A mountainous area tourism route planning method based on weather and natural disasters

Abstract

The application discloses a mountainous area tourism route planning method based on weather and natural disasters, comprising the following steps: step one, a Markov game model is constructed based on POMDP, a dynamic risk matrix integrating geological disaster probability, weather trend and ecological sensitivity is designed by combining Mamba multi-modal feature compression and RAG semantic enhancement, and partial observable environment state perception is provided; step two, a Nash equilibrium constraint optimization model is established, extreme weather strategies are shared through federated learning, an NSGA-II-Actor-Critic architecture is introduced to optimize unmanned aerial vehicle inspection and path adjustment, and dynamic communication is triggered when rainfall exceeds a threshold; step three, a reward function is designed by fusing safety and experience benefits, an improved DDPG is used to process continuous action planning, and NSGA-II is combined to accelerate the convergence of the multi-objective Pareto frontier; step four, a user weather sensitivity portrait is constructed, the weight balance benefit is dynamically adjusted, and an AR simulation preview is preferred; and step five, an agent and environment features are complementary based on a GIS closed-loop decision engine, a blockchain records emergency logs, and a meta-learning constructs a disaster case library.

Description

Technical Field

[0001] This invention relates to the field of route planning technology, and in particular to a method for planning mountain tourism routes based on weather and natural disasters. Background Technology

[0002] In recent years, mountain tourism has rapidly emerged as a major growth engine in the global tourism market due to its unique natural landscapes and adventure experiences. However, the complexity of the mountain environment, such as steep terrain, variable climate, and ecological fragility, poses a serious challenge to ensuring the safety of tourists.

[0003] Over time, the shortcomings of traditional route planning methods in terms of dynamic risk perception have become increasingly apparent. These methods rely heavily on static geological maps and historical weather data, lacking real-time awareness of geological disaster probabilities, sudden weather changes, and ecological sensitivity, leading to delayed decision-making. For example, in 2023, during continuous rainfall in a mountainous scenic area in Sichuan, the traditional system failed to detect the expansion of local mountain cracks in time, resulting in three tourists being blocked by a landslide while following the recommended route and forced to remain in the danger zone for up to six hours. Therefore, this paper proposes a mountain tourism route planning method based on weather and natural disasters. This method constructs a closed-loop framework that integrates a dynamic risk matrix with multi-agent game decision-making and introduces a user profile-driven dynamic weight adjustment mechanism to improve dynamic risk perception and achieve real-time and accurate balance between safety and experience in mountain tourism routes. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a method for planning mountain tourism routes based on weather and natural disasters.

[0005] The core concepts of this invention are as follows: First, by integrating multimodal perception data and semantic enhancement technology, a dynamically updated risk assessment matrix is ​​constructed, solving the problem of lag in static environment model perception; Second, a collaborative decision-making mechanism based on Nash equilibrium constraints and NSGA-II multi-objective optimization is designed to ensure the decision-making efficiency and stability of the multi-agent system in complex environments; Third, a user profile-driven dynamic weight adjustment algorithm is introduced to objectively quantify subjective preferences, achieving a precise balance between security and user experience; Finally, a closed-loop decision-making and evolution framework is constructed through GIS, blockchain, and meta-learning technologies, enhancing the system's autonomy and adaptability.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A method for planning mountain tourism routes based on weather and natural disasters, comprising:

[0008] Step 1 (Dynamic Risk Assessment and State Awareness Module): Based on the POMDP model, a Markov game-theoretic decision-making model for multi-agent collaborative optimization is constructed. This model integrates natural disaster monitoring data (such as landslide displacement sensors and rainfall radar) and ecological environment indicators (vegetation coverage and soil moisture) from along transportation routes and tourist destinations. Combining the Mamba multimodal feature representation learning method, and after multimodal feature compression and RAG semantic enhancement, a dynamic risk matrix (outputting G, W, E, R) is constructed. This provides a state awareness foundation for the multi-agent system under partially observable environments.

[0009] Step Two (Multi-Agent Collaborative Optimization Module): Optimize the UAV inspection strategy through federated learning and the NSGA-II algorithm, establishing an optimization target model that satisfies Nash equilibrium constraints. A dedicated agent is designed for weather perception needs, and a lightweight YOLOv8 model is deployed to identify weather phenomena. Simultaneously, a communication and feature sharing mechanism among multiple agents is proposed. The NSGA-II algorithm is introduced into the Actor-Critic architecture to optimize the UAV inspection frequency and path adjustment strategy, ensuring that a dynamic communication protocol is automatically triggered when rainfall intensity exceeds a threshold, thereby improving the efficiency of collaborative perception in a high-dimensional state space.

[0010] Step 3 (Path Planning Decision Module): Design a reward function that integrates safety and user experience benefits, and output the final path instruction (θ, v, α) using the improved DDPG algorithm. This module directly supports the claims for "Improved DDPG for handling continuous action planning" and "Designing a reward function that integrates safety and user experience benefits".

[0011] Step Four (User Profiling and Adjustment Module): Based on historical user behavior data, user profiles are constructed, and dynamic weight adjustments and AR simulation previews are implemented. The environment and user feedback loop in the diagram illustrate the characteristics of "dynamically adjusting weights to balance benefits" and "AR simulation preview preferences." When a user selects a specific mode, the weight of the corresponding benefit item is automatically increased. An AR weather simulation function is developed to preview preferences. Finally, a joint game theory approach is used to generate an "observation-behavior" mapping strategy, strengthening geographic awareness and recommendation capabilities.

[0012] Step 5 (Closed-Loop Management and Learning Module): Based on the GIS geographic context reasoning framework, a closed-loop decision engine with "perceptual input - cognitive output" capability is formed. Feature complementarity is achieved through the feedback mechanism between the intelligent agent and the environment. Blockchain technology is introduced to record emergency decision logs, and a disaster emergency case library is built in combination with the meta-learning framework, enabling the system to quickly adapt to new climate regions. The "reward-benefit" balance mechanism is used to complete spatiotemporal evolution intelligent reasoning, improve the robustness of group decision-making in continuous action space, and provide tourists with safer and more sustainable mountain tourism service guarantees.

[0013] The above technical solution further includes:

[0014] Furthermore, the design of the dynamic risk assessment matrix includes the following steps:

[0015] Deploy a geological monitoring sensor network (crack displacement gauges, pore water pressure gauges, etc.) to collect data related to disasters such as landslides and debris flows; access meteorological radar and satellite cloud image data to obtain weather indicators such as rainfall intensity, visibility, and wind speed; and unify data from different sources into the WGS84 coordinate system through GIS geographic registration for spatial alignment.

[0016] A state-space model is used to process time series data, constructing an implicit representation of continuous time series.

[0017] Formula expression: Let the geological sensor sequence be... Mapped to the latent space by a linear transformation W:

[0018]

[0019] in, Let be the hidden state vector at time step t; The linear transformation matrix of the input data; Let be the linear transformation matrix of the historical hidden states; It is a bias vector with dimension 64;

[0020] The final output is a 64-dimensional geological feature vector (original data dimension 1024 → 64).

[0021] Construct a disaster domain knowledge graph to map text reports (such as "local cracks in the mountain are widening") into structured indicators; use LLM for event extraction, for example:

[0022] Input text: "A new crack, 3cm wide, was found on the northeast side of monitoring point A".

[0023] Output structured data: {"Crack displacement": 0.03, "Location": "Northeast side of monitoring point A", "Risk level": "High"};

[0024] Define matrix dimensions:

[0025] Geological disaster probability (G): 0-100% (based on historical data and real-time sensor trends);

[0026] Weather risk level (W): 1-5 (Level 1 is safe, Level 5 is extreme weather).

[0027] Ecological sensitivity (E): 0-100% (high sensitivity is triggered when vegetation cover is <30%).

[0028] Formula for calculating the comprehensive risk value:

[0029]

[0030] A red alert is triggered when R ≥ 75;

[0031] Deploy the lightweight ONNX inference engine on a Raspberry Pi 4B device:

[0032] Data preprocessing: 10ms / frame

[0033] Feature compression: 8ms / frame

[0034] Risk calculation: 2ms / frame

[0035] The update cycle is strictly controlled within 5 minutes to ensure timely disaster response; the dimensional weights are automatically adjusted according to seasonal changes.

[0036] By compressing high-dimensional monitoring data into computable quantitative indicators and combining it with real-time semantic enhancement technology, a risk assessment system with dynamic adaptability was constructed, providing accurate environmental state input for subsequent multi-agent collaborative decision-making.

