition is reaching a decision point, path deviation, or periodic update, and the user is still in the same macro node, it means that only a new bottom-level tactical decision needs to be made. At this time, the process directly returns and re-executes step S5, or first executes S4 and then executes S5 to generate and issue a new next-step action instruction using the user's latest accurate position and the latest local risk map. This monitoring-triggering-redetermination cycle mechanism naturally extends a single instruction into a continuous instruction stream. For example, a user receives an instruction to turn right at the intersection ahead. When the user approaches the intersection, the reaching a decision point condition is met, the cycle is triggered, and step S5 is called again to generate a new instruction to go straight on the current road for 200 meters according to the latest state. This process is repeated until the user finally reaches the shelter.

[0058] In summary, the mountain flood disaster space-earth-air coordination real-time deduction and disaster avoidance path planning method based on the embodiments of the present application is illustrated, which proposes a hierarchical decision-making path planning to solve the real-time-optimization trade-off problem in a dynamic environment. Specifically, the complex global routing problem is divided into two levels: high-level macro strategy and bottom-level local execution. The high-level strategy network quickly determines a strategic sub-target on the abstract simplified road network graph combined with future risk prediction, avoiding high-time-consuming global recalculation on the complete road network and solving the problem of slow planning response. Then, the bottom-level strategy network only makes decisions in the local area where it is currently located according to the detailed real-time risk map to calculate the specific next-step action, so as to quickly respond to local environmental mutations. This high-level direction-determining and bottom-level detail-determining architecture greatly reduces the computational complexity, making real-time dynamic planning possible, thereby efficiently generating a disaster avoidance path that takes into account safety and efficiency in a rapidly changing disaster environment.

[0059] The above specific embodiments further illustrate the purpose, technical solutions and beneficial effects of the present application. It should be understood that the above is only a specific embodiment of the present application and is not intended to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for real-time simulation and evacuation route planning of flash flood disasters using a combined air-space-ground approach, characterized in that, include: Based on the flood simulation model, dynamic risk field perception and time-series prediction are performed on the raw sensor data to obtain a series of risk prediction maps for the next N time steps. Based on road network data and topology feature data, construct a set of high-level abstract graphs and low-level local graphs; Input the current macro node, the macro node where the shelter is located, the high-level abstract graph, and a series of risk prediction graphs for the next N time steps into the pre-trained high-level policy network to obtain the sub-target macro nodes; Extract the risk prediction map of the current macro node region and the underlying local map of the current macro node region from the set of underlying local maps and the series of risk prediction maps for the next N time steps; The current precise location, the current macro node region risk prediction map, and the current macro node region bottom-level local map are input into the pre-trained bottom-level policy network to obtain the next action instruction; Based on the next action instructions and user status feedback, a continuous navigation instruction stream is generated.

2. The method for real-time simulation and evacuation route planning of flash flood disasters using a combined air-space-ground approach, as described in claim 1, is characterized in that... The raw sensor data includes satellite remote sensing imagery, drone video streams, rainfall station data, and water level station data.

3. The method for real-time simulation and evacuation route planning of flash flood disasters using a combined air-space-ground approach, as described in claim 2, is characterized in that... Based on the flood projection model, dynamic risk field perception and time-series prediction are performed on the raw sensor data to obtain a series of risk prediction maps for the next N time steps, including: Water body boundary extraction is performed on satellite remote sensing imagery and UAV video streams to obtain water body boundary vectors; Inundation depth inversion is performed based on water body boundary vectors, water level station data, and digital elevation model data to obtain an inundation depth map at the current moment. Spatial interpolation of the rainfall field is performed based on rainfall station data to obtain the rainfall field map at the current moment; The current inundation depth map, the current rainfall field map, digital elevation model data, and future rainfall forecast data are input into the pre-trained flood extrapolation model to perform recursive time-series predictions to obtain a series of water depth prediction maps for the next N time steps. Based on digital elevation model data, dynamic risk quantification and spatiotemporal risk field generation are performed on a series of water depth prediction maps for the next N time steps to obtain the series of risk prediction maps for the next N time steps.

4. The method for real-time simulation and evacuation route planning of flash flood disasters using a combined air-space-ground approach, as described in claim 3, is characterized in that... The pre-trained flood projection model has a CNN-LSTM architecture.

5. The method for real-time simulation and evacuation route planning of flash flood disasters using a combined air-space-ground approach, as described in claim 1, is characterized in that... The current macro node, the refuge macro node, the high-level abstract graph, and a series of risk prediction graphs for the next N time steps are input into a pre-trained high-level policy network to obtain sub-target macro nodes, including: Macro-path time-series risk cost aggregation is performed on a series of risk prediction maps and high-level abstract maps for the next N time steps to obtain a macro-edge time-series cost matrix; A state vector is constructed based on the macro-edge time-series cost matrix, the current macro node, the macro node where the shelter is located, and the high-level abstract graph. The state vector is input into a pre-trained high-level policy network to obtain the sub-target macro nodes.

6. The method for real-time simulation and evacuation route planning of flash flood disasters using a combined air-space-ground approach, as described in claim 5, is characterized in that... Based on the macro-edge time-series cost matrix, the current macro node, the macro node where the shelter is located, and the high-level abstract graph, a state vector is constructed, including: Encode the location information of the current macro node and the macro node where the shelter is located to obtain the location information encoding vector of the current macro node and the location information encoding vector of the macro node where the shelter is located. Based on the high-level abstract graph and the macro-edge temporal cost matrix, the action space cost is encoded for the current macro node to obtain the action cost encoding vector of the current macro node. The state vector is obtained by concatenating the encoded vectors of the current macro node location information, the macro node location information, and the macro node action cost.

7. The method for real-time simulation and evacuation route planning of flash flood disasters using a combined air-space-ground approach, as described in claim 5, is characterized in that... Inputting the state vector into a pre-trained high-level policy network to obtain the sub-target macro nodes includes: The state vector is input into a pre-trained high-level policy network to obtain the action Q-value vector; Optimal action selection is performed based on the action Q-value vector to obtain the optimal action index; Based on the optimal action index, subtarget mapping is performed on the list of actual neighboring nodes of the current macro node to obtain the subtarget macro node.

8. The method for real-time simulation and evacuation route planning of flash flood disasters using a combined air-space-ground approach, as described in claim 1, is characterized in that... The current precise location, the current macro-node region risk prediction map, and the current macro-node region low-level local map are input into the pre-trained low-level policy network to obtain the next action instructions, including: The current precise location, the current macro node region risk prediction map, and the current macro node region bottom-level local map are input into the pre-trained bottom-level policy network to obtain the action probability distribution. Select the action with the highest probability value from the action probability distribution as the selected action; The selected action is converted into an instruction to obtain the instruction for the next action.

Images