[0037] Furthermore, the construction of the communication and feature sharing mechanism among multiple agents, by introducing the NSGA-II algorithm into the Actor-Critic architecture, includes the following steps:

[0038] Based on the dynamic risk matrix output in step one, the communication threshold is set as follows: triggered when the comprehensive risk value R > 60 or the rainfall intensity > 20 mm / h. When the threshold is triggered, the agent automatically switches to high-frequency communication mode; a hybrid communication topology is adopted, using distributed communication under normal conditions (reducing the pressure on the central node), and automatically switching to centralized communication during disaster warnings (ensuring command consistency); the high-definition images collected by the UAV are compressed using the JPEG2000 compression algorithm, compressing the data volume to 15% of the original size while maintaining 95% image quality.

[0039] Each agent (such as a drone or ground sensor) trains a lightweight YOLOv8 model on a local dataset, outputting a feature vector with a dimension of 256. Local data privacy is protected using differential privacy techniques (adding noise with a variance σ=0.1). Model parameters are aggregated on the server side using the FedAvg algorithm.

[0040]

[0041] in For the local model parameters of the k-th agent, | | represents the local data volume;

[0042] The maximum mean difference (MMD) metric is used to measure the consistency of feature distributions, ensuring that the difference in the feature space after aggregation is less than 0.05.

[0043] The visual features of the UAV (256-dimensional), geological sensor features (64-dimensional), and meteorological data features (32-dimensional) are projected into a unified 128-dimensional space; the importance of each feature is calculated using a self-attention mechanism.

[0044]

[0045]

[0046] in, Let be the attention weight for the i-th feature. Let be the energy score of the i-th feature, W be the learnable weight matrix, and b be the bias term. This is the transposed query vector. Let i be the i-th feature vector, and the final fused features are: ;

[0047] NSGA-II-Actor-Critic Multi-Objective Optimization:

[0048] Safety benefits (G represents the probability of geological disasters);

[0049] Efficiency Benefits (v represents the actual inspection speed)

[0050] Ecological benefits (E represents ecological sensitivity)

[0051] The Actor network generates an initial policy π(s|θ), then NSGA-II performs crossover and mutation operations on θ to generate a set of candidate policies. Finally, the Critic network evaluates the Q-value of each policy.

[0052] The winning strategy is selected through non-dominated sorting and crowding calculation, and the Actor parameters are updated.

[0053] Introduce constraints during policy updates:

[0054]

[0055] in, Let i be the reward function for the i-th agent; The changed policy for the i-th agent; It is a set of policies for other agents; ensuring that no agent can gain a higher benefit by unilaterally changing its policy.

[0056] The UAV adjusts its inspection path based on the fusion feature F. When a new risk point is detected, it sends a priority interruption signal to the ground agent via the V2X protocol. It uses cognitive radio technology to automatically switch between the 2.4GHz and 5GHz frequency bands to avoid communication congestion (switching is triggered when the channel utilization rate is >85%).

[0057] Furthermore, the design integrates a reward function that combines safety and experience benefits, and uses an improved DDPG algorithm to handle path planning problems in a continuous action space, including the following steps:

[0058] Safety benefits:

[0059] Geological disaster risk: (G is the probability of geological disaster output in step one);

[0060] Weather risk: (W represents the weather risk level);

[0061] Ecological sensitivity: (E represents ecological sensitivity)

[0062] Overall security benefits:

[0063] Experience benefits:

[0064] Landscape aesthetic value: The landscape along the route, such as terraced fields and sea of ​​clouds, is scored using a pre-trained CNN model;

[0065] User preference matching degree: Based on the weather sensitivity profile generated in step four, calculate the similarity between the path and user preferences, such as the weighted preference of "rain scene shooting type" users for rain and fog landscapes.

[0066] Overall experience benefits: (A represents aesthetic rating, and S represents preference similarity)

[0067] Total reward function:

[0068] Where α is a dynamic weighting coefficient, adjusted according to the real-time risk value:

[0069] when When α < 0.7, α = 0.7 (forced priority safety).

[0070] when When the value is ≥0.7, α=0.3 (prioritizing trial and error).

[0071] The Actor network takes the dynamic risk matrix and user preference vector output from step one as input and outputs continuous actions (path direction angle θ∈[0°, 360°) and velocity v∈[0, 8m / s]); the Critic network uses dual Q-value estimation to reduce overestimation bias, takes state-action pairs as input, and outputs Q-value estimates; adaptive Ornstein-Uhlenbeck noise is introduced, and the noise amplitude decreases with training epochs.

[0072] (Initial noise σ0=0.5, attenuation rate λ=0.001)

[0073] Based on the TD error calculation sampling priority, the sampling probability of samples with large errors is increased by 30%; the experience data of different agents are stored in a shared replay pool to break the data isolation;

[0074] A feasible area mask is generated using DEM elevation data to prevent the path from crossing dangerous areas such as cliffs; when communication between the UAV and the ground intelligent agent is interrupted, a conservative path mode is automatically triggered (speed is reduced by 50%, and the heading angle is limited to ±30°).

[0075] When the overall risk value R output in step one is greater than 75, α is forcibly set to 0.7, and the route planning prioritizes avoiding high-risk areas; when the user selects the "landscape priority" mode, α is dynamically adjusted to 0.3, allowing the route to detour to areas with high aesthetic scores, such as cloud-shrouded ridgelines; the hybrid mode defaults to α=0.5, balancing safety and experience.

[0076] Deploying an asynchronous Actor-Learner architecture on four GPUs improves training speed by 3 times; the pre-trained model is initialized in a simulated mountain environment, reducing actual training time; a hard constraint layer is added after the Actor output to forcibly exclude all illegal actions that violate physical rules (such as aerial paths crossing mountains).

[0077] Furthermore, the multi-objective optimization model established by integrating NSGA-II and the Actor-Critic architecture accelerates Pareto front convergence and generates joint game decision results, including the following steps:

[0078] 50% of the strategies were randomly generated through the Actor network, and 50% were generated through NSGA-II crossover mutation. The top 10% of high-fitness strategies were retained as the initial population.

[0079] Based on the current state s, output continuous actions a (path direction angle θ and velocity v); use SBX crossover (distribution exponent η=15) to generate offspring policies; apply multinomial mutation (mutation probability PM=0.1) to explore the new policy space; calculate the Q value of each policy: Pareto fronts are divided into levels based on Q-values ​​and constraints (such as path feasibility); within the same front, strategies with greater diversity are selected using distance metrics.

[0080] The policy parameter θ is updated along the Q-value gradient direction using the policy gradient method:

[0081]

[0082] The strategy of retaining the top 20% of elites is combined with newly generated strategies to form the next generation of the population;

[0083] When the overall risk value R ≥ 75, the weight of safety benefits increases by 30%; when a "landscape priority" request is detected, the weight of experience benefits increases by 20%.

[0084] Under the federated learning framework, each agent shares the top 5% of elite policies, and policy privacy is protected through differential privacy (noise variance σ=0.05); the shared policies are adjusted by local fitness and then re-participate in NSGA-II evolution.

[0085] A repeated game mechanism is adopted, in which each agent proposes a strategy in turn, and the Critic evaluates whether equilibrium has been reached; when the rate of change of all agents' payoffs is less than 2% after 5 consecutive rounds of strategy updates, the game equilibrium is determined to have been reached.

[0086] The final decision is a triplet containing the path direction angle θ, velocity v, and risk aversion coefficient α.

[0087]

[0088] The decision simultaneously satisfies multi-objective optimization and Nash equilibrium constraints;

[0089] Deploying asynchronous NSGA-II evolution on a GPU cluster reduces the policy evaluation time per round to 0.3 seconds; adding a hard constraint layer after the Actor output forces the exclusion of all illegal actions that violate physical rules, such as crossing mountains; and dynamically adjusting the weights of multiple objectives based on real-time risk values ​​and user preferences.

[0090] Furthermore, the process of constructing a user weather sensitivity profile and generating preference tags by combining historical behavioral data, employing a dynamic weight adjustment algorithm, includes the following steps:

[0091] Collect user historical behavior data, including: travel records (trip dates, route selection, stay duration), weather feedback (user-initiated "comfortable / uncomfortable" weather type), and social media comments (such as weather-related descriptions like "great photos taken in the rain"); remove irrelevant information (such as advertisements and emojis) through text cleaning, and retain key weather-related expressions;

[0092] User-selected weather modes (e.g., "rain-preferred") and trip adjustment records (number of trips canceled due to rain) are analyzed using NLP to assess the sentiment in user reviews. For example, "low visibility due to heavy fog, poor experience" is marked as negative sentiment. The K-means algorithm is used to categorize users into three groups:

[0093] Rain-sheltered type: Over 70% of the historical itineraries have had their routes adjusted due to rainfall;

[0094] Rain scene photography type: Actively choose rainy day trips and the keyword "rain scene" appears more than 5 times in the comments;

[0095] Climate-adaptive type: No obvious weather preference, and the correlation between itinerary selection and weather is less than 30%;

[0096] After each user completes a trip, real-time feedback is collected through questionnaires to update user profile tags; if a user's behavioral data has not been updated for 3 months, the tag weight decreases by 10% each month.

[0097] Weight adjustment trigger conditions:

[0098] User-initiated selection: When the user explicitly selects "Rain Scene Priority" or "Safety Priority" mode through the APP interface;

[0099] Environmental change trigger: When the comprehensive risk value R output in step one is greater than or equal to 75, the system will be forced to enter the safety priority mode.

[0100] Hybrid mode: When there are no explicit user instructions and the environment is safe, the default balanced mode is used;

[0101] The weighting adjustment rules are as follows: for users seeking shelter from rain, the safety benefit weighting is increased to 70% (from the default of 50%), the experience benefit weighting is reduced to 30%, and route planning is prohibited from passing through areas with a rainfall probability >40%.

[0102] The weighting of user experience benefits for rain scene photography has been increased to 60% (from the default of 50%), while the weighting of safety benefits has been reduced to 30%. Path detours to areas with a rainfall probability >50% are now allowed (provided that the ecological sensitivity is <50%).

[0103] Climate-adaptive users maintain the default weights (50% safety, 50% experience), and are only forced to switch to safety mode when the environmental risk value R ≥ 80;

[0104] The weighting adjustment is constrained to ensure that the safety benefit weight is no less than 30% (to ensure a basic safety baseline) and the experience benefit weight is no more than 60% (to avoid excessive pursuit of experience while ignoring risks). After the weighting adjustment, the feasibility of the strategy must be verified through the Critic network, such as whether the path is passable.

[0105] After the user selects a preferred mode, AR technology is used to display the weather effects along the route in real time: when the "Rain Scene Priority" mode is selected, the virtual rainfall effect and possible shooting locations are displayed; when the "Safety Priority" mode is selected, areas at risk of geological disasters are highlighted. Users can adjust the simulation parameters, such as rainfall intensity and visibility, through gesture interaction.

[0106] Users can rate the display effect in the simulation interface, and the rating data is fed back to the profile update module in real time to adjust the subsequent weight allocation strategy.

[0107] Furthermore, the GIS-based geographic context reasoning framework forms a closed-loop decision engine, which complements features through feedback between the agent and the environment, including the following steps:

[0108] Integrate high-precision DEM (Digital Elevation Model) data, vegetation cover layer (NDVI index), water system distribution layer, and historical geological disaster point data; and generate a composite layer containing slope (greater than 30° marked as dangerous), catchment area (area greater than 1km²), and ecological red line area (prohibited development area) through GIS spatial analysis tools.

[0109] The dynamic risk matrix (geological, weather, and ecological dimensions) from step one is overlaid with the GIS layer to generate a real-time geographic scenario model; the model is updated every 5 minutes, keeping pace with the data refresh rate in step one.

[0110] The terrain is rendered using a 3D GIS engine, highlighting high-risk areas (red), user-preferred paths (blue), and the agent's current location (green marker); AR technology is used to overlay virtual paths onto the real environment, allowing users to intuitively view the effects of their decisions.

[0111] The joint game decision (path direction angle θ, velocity v, risk avoidance coefficient α) generated in step 3 is sent to the UAV and ground intelligent agent; during the execution of the path by the intelligent agent, environmental data is collected in real time, such as new cracks and changes in rainfall intensity; the environmental change data is sent back to the decision engine via the V2X protocol to update the GIS geographic context model.

[0112] Based on the updated model, the path planning parameters are recalculated to form a closed-loop adjustment.

[0113] Dynamically adjust trigger conditions:

[0114] Environmental mutation: When a geological disaster probability G ≥ 80% or a weather risk level W ≥ 4 is detected, route replanning will be forcibly triggered;

[0115] User intervention: When a user manually adjusts their route preferences via the app (such as "avoid the current area"), a new decision is generated immediately;

[0116] Periodic optimization: Lightweight path optimization is automatically performed every 15 minutes to balance computing resources and decision-making timeliness;

[0117] The drone perspective provides a high-altitude macroscopic view to identify large-scale geological changes (such as the expansion of cracks in mountains); ground sensors provide microscopic data such as soil moisture and surface displacement; and meteorological station data supplements local weather changes such as sudden gusts and sharp drops in visibility.

[0118] By using GIS coordinate transformation, UAV orthophotos and ground sensor data are unified into the same coordinate system; the PTP clock synchronization protocol is adopted to ensure that the timestamp error of multi-source data is less than 10ms; and weights are assigned according to the reliability of the data source, such as 0.8 for LiDAR data and 0.6 for visual data.

[0119] Areas of geological change identified by drones need to be confirmed by ground sensor data. For example, if the displacement of cracks is greater than 5mm, secondary verification is triggered. When a single data source shows an anomaly, such as a sudden drop in temperature of a sensor but normal operation of surrounding sensors, a self-check is initiated and the data is marked as suspicious.

[0120] Each decision's input data (environmental status), output actions (path parameters), and execution results (user feedback) are packaged and uploaded to the blockchain; the on-chain data is stored using a Merkle tree structure to ensure data integrity and traceability.

[0121] Historical decision-making cases are extracted from the blockchain to construct a case library containing typical scenarios (such as "mountain patrol during rainy season" and "user landscape priority mode"). The MAML (Model-Independent Meta-Learning) algorithm is employed, with its policy network structure identical to the Actor network in the improved DDPG algorithm. During the meta-training phase, the model learns to quickly adapt to different disaster scenario tasks (such as landslides under different rainfall levels and path adjustments under different wind speeds), with its loss function being the negative expected value of the cumulative reward for each task. Rapid adaptation to new scenarios using a small number of samples reduces online learning time. Combined with GIS spatiotemporal analysis tools, environmental change patterns are identified, such as "the probability of a landslide increases by 30% after 5 consecutive days of rainfall." These spatiotemporal patterns are encoded into a rule engine to assist the decision engine in predicting environmental trends.

[0122] Furthermore, the introduction of blockchain technology to record emergency decision logs and the construction of a disaster emergency case library using a meta-learning framework includes the following steps:

[0123] The system uses environmental state snapshots to record the dynamic risk matrix (geological disaster probability G, weather risk level W, ecological sensitivity E) and GIS geographic context model during decision-making; decision action details include path direction angle θ, speed v, risk avoidance coefficient α, and multi-agent communication protocol status; execution result feedback includes actual user travel data (such as path completion rate, stop point selection), environmental changes monitored by the system, such as new cracks and sudden changes in rainfall intensity; user intervention records are user manual adjustments to path preferences (such as "avoid current area") and system responses;

[0124] Each decision is fully logged and encapsulated as a blockchain transaction, with timestamps and digital signatures added; data is organized using a Merkle tree structure to ensure data integrity and rapid verification capabilities.

[0125] Access control:

[0126] Write permissions: Only for the decision engine and user terminals (signed via private key);

[0127] Read access: Granted to authorized agencies (such as scenic area management departments) for post-event analysis;

[0128] The blockchain explorer allows users to query decision logs at any point in time, supporting multi-dimensional searches by user ID, time range, risk level, and more. In the event of a geological disaster, it can quickly pinpoint the decision-making path and environmental conditions up to 5 minutes before the incident, aiding in liability determination.

[0129] Extract historical decision-making cases from the blockchain and categorize them by scenario type:

[0130] Typical disaster scenarios: such as "landslide warning after 7 consecutive days of rainfall" and "typhoon path adjustment";

[0131] User preference scenarios: such as "rain scene photography users' trips" and "safety-first users' risk avoidance";

[0132] Hybrid scenarios: such as "users prioritizing the landscape in high-risk environments";

[0133] Each case study is labeled with key features including environmental risk value R, user preference tags, decision-making action sequence, and execution result score.

[0134] Employing the MAML (Model Independent Meta-Learning) algorithm, it can quickly adapt to new scenarios using a small number of samples:

[0135] Meta-training phase: Simulate different disaster scenarios in the historical case library to train the model to learn "how to learn quickly";

[0136] Fine-tuning phase: When encountering new scenarios, such as the first appearance of a "hail + dense tourist" composite scenario, only 10-20 samples are needed to adjust the model parameters;

[0137] The matching degree between the new scenario and the case library is measured by Euclidean distance. When the similarity is greater than 85%, the historical strategy is directly invoked.

[0138] When a new risk scenario is detected, the three most similar cases in the case library are quickly retrieved to generate a candidate strategy set; the Q-value of the candidate strategies is evaluated through the Critic network, and the optimal strategy is selected and sent to the agent; after each emergency decision is executed, the new case is added to the case library, and outdated cases that have not been accessed for more than two years are eliminated.

[0139] The decision logs recorded by the blockchain provide real-world data for meta-learning, and the strategies optimized by meta-learning are traceable and verifiable through the blockchain. When the meta-learning model generates a new strategy, the strategy version number and training sample set need to be recorded in the blockchain to ensure decision transparency.

[0140] User behavior data, such as path selection and preference tags, is anonymized before being stored in the blockchain, such as by hashing user IDs. During meta-learning training, a federated learning framework is used, where each agent retains its original data locally and only shares the amount of model parameter updates.

[0141] The present invention has the following beneficial effects:

[0142] 8. Real-time and accurate perception: By adopting a dynamic risk matrix construction method based on Mamba multimodal feature compression and RAG semantic enhancement, the system achieves minute-level (within 5 minutes) updates and fusion assessments of geological hazards, weather changes and ecological sensitivity. Compared with traditional methods that rely on static maps, this significantly improves the timeliness and accuracy of risk perception.

[0143] 9. Efficient Collaborative Decision Making: By proposing a communication mechanism with the NSGA-II-Actor-Critic hybrid architecture and Nash equilibrium constraints, UAVs and ground agents can not only efficiently approach the Pareto front of multiple objectives such as safety, efficiency, and ecology in path planning, but also avoid policy conflicts between agents, ensuring the collaboration and stability of system decision making.

[0144] 10. Personalized Experience Optimization: By constructing a user weather sensitivity profile and designing a dynamic weight adjustment algorithm, the system can understand and quantify users' subjective preferences, thereby generating travel routes that truly meet users' personalized needs while ensuring basic safety, thus improving the user experience.

[0145] 11. System Autonomy and Evolution: Due to the closed-loop decision engine formed by the GIS geographic context reasoning framework and the introduction of blockchain evidence storage and meta-learning case library, the system has the ability to continuously learn and evolve from historical decisions, and can constantly adapt to new climate environments and disaster scenarios, thereby enhancing the robustness and long-term effectiveness of the system. Attached Figure Description

[0146] Figure 1 This is a schematic diagram illustrating the system working principle and process of a mountain tourism route planning method based on weather and natural disasters proposed in this invention. Detailed Implementation

[0147] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0148] Please see Figure 1 As shown in the system flowchart, this invention is a method for planning mountain tourism routes based on weather and natural disasters, including:

[0149] Step 1 (Dynamic Risk Assessment and Status Awareness Module): For example... Figure 1 The above describes the input of multi-source data (geological, meteorological, and ecological). Based on the POMDP model, a Markov game-theoretic decision-making model for multi-agent collaborative optimization is constructed. By integrating natural disaster monitoring data (such as landslide displacement sensors and rainfall radar) and ecological environment indicators (vegetation coverage and soil moisture) from transportation routes and tourist destinations, and combining the multimodal feature representation learning method implemented by Mamba, the model undergoes multimodal feature compression and RAG semantic enhancement processing to finally construct a dynamic risk matrix (output G, W, E, R). This provides a state perception basis for multi-agent systems in partially observable environments.

[0150] Step Two (Multi-Agent Cooperative Optimization Module): For example... Figure 1As shown in the upper left corner, this describes a communication protocol triggered by a risk matrix, and the optimization of UAV inspection strategies through federated learning and the NSGA-II algorithm to establish an optimized objective model that satisfies Nash equilibrium constraints. A dedicated agent is designed for weather perception needs, and a lightweight YOLOv8 model is deployed to identify weather phenomena such as rain, fog, and hail. Furthermore, a communication mechanism and feature sharing mechanism among multiple agents are proposed. The NSGA-II algorithm is introduced into the Actor-Critic architecture to optimize the UAV inspection frequency and path adjustment strategy, ensuring that a dynamic communication protocol is automatically triggered when rainfall intensity exceeds a threshold, thereby improving the efficiency of collaborative perception in a high-dimensional state space.

[0151] Step 3 (Path Planning Decision Module): For example... Figure 1 As shown in the middle section, it describes the process of receiving risk data and collaborative strategies, designing a reward function that integrates security and experience benefits, and outputting the final path instruction (θ, v, α) through an improved DDPG algorithm. This module directly supports the claims for "improved DDPG processing of continuous action planning" and "designing a reward function that integrates security and experience benefits".

[0152] Step Four (User Profile and Adjustment Module): For example... Figure 1 As shown in the lower right corner, this diagram illustrates the process of building user profiles based on historical behavior data, dynamically adjusting weights, and performing AR simulation previews. The environmental feedback and user feedback loops in the diagram demonstrate the technical features of "dynamically adjusting weights to balance benefits" and "AR simulation preview preferences." When a user selects a specific mode, the weight of the corresponding benefit item is automatically increased, and an AR weather simulation function is developed to preview preferences. Finally, a joint game theory approach is used to generate an "observation-behavior" mapping strategy, strengthening geographic awareness and recommendation capabilities.

[0153] Step 5 (Closed-Loop Management and Learning Module): For example... Figure 1 As shown in the lower left corner, this module describes the closed-loop process of path execution, environmental feedback, GIS context model updates, blockchain notarization, and the construction of a meta-learning case library. This module fully embodies three core features: a GIS-based closed-loop decision engine, blockchain-recorded emergency logs, and meta-learning-built disaster case library. Specifically, based on a GIS geographic context reasoning framework, a closed-loop decision engine with "perceptual input—cognitive output" capabilities is formed, achieving feature complementarity through a feedback mechanism between the agent and the environment. Blockchain technology is introduced to record emergency decision logs, and a disaster emergency case library is constructed using a meta-learning framework, enabling the system to quickly adapt to new climate regions. A "reward-benefit" balancing mechanism is used to complete spatiotemporal evolution intelligent reasoning, improving the robustness of group decision-making in continuous action spaces and providing tourists with safer and more sustainable mountain tourism service guarantees.

[0154] In one embodiment, designing the dynamic risk assessment matrix includes the following steps:

[0155] Deploy a geological monitoring sensor network (crack displacement gauges, pore water pressure gauges, etc.) to collect data related to disasters such as landslides and debris flows; access meteorological radar and satellite cloud image data to obtain weather indicators such as rainfall intensity, visibility, and wind speed; and unify data from different sources into the WGS84 coordinate system through GIS geographic registration for spatial alignment.

[0156] A state-space model is used to process time series data, constructing an implicit representation of continuous time series.

[0157] Formula expression: Let the geological sensor sequence be... Mapped to the latent space by a linear transformation W:

[0158]

[0159] in, Let be the hidden state vector at time step t; The linear transformation matrix of the input data; Let be the linear transformation matrix of the historical hidden states; It is a bias vector with dimension 64;

[0160] The final output is a 64-dimensional geological feature vector (original data dimension 1024 → 64).

[0161] Construct a disaster domain knowledge graph to map text reports (such as "local cracks in the mountain are widening") into structured indicators; use LLM for event extraction, for example:

[0162] Input text: "A new crack, 3cm wide, was found on the northeast side of monitoring point A".

[0163] Output structured data: {"Crack displacement": 0.03, "Location": "Northeast side of monitoring point A", "Risk level": "High"};

[0164] Define matrix dimensions:

[0165] Geological disaster probability (G): 0-100% (based on historical data and real-time sensor trends);

[0166] Weather risk level (W): 1-5 (Level 1 is safe, Level 5 is extreme weather).

[0167] Ecological sensitivity (E): 0-100% (high sensitivity is triggered when vegetation cover is <30%).

[0168] Formula for calculating the comprehensive risk value:

[0169]

[0170] A red alert is triggered when R ≥ 75;

[0171] Deploy the lightweight ONNX inference engine on a Raspberry Pi 4B device:

[0172] Data preprocessing: 10ms / frame

[0173] Feature compression: 8ms / frame

[0174] Risk calculation: 2ms / frame

[0175] The update cycle is strictly controlled within 5 minutes to ensure timely disaster response; the dimensional weights are automatically adjusted according to seasonal changes.

[0176] By compressing high-dimensional monitoring data into computable quantitative indicators and combining it with real-time semantic enhancement technology, a risk assessment system with dynamic adaptability was constructed, providing accurate environmental state input for subsequent multi-agent collaborative decision-making.

[0177] In one embodiment, the construction of a communication and feature sharing mechanism among multiple agents, by introducing the NSGA-II algorithm into the Actor-Critic architecture, includes the following steps:

[0178] Based on the dynamic risk matrix output in step one, the communication threshold is set as follows: triggered when the comprehensive risk value R > 60 or the rainfall intensity > 20 mm / h. When the threshold is triggered, the agent automatically switches to high-frequency communication mode; a hybrid communication topology is adopted, using distributed communication under normal conditions (reducing the pressure on the central node), and automatically switching to centralized communication during disaster warnings (ensuring command consistency); the high-definition images collected by the UAV are compressed using the JPEG2000 compression algorithm, compressing the data volume to 15% of the original size while maintaining 95% image quality.

[0179] Each agent (such as a drone or ground sensor) trains a lightweight YOLOv8 model on a local dataset, outputting a feature vector with a dimension of 256. Local data privacy is protected using differential privacy techniques (adding noise with a variance σ=0.1). Model parameters are aggregated on the server side using the FedAvg algorithm.

[0180]

[0181] in For the local model parameters of the k-th agent, | | represents the local data volume;

[0182] The maximum mean difference (MMD) metric is used to measure the consistency of feature distributions, ensuring that the difference in the feature space after aggregation is less than 0.05.

[0183] The visual features of the UAV (256-dimensional), geological sensor features (64-dimensional), and meteorological data features (32-dimensional) are projected into a unified 128-dimensional space; the importance of each feature is calculated using a self-attention mechanism.

[0184]

[0185]

[0186] in, Let be the attention weight for the i-th feature. Let be the energy score of the i-th feature, W be the learnable weight matrix, and b be the bias term. This is the transposed query vector. Let i be the i-th feature vector, and the final fused features are: ;

[0187] NSGA-II-Actor-Critic Multi-Objective Optimization:

[0188] Safety benefits (G represents the probability of geological disasters);

[0189] Efficiency Benefits (v represents the actual inspection speed)

[0190] Ecological benefits (E represents ecological sensitivity)

[0191] The Actor network generates an initial policy π(s|θ), then NSGA-II performs crossover and mutation operations on θ to generate a set of candidate policies. Finally, the Critic network evaluates the Q-value of each policy.

[0192] The winning strategy is selected through non-dominated sorting and crowding calculation, and the Actor parameters are updated.

[0193] Introduce constraints during policy updates:

[0194]

[0195] in, Let i be the reward function for the i-th agent; The changed policy for the i-th agent; It is a set of policies for other agents; ensuring that no agent can gain a higher benefit by unilaterally changing its policy.

[0196] The UAV adjusts its inspection path based on the fusion feature F. When a new risk point is detected, it sends a priority interruption signal to the ground agent via the V2X protocol. It uses cognitive radio technology to automatically switch between the 2.4GHz and 5GHz frequency bands to avoid communication congestion (switching is triggered when the channel utilization rate is >85%).

[0197] In one embodiment, the design integrates a reward function that combines safety and experience benefits, and uses an improved DDPG algorithm to handle path planning problems in a continuous action space, including the following steps:

[0198] Safety benefits:

[0199] Geological disaster risk: (G is the probability of geological disaster output in step one);

[0200] Weather risk: (W represents the weather risk level);

[0201] Ecological sensitivity: (E represents ecological sensitivity)

[0202] Overall security benefits:

[0203] Experience benefits:

[0204] Landscape aesthetic value: The landscape along the route, such as terraced fields and sea of ​​clouds, is scored using a pre-trained CNN model;

[0205] User preference matching degree: Based on the weather sensitivity profile generated in step four, calculate the similarity between the path and user preferences, such as the weighted preference of "rain scene shooting type" users for rain and fog landscapes.

[0206] Overall experience benefits: (A represents aesthetic rating, and S represents preference similarity)

[0207] Total reward function:

[0208] Where α is a dynamic weighting coefficient, adjusted according to the real-time risk value:

[0209] when When α < 0.7, α = 0.7 (forced priority safety).

[0210] when When the value is ≥0.7, α=0.3 (prioritizing trial and error).

[0211] The Actor network takes the dynamic risk matrix and user preference vector output from step one as input and outputs continuous actions (path direction angle θ∈[0°, 360°) and velocity v∈[0, 8m / s]); the Critic network uses dual Q-value estimation to reduce overestimation bias, takes state-action pairs as input, and outputs Q-value estimates; adaptive Ornstein-Uhlenbeck noise is introduced, and the noise amplitude decreases with training epochs.

[0212] (Initial noise σ0=0.5, attenuation rate λ=0.001)

[0213] Based on the TD error calculation sampling priority, the sampling probability of samples with large errors is increased by 30%; the experience data of different agents are stored in a shared replay pool to break the data isolation;

[0214] A feasible area mask is generated using DEM elevation data to prevent the path from crossing dangerous areas such as cliffs; when communication between the UAV and the ground intelligent agent is interrupted, a conservative path mode is automatically triggered (speed is reduced by 50%, and the heading angle is limited to ±30°).

[0215] When the overall risk value R output in step one is greater than 75, α is forcibly set to 0.7, and the route planning prioritizes avoiding high-risk areas; when the user selects the "landscape priority" mode, α is dynamically adjusted to 0.3, allowing the route to detour to areas with high aesthetic scores, such as cloud-shrouded ridgelines; the hybrid mode defaults to α=0.5, balancing safety and experience.

[0216] Deploying an asynchronous Actor-Learner architecture on four GPUs improves training speed by 3 times; the pre-trained model is initialized in a simulated mountain environment, reducing actual training time; a hard constraint layer is added after the Actor output to forcibly exclude all illegal actions that violate physical rules (such as aerial paths crossing mountains).

[0217] In one embodiment, the fusion of NSGA-II and Actor-Critic architectures to establish a multi-objective optimization model, accelerate Pareto front convergence, and generate joint game decision results includes the following steps:

[0218] 50% of the strategies were randomly generated through the Actor network, and 50% were generated through NSGA-II crossover mutation. The top 10% of high-fitness strategies were retained as the initial population.

[0219] Based on the current state s, output continuous actions a (path direction angle θ and velocity v); use SBX crossover (distribution exponent η=15) to generate offspring policies; apply multinomial mutation (mutation probability PM=0.1) to explore the new policy space; calculate the Q value of each policy: Pareto fronts are divided into levels based on Q-values ​​and constraints (such as path feasibility); within the same front, strategies with greater diversity are selected using distance metrics.

[0220] The policy parameter θ is updated along the Q-value gradient direction using the policy gradient method:

[0221]

[0222] The strategy of retaining the top 20% of elites is combined with newly generated strategies to form the next generation of the population;

[0223] When the overall risk value R ≥ 75, the weight of safety benefits increases by 30%; when a "landscape priority" request is detected, the weight of experience benefits increases by 20%.

[0224] Under the federated learning framework, each agent shares the top 5% of elite policies, and policy privacy is protected through differential privacy (noise variance σ=0.05); the shared policies are adjusted by local fitness and then re-participate in NSGA-II evolution.

[0225] A repeated game mechanism is adopted, in which each agent proposes a strategy in turn, and the Critic evaluates whether equilibrium has been reached; when the rate of change of all agents' payoffs is less than 2% after 5 consecutive rounds of strategy updates, the game equilibrium is determined to have been reached.

[0226] The final decision is a triplet containing the path direction angle θ, velocity v, and risk aversion coefficient α.

[0227]

[0228] The decision simultaneously satisfies multi-objective optimization and Nash equilibrium constraints;

[0229] Deploying asynchronous NSGA-II evolution on a GPU cluster reduces the policy evaluation time per round to 0.3 seconds; adding a hard constraint layer after the Actor output forces the exclusion of all illegal actions that violate physical rules, such as crossing mountains; and dynamically adjusting the weights of multiple objectives based on real-time risk values ​​and user preferences.

[0230] In one embodiment, the process of constructing a user weather sensitivity profile and generating preference labels by combining historical behavioral data, employing a dynamic weight adjustment algorithm, includes the following steps:

[0231] Collect user historical behavior data, including: travel records (trip dates, route selection, stay duration), weather feedback (user-initiated "comfortable / uncomfortable" weather type), and social media comments (such as weather-related descriptions like "great photos taken in the rain"); remove irrelevant information (such as advertisements and emojis) through text cleaning, and retain key weather-related expressions;

[0232] User-selected weather modes (e.g., "rain-preferred") and trip adjustment records (number of trips canceled due to rain) are analyzed using NLP to assess the sentiment in user reviews. For example, "low visibility due to heavy fog, poor experience" is marked as negative sentiment. The K-means algorithm is used to categorize users into three groups:

[0233] Rain-sheltered type: Over 70% of the historical itineraries have had their routes adjusted due to rainfall;

[0234] Rain scene photography type: Actively choose rainy day trips and the keyword "rain scene" appears more than 5 times in the comments;

[0235] Climate-adaptive type: No obvious weather preference, and the correlation between itinerary selection and weather is less than 30%;

[0236] After each user completes a trip, real-time feedback is collected through questionnaires to update user profile tags; if a user's behavioral data has not been updated for 3 months, the tag weight decreases by 10% each month.

[0237] Weight adjustment trigger conditions:

[0238] User-initiated selection: When the user explicitly selects "Rain Scene Priority" or "Safety Priority" mode through the APP interface;

[0239] Environmental change trigger: When the comprehensive risk value R output in step one is greater than or equal to 75, the system will be forced to enter the safety priority mode.

[0240] Hybrid mode: When there are no explicit user instructions and the environment is safe, the default balanced mode is used;

[0241] The weighting adjustment rules are as follows: for users seeking shelter from rain, the safety benefit weighting is increased to 70% (from the default of 50%), the experience benefit weighting is reduced to 30%, and route planning is prohibited from passing through areas with a rainfall probability >40%.

[0242] The weighting of user experience benefits for rain scene photography has been increased to 60% (from the default of 50%), while the weighting of safety benefits has been reduced to 30%. Path detours to areas with a rainfall probability >50% are now allowed (provided that the ecological sensitivity is <50%).

[0243] Climate-adaptive users maintain the default weights (50% safety, 50% experience), and are only forced to switch to safety mode when the environmental risk value R ≥ 80;

[0244] The weighting adjustment is constrained to ensure that the safety benefit weight is no less than 30% (to ensure a basic safety baseline) and the experience benefit weight is no more than 60% (to avoid excessive pursuit of experience while ignoring risks). After the weighting adjustment, the feasibility of the strategy must be verified through the Critic network, such as whether the path is passable.

[0245] After the user selects a preferred mode, AR technology is used to display the weather effects along the route in real time: when the "Rain Scene Priority" mode is selected, the virtual rainfall effect and possible shooting locations are displayed; when the "Safety Priority" mode is selected, areas at risk of geological disasters are highlighted. Users can adjust the simulation parameters, such as rainfall intensity and visibility, through gesture interaction.

[0246] Users can rate the display effect in the simulation interface, and the rating data is fed back to the profile update module in real time to adjust the subsequent weight allocation strategy.

[0247] In one embodiment, the closed-loop decision engine formed based on the GIS geographic context reasoning framework, which complements features through feedback between the agent and the environment, includes the following steps:

[0248] Integrate high-precision DEM (Digital Elevation Model) data, vegetation cover layer (NDVI index), water system distribution layer, and historical geological disaster point data; and generate a composite layer containing slope (greater than 30° marked as dangerous), catchment area (area greater than 1km²), and ecological red line area (prohibited development area) through GIS spatial analysis tools.

[0249] The dynamic risk matrix (geological, weather, and ecological dimensions) from step one is overlaid with the GIS layer to generate a real-time geographic scenario model; the model is updated every 5 minutes, keeping pace with the data refresh rate in step one.

[0250] The terrain is rendered using a 3D GIS engine, highlighting high-risk areas (red), user-preferred paths (blue), and the agent's current location (green marker); AR technology is used to overlay virtual paths onto the real environment, allowing users to intuitively view the effects of their decisions.

[0251] The joint game decision (path direction angle θ, velocity v, risk avoidance coefficient α) generated in step 3 is sent to the UAV and ground intelligent agent; during the execution of the path by the intelligent agent, environmental data is collected in real time, such as new cracks and changes in rainfall intensity; the environmental change data is sent back to the decision engine via the V2X protocol to update the GIS geographic context model.

[0252] Based on the updated model, the path planning parameters are recalculated to form a closed-loop adjustment.

[0253] Dynamically adjust trigger conditions:

[0254] Environmental mutation: When a geological disaster probability G ≥ 80% or a weather risk level W ≥ 4 is detected, route replanning will be forcibly triggered;

[0255] User intervention: When a user manually adjusts their route preferences via the app (such as "avoid the current area"), a new decision is generated immediately;

[0256] Periodic optimization: Lightweight path optimization is automatically performed every 15 minutes to balance computing resources and decision-making timeliness;

[0257] The drone perspective provides a high-altitude macroscopic view to identify large-scale geological changes (such as the expansion of cracks in mountains); ground sensors provide microscopic data such as soil moisture and surface displacement; and meteorological station data supplements local weather changes such as sudden gusts and sharp drops in visibility.

[0258] By using GIS coordinate transformation, UAV orthophotos and ground sensor data are unified into the same coordinate system; the PTP clock synchronization protocol is adopted to ensure that the timestamp error of multi-source data is less than 10ms; and weights are assigned according to the reliability of the data source, such as 0.8 for LiDAR data and 0.6 for visual data.

[0259] Areas of geological change identified by drones need to be confirmed by ground sensor data. For example, if the displacement of cracks is greater than 5mm, secondary verification is triggered. When a single data source shows an anomaly, such as a sudden drop in temperature of a sensor but normal operation of surrounding sensors, a self-check is initiated and the data is marked as suspicious.

[0260] Each decision's input data (environmental status), output actions (path parameters), and execution results (user feedback) are packaged and uploaded to the blockchain; the on-chain data is stored using a Merkle tree structure to ensure data integrity and traceability.

[0261] Historical decision-making cases are extracted from the blockchain to build a case library containing typical scenarios (such as "mountain patrol during the rainy season" and "user landscape priority mode"); the MAML (model-independent meta-learning) algorithm is adopted to quickly adapt to new scenarios with a small number of samples, reducing online learning time; combined with GIS spatiotemporal analysis tools, environmental change patterns are identified, such as "the probability of landslide increases by 30% after 5 consecutive days of rainfall"; spatiotemporal patterns are encoded into a rule engine to assist the decision engine in predicting environmental trends.

[0262] In one embodiment, the introduction of blockchain technology to record emergency decision logs and the construction of a disaster emergency case library in conjunction with a meta-learning framework includes the following steps:

[0263] The system uses environmental state snapshots to record the dynamic risk matrix (geological disaster probability G, weather risk level W, ecological sensitivity E) and GIS geographic context model during decision-making; decision action details include path direction angle θ, speed v, risk avoidance coefficient α, and multi-agent communication protocol status; execution result feedback includes actual user travel data (such as path completion rate, stop point selection), environmental changes monitored by the system, such as new cracks and sudden changes in rainfall intensity; user intervention records are user manual adjustments to path preferences (such as "avoid current area") and system responses;

[0264] Each decision is fully logged and encapsulated as a blockchain transaction, with timestamps and digital signatures added; data is organized using a Merkle tree structure to ensure data integrity and rapid verification capabilities.

[0265] Access control:

[0266] Write permissions: Only for the decision engine and user terminals (signed via private key);

[0267] Read access: Granted to authorized agencies (such as scenic area management departments) for post-event analysis;

[0268] The blockchain explorer allows users to query decision logs at any point in time, supporting multi-dimensional searches by user ID, time range, risk level, and more. In the event of a geological disaster, it can quickly pinpoint the decision-making path and environmental conditions up to 5 minutes before the incident, aiding in liability determination.

[0269] Extract historical decision-making cases from the blockchain and categorize them by scenario type:

[0270] Typical disaster scenarios: such as "landslide warning after 7 consecutive days of rainfall" and "typhoon path adjustment";

[0271] User preference scenarios: such as "rain scene photography users' trips" and "safety-first users' risk avoidance";

[0272] Hybrid scenarios: such as "users prioritizing the landscape in high-risk environments";

[0273] Each case study is labeled with key features including environmental risk value R, user preference tags, decision-making action sequence, and execution result score.

[0274] Employing the MAML (Model Independent Meta-Learning) algorithm, it can quickly adapt to new scenarios using a small number of samples:

[0275] Meta-training phase: Simulate different disaster scenarios in the historical case library to train the model to learn "how to learn quickly";

[0276] Fine-tuning phase: When encountering new scenarios, such as the first appearance of a "hail + dense tourist" composite scenario, only 10-20 samples are needed to adjust the model parameters;

[0277] The matching degree between the new scenario and the case library is measured by Euclidean distance. When the similarity is greater than 85%, the historical strategy is directly invoked.

[0278] When a new risk scenario is detected, the three most similar cases in the case library are quickly retrieved to generate a candidate strategy set; the Q-value of the candidate strategies is evaluated through the Critic network, and the optimal strategy is selected and sent to the agent; after each emergency decision is executed, the new case is added to the case library, and outdated cases that have not been accessed for more than two years are eliminated.

[0279] The decision logs recorded by the blockchain provide real-world data for meta-learning, and the strategies optimized by meta-learning are traceable and verifiable through the blockchain. When the meta-learning model generates a new strategy, the strategy version number and training sample set need to be recorded in the blockchain to ensure decision transparency.

[0280] User behavior data, such as path selection and preference tags, is anonymized before being stored in the blockchain, such as by hashing user IDs. During meta-learning training, a federated learning framework is used, where each agent retains its original data locally and only shares the amount of model parameter updates.

[0281] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for planning mountain tourism routes based on weather and natural disasters, characterized in that, include: Step 1: Based on the POMDP model, construct a Markov game decision model for multi-agent collaborative optimization, integrate natural disaster monitoring data and ecological environment indicators from transportation routes and tourist destinations, combine the Mamba multimodal feature representation learning method, and construct a dynamic risk matrix through multimodal feature compression and RAG semantic enhancement processing to provide a state perception basis for multi-agent systems in partially observable environments; Step 2: Optimize the UAV inspection strategy through federated learning and the NSGA-II algorithm, and establish an optimization target model that satisfies the Nash equilibrium constraint; design a dedicated agent for weather perception needs, deploy a lightweight YOLOv8 model to identify weather phenomena, and propose a communication and feature sharing mechanism among multiple agents. Introduce the NSGA-II algorithm into the Actor-Critic architecture to optimize the UAV inspection frequency and path adjustment strategy, so that the dynamic communication protocol is automatically triggered when the rainfall intensity exceeds the threshold. Step 3: Design a reward function that integrates security and user experience benefits, and output the final path instruction through an improved DDPG algorithm; Step 4: Build user profiles based on historical user behavior data, perform dynamic weight adjustments and AR simulation previews; automatically increase the weight of corresponding benefit items when users select specific modes, develop AR weather simulation functions for preference previews, and finally generate observation-behavior mapping strategies through joint game theory. Step 5: Based on the GIS geographic context reasoning framework, a closed-loop decision engine with perceptual input and cognitive output capabilities is formed. Feature complementarity is achieved through the feedback mechanism between the agent and the environment. Blockchain technology is introduced to record emergency decision logs, and a disaster emergency case library is built in combination with the meta-learning framework, enabling the system to quickly adapt to new climate regions. The reward-benefit balance mechanism is used for spatiotemporal evolution intelligent reasoning to improve the robustness of group decision-making in continuous action space.

2. The method for planning mountain tourism routes based on weather and natural disasters according to claim 1, characterized in that, The construction of the dynamic risk matrix includes the following steps: Deploy a geological monitoring sensor network to collect relevant data on landslides and debris flows; access meteorological radar and satellite cloud image data to obtain weather indicators including rainfall intensity, visibility, and wind speed; and use GIS geographic registration to unify data from different sources into the WGS84 coordinate system for spatial alignment. A state-space model is used to process time series data, and an implicit representation of continuous time series is constructed. Let the geological sensor sequence be... Mapped to the latent space by a linear transformation W: in, Let be the hidden state vector at time step t; The linear transformation matrix of the input data; Let be the linear transformation matrix of the historical hidden states; It is a bias vector with dimension 64; The final output is a 64-dimensional geological feature vector; Construct a disaster domain knowledge graph to map text reports into structured indicators; use LLM for event extraction; The matrix dimensions are defined by factors including geological hazard probability G, weather risk level Y, and ecological sensitivity E. Overall risk value calculation: A red alert is triggered when R ≥ 75; Deploy a lightweight ONNX inference engine to perform data preprocessing, feature compression, and risk calculation on a Raspberry Pi 4B device; The update cycle is strictly controlled within 5 minutes, and the dimension weights are automatically adjusted according to seasonal changes. By compressing high-dimensional monitoring data into computable quantitative indicators and combining it with real-time semantic enhancement technology, a risk assessment system with dynamic adaptability is constructed, providing accurate environmental state input for subsequent multi-agent collaborative decision-making.

3. The method for planning mountain tourism routes based on weather and natural disasters according to claim 2, characterized in that, The proposed communication and feature sharing mechanism among multiple agents, which introduces the NSGA-II algorithm into the Actor-Critic architecture, includes the following steps: Based on the dynamic risk matrix output in step one, the communication threshold is set as follows: it is triggered when the comprehensive risk value R is greater than 60 or the rainfall intensity is greater than 20 mm / h. When the threshold is triggered, the agent automatically switches to high-frequency communication mode. A hybrid communication topology is adopted, using distributed communication under normal conditions and automatically switching to centralized communication during disaster warning. The high-definition images collected by the UAV are compressed using the JPEG2000 compression algorithm, compressing the data volume to 15% of the original size while maintaining 95% of the image quality. Each agent trains a lightweight YOLOv8 model on a local dataset, protects local data privacy using differential privacy techniques, and aggregates model parameters on the server side using the FedAvg algorithm. in, These are the model parameters after global aggregation. For the local model parameters of the k-th agent, | | represents the local data volume; The maximum mean difference metric is used to measure the consistency of feature distributions. The visual features of the UAV, geological sensor features, and meteorological data features are projected into a unified 128-dimensional space; a self-attention mechanism is used to calculate the importance of each feature. in, Let be the attention weight for the i-th feature. Let be the energy score of the i-th feature, W be the learnable weight matrix, and b be the bias term. This is the transposed query vector. Let i be the i-th feature vector, and the final fused features are: ; NSGA-II-Actor-Critic Multi-Objective Optimization: Safety benefits G represents the probability of geological disasters; Efficiency Benefits v represents the actual inspection speed; Ecological benefits E represents ecological sensitivity; The Actor network generates an initial policy π(s|θ), then NSGA-II performs crossover and mutation operations on θ to generate a set of candidate policies. Finally, the Critic network evaluates the Q-value of each policy. The winning strategy is selected through non-dominated sorting and crowding calculation, and the Actor parameters are updated. Introduce constraints during policy updates: in, Let i be the reward function for the i-th agent; The changed policy for the i-th agent; It is a set of policies for other agents; ensuring that no agent can gain a higher benefit by unilaterally changing its policy. The drone adjusts its inspection path based on the fusion feature F. When a new risk point is detected, it sends a priority interruption signal to the ground agent via the V2X protocol. It uses cognitive radio technology to automatically switch between the 2.4GHz and 5GHz frequency bands to avoid communication congestion.

4. The method for planning mountain tourism routes based on weather and natural disasters according to claim 1, characterized in that, The process of designing a reward function that integrates security and user experience benefits, and outputting the final path instruction through an improved DDPG algorithm, includes the following steps: Geological disaster risk: G represents the probability of geological disasters; Weather risks; Y represents the weather risk level; Ecological sensitivity; E represents ecological sensitivity; Overall security benefits: ; The landscape along the route is scored using a pre-trained CNN model; the similarity between the route and user preferences is calculated based on the weather sensitivity profile generated in step four. Overall experience benefits: Where A is the landscape aesthetic score output by the pre-trained CNN model, with a value range of [0, 1]; S is the path preference matching degree calculated based on user profile, using cosine similarity, with a value range of [0, 1]. Total reward function: Where α is the dynamic weighting coefficient, adjusted according to the real-time risk value: when When <0.7, α=0.7; when When ≥0.7, α=0.3; The Actor network takes the dynamic risk matrix and user preference vector output from step one as input and outputs continuous actions; the Critic network uses double Q-value estimation to reduce overestimation bias, takes state-action pairs as input and outputs Q-value estimates; adaptive Ornstein-Uhlenbeck noise is introduced, with the noise amplitude decreasing with training epochs. The initial noise σ0 = 0.5 and the attenuation rate λ = 0.

001. Based on the TD error calculation sampling priority, the sampling probability of samples with large errors is increased by 30%; the experience data of different agents are stored in a shared replay pool; A feasible area mask is generated using DEM elevation data to prevent paths from crossing dangerous cliff areas; a conservative path mode is automatically triggered when communication between the UAV and the ground agent is interrupted. When the overall risk value R output in step one is greater than 75, α is forcibly set to 0.7, and the route planning prioritizes avoiding high-risk areas; when the user selects the landscape priority mode, α is dynamically adjusted to 0.3, allowing the route to detour to areas with high aesthetic scores; the hybrid mode defaults to α=0.5, balancing safety and experience. An asynchronous Actor-Learner architecture pre-trained model is deployed on 4 GPUs and initialized in a simulated mountainous environment. After the Actor output, a hard constraint layer is added to forcibly exclude all illegal actions that violate the physical rules.

5. The method for planning mountain tourism routes based on weather and natural disasters according to claim 1, characterized in that, The introduction of the NSGA-II algorithm into the Actor-Critic architecture includes the following steps: 50% of the strategies were randomly generated through the Actor network, and 50% were generated through NSGA-II crossover mutation. The top 10% of high-fitness strategies were retained as the initial population. Based on the current state s, output continuous actions a; use SBX crossover to generate offspring policies; apply polynomial mutation to explore the new policy space; calculate the Q-value of each policy: The Pareto front is divided into levels based on Q-values ​​and constraints; within the same front, strategies with better diversity are selected using distance metrics. The policy parameter θ is updated along the Q-value gradient direction using the policy gradient method: The strategy of retaining the top 20% of elites is combined with newly generated strategies to form the next generation of the population; When the overall risk value R ≥ 75, the weight of safety benefits increases by 30%; when a "landscape priority" request is detected, the weight of experience benefits increases by 20%. Under the federated learning framework, each agent shares the top 5% of elite policies, and the privacy of the policies is protected through differential privacy; after the shared policies are adjusted by local fitness, they re-participate in NSGA-II evolution. A repeated game mechanism is adopted, in which each agent proposes a strategy in turn, and the Critic evaluates whether an equilibrium has been reached; when the rate of change of all agents' payoffs is less than 2% after 5 consecutive rounds of strategy updates, the game equilibrium is determined to have been reached. The final generation includes path direction angles. ,speed Risk aversion coefficient Triple decision-making: The decision simultaneously satisfies multi-objective optimization and Nash equilibrium constraints.

6. The method for planning mountain tourism routes based on weather and natural disasters according to claim 1, characterized in that, The process of building a user profile based on historical behavior data, dynamically adjusting weights, and performing AR simulation previews includes the following steps: Collect user historical behavior data, including: travel records, weather feedback, and social media comments; remove irrelevant information through text cleaning, retaining key weather-related statements; User-selected weather modes and travel itinerary records were analyzed using NLP to assess sentiment in user comments; users were then categorized into three groups using the K-means algorithm. Rain-sheltered type: Over 70% of the historical itineraries have had their routes adjusted due to rainfall; Rain scene photography type: Actively choose rainy day trips and the keyword "rain scene" appears more than 5 times in the comments; Climate-adaptive type: No obvious weather preference, and the correlation between itinerary selection and weather is less than 30%; After each user completes a trip, real-time feedback is collected through questionnaires to update user profile tags; if a user's behavioral data has not been updated for 3 months, the tag weight decreases by 10% each month. The weight adjustment is triggered when the user explicitly selects either rain scene priority or safety priority mode through the APP interface; or when the comprehensive risk value R output in step one is greater than 75, the user is forced to enter the safety priority mode; when there is no explicit user instruction and the environment is safe, the default balanced mode is used. The weighting adjustment rules are as follows: for rain-avoidance users, the safety benefit weight is increased to 70%, the experience benefit weight is reduced to 30%, and route planning is prohibited from passing through areas with a rainfall probability greater than 40%; for rain scene photography users, the user experience benefit weight is increased to 60%, the safety benefit weight is reduced to 30%, and route detours to areas with a rainfall probability greater than 50% are allowed; for climate-adaptive users, the default weight is maintained, and they are only forcibly adjusted to safety mode when the comprehensive risk value R is greater than 80. The weighting adjustment constraints are that the minimum weighting of safety benefits is no less than 30%, and the maximum weighting of experience benefits is no more than 60%. After the weighting adjustment, the feasibility of the strategy must be verified through the Critic network. After the user selects a preferred mode, AR technology is used to display the weather effects along the route in real time: when the rain scene priority mode is selected, the virtual rainfall effect and possible shooting points are displayed; when the safety priority mode is selected, the geological disaster risk areas are highlighted. The user can adjust the simulation parameters through gesture interaction. Users can rate the display effect in the simulation interface, and the rating data is fed back to the profile update module in real time to adjust the subsequent weight allocation strategy.

7. The method for planning mountain tourism routes based on weather and natural disasters according to claim 1, characterized in that, The GIS-based geographic context reasoning framework forms a closed-loop decision engine with perceptual input and cognitive output capabilities. It achieves feature complementarity through a feedback mechanism between the agent and the environment, including the following steps: Integrate high-precision DEM data, vegetation cover layers, water system distribution layers, and historical geological disaster point data; generate composite layers containing slope, catchment areas, and ecological red line areas using GIS spatial analysis tools; The dynamic risk matrix from step one is overlaid with the GIS layer to generate a real-time geographic scenario model; the model is updated every 5 minutes, keeping pace with the data refresh rate in step one. The system uses a 3D GIS engine to render the terrain, highlighting high-risk areas, user-preferred paths, and the current location of the agent; it also uses AR technology to overlay virtual paths onto the real environment, allowing users to intuitively view the effects of their decisions. The joint game decision generated in step three is sent to the drone and the ground intelligent agent; during the execution of the path by the intelligent agent, environmental data is collected in real time; environmental change data is sent back to the decision engine via the V2X protocol to update the GIS geographic context model. Based on the updated model, the path planning parameters are recalculated to form a closed-loop adjustment. Dynamically adjust trigger conditions: Environmental mutation: When a geological disaster probability G ≥ 80% or a weather risk level W ≥ 4 is detected, route replanning will be forcibly triggered; User intervention: When a user manually adjusts their path preferences through the app, a new decision is generated immediately; Periodic optimization: Lightweight path optimization is automatically performed every 15 minutes to balance computing resources and decision-making timeliness; Drones provide a high-altitude macroscopic view to identify large-scale geological changes; ground sensors provide microscopic data; and weather station data supplements local weather changes. By using GIS coordinate transformation, the UAV orthophotos and ground sensor data are unified into the same coordinate system; the PTP clock synchronization protocol is adopted, and weights are allocated according to the reliability of the data source. The geological change areas identified by the drone are confirmed by ground sensor data; when a single data source shows an anomaly, a self-check is initiated and the data is marked as suspicious. Each decision's input data, output actions, and execution results are packaged and uploaded to the blockchain; the on-chain data is stored using a Merkle tree structure, ensuring data integrity and traceability. Historical decision-making cases are extracted from the blockchain to build a case library containing typical scenarios; the MAML algorithm is used to quickly adapt to new scenarios with a small number of samples, reducing online learning time; combined with GIS spatiotemporal analysis tools, environmental change patterns are identified; and spatiotemporal patterns are encoded into a rule engine to assist the decision engine in predicting environmental trends.

8. The method for planning mountain tourism routes based on weather and natural disasters according to claim 1, characterized in that, The process of introducing blockchain technology to record emergency decision logs and building a disaster emergency case library using a meta-learning framework includes the following steps: The system uses environmental state snapshots to record dynamic risk matrices and GIS geographic context models during decision-making; decision action details include path direction angle θ, speed v, risk avoidance coefficient α, and multi-agent communication protocol status; execution result feedback includes actual user travel data and environmental changes monitored by the system; user intervention records include user manual adjustments to path preferences and system responses. Each decision is fully logged and encapsulated as a blockchain transaction, with timestamps and digital signatures added, and the data is organized using a Merkle tree structure. Access control: Write permissions: Decision engine and user terminal only; Read access: Granted to authorized organizations for post-event analysis; The blockchain explorer allows users to query decision logs at any point in time, supporting multi-dimensional searches by user ID, time range, and risk level. In the event of a geological disaster, it can quickly pinpoint the decision-making path and environmental conditions 5 minutes before the incident, assisting in liability determination. Historical decision-making cases are extracted from the blockchain and categorized by scenario type into typical disaster scenarios, user preference scenarios, and mixed scenarios. Each case is labeled with key features including a comprehensive risk value R, user preference tags, decision action sequence, and execution result score. The MAML algorithm is used to quickly adapt to new scenarios with a small number of samples. Meta-training phase: Simulate different disaster scenarios in a historical case library to train the model to learn how to learn quickly; Fine-tuning phase: When encountering new scenarios, adjust the model parameters using 10-20 samples; The matching degree between the new scenario and the case library is measured by Euclidean distance. When the similarity is greater than 85%, the historical strategy is directly invoked. When a new risk scenario is detected, the three most similar cases in the case library are quickly retrieved to generate a candidate strategy set; the Q-value of the candidate strategies is evaluated through the Critic network, and the optimal strategy is selected and sent to the agent; after each emergency decision is executed, the new case is added to the case library, and outdated cases that have not been accessed for more than two years are eliminated. The decision logs recorded by the blockchain provide real-world data for meta-learning, and the strategies optimized by meta-learning are traceable and verifiable through the blockchain. When the meta-learning model generates a new strategy, the strategy version number and training sample set are recorded in the blockchain to improve decision transparency. User behavior data is anonymized before being stored in the blockchain; during meta-learning training, a federated learning framework is used, where each agent retains the original data locally and only shares the model parameter update amount.

Images