Multi-source fusion-based intelligent agent landscape probability prediction method and system for tourism meteorology

Abstract

This invention discloses a method and system for probabilistic landscape forecasting based on multi-source fusion using intelligent agents for tourism meteorology, relating to the field of tourism meteorological service technology. The method includes: acquiring multi-source data of each landscape node in a target scenic area and constructing a multi-dimensional feature tensor for each landscape node; generating a heat map of suitable landscape areas for each landscape type during the forecast period based on the multi-dimensional feature tensor; constructing a multi-agent collaborative framework, obtaining the landscape occurrence probability and confidence interval of each landscape node based on the framework; identifying anomalous clustering events based on multi-source data, calculating the forecast deviation of each landscape node based on these events, and updating the multi-agent collaborative framework; and generating a visualized forecast result interface based on the landscape suitable area heat map, the landscape occurrence probability of each landscape node, and the confidence interval. This invention addresses the problems of coarse forecast granularity, lack of spatial refinement capabilities, limited data sources, and insufficient utilization of multi-source heterogeneous data in existing technologies.

Description

Technical Field

[0001] This invention relates to the field of tourism meteorological service technology, and more specifically, to a method and system for predicting the probability of tourism meteorological landscapes based on multi-source fusion intelligent bodies. Background Technology

[0002] With the rapid development of the tourism industry, the demand for tourism meteorological services is increasing, especially the probability forecast of natural phenomena such as sea of ​​clouds, sunrise, starry sky, and Buddha's light, which is of great significance for tourists' travel decisions and the improvement of their travel experience.

[0003] For example, Chinese invention patent application CN106772697A discloses a method and system for forecasting the natural landscape of a sea of ​​clouds. This method involves setting up a first automatic weather station at an altitude higher than the cloud top and a second automatic weather station within the sea of ​​clouds to collect meteorological data such as temperature, humidity, wind speed, and rainfall for the day. Combined with weather forecast data issued by the meteorological bureau and the occurrence of the sea of ​​clouds that day, a forecast equation is pre-established using logistic regression analysis. Substituting the aforementioned data into the forecast equation yields the forecast result for the sea of ​​clouds the following day. This method can provide a highly accurate forecast of the probability of the sea of ​​clouds appearing, facilitating travel plans for tourists and photography enthusiasts and promoting the development of local tourism.

[0004] However, the above method has the following shortcomings: (1) The forecast granularity is coarse and lacks spatial refinement capabilities: the forecast results can only reflect whether the sea of ​​clouds appears, but cannot determine the spatial distribution of the sea of ​​clouds in the scenic area. Moreover, the landscape type coverage is singular, only targeting the single landscape type of sea of ​​clouds, which limits its applicability in different tourism scenarios.

[0005] (2) Limited data sources and insufficient utilization of multi-source heterogeneous data: relying only on weather forecast data from two automatic weather stations and the meteorological bureau, the data sources of satellite remote sensing data, high-precision topographic elevation data, real-time monitoring images of scenic spots and tourist location data are not fully utilized, which limits the accuracy and reliability of forecasts.

[0006] To address the aforementioned shortcomings, there is an urgent need for a landscape probability forecasting method for tourism meteorological intelligent agents that offers high forecast accuracy, spatial refinement capabilities, and the ability to fully utilize multi-source heterogeneous data to improve the accuracy of landscape probability forecasts. Summary of the Invention

[0007] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method and system for probabilistic forecasting of landscape based on multi-source fusion of intelligent tourism meteorological agents, in order to solve the problems of coarse forecast granularity, lack of spatial refinement capability, limited data sources, and insufficient utilization of multi-source heterogeneous data in the prior art.

[0008] To achieve the above objectives, this application provides the following technical solution: A first aspect of the present invention provides a probabilistic landscape forecasting method based on multi-source fusion for intelligent tourism meteorological agents, comprising the following steps: Acquire multi-source data of each landscape node in the target scenic area, and construct a multi-dimensional feature tensor of the landscape node based on the multi-source data; Based on the multidimensional feature tensor, a meteorological and topographic field of the target scenic area is constructed during the forecast period. Based on the meteorological and topographic field, suitable distribution areas of landscape types are identified, and a heat map of suitable areas for each landscape type during the forecast period is generated based on the suitable distribution areas. A multi-agent collaborative framework is constructed, and multi-dimensional feature tensors and three-dimensional encoding vectors from landscape suitability zone heatmaps are extracted based on the multi-agent collaborative framework. The three-dimensional encoding vectors are then fused to obtain the landscape occurrence probability and confidence interval of landscape nodes. Based on the multi-source data, abnormal clustering events are identified, and the implicit labels of each landscape node are determined based on the abnormal clustering events. The prediction deviation of each landscape node is calculated based on the implicit labels and the probability of landscape occurrence, and the multi-agent cooperation framework is updated based on the prediction deviation. A visualized forecast result interface is generated based on the heat map of suitable landscape areas, the probability of landscape occurrence at each landscape node, and the confidence interval.

[0009] In a preferred embodiment, constructing a multidimensional feature tensor for landscape nodes based on multi-source data includes: The multi-source data includes meteorological observation data, numerical weather prediction data, satellite remote sensing data, topographic elevation data, scenic area monitoring images, and tourist location data; The multi-source data is time-aligned and standardized, and meteorological features, remote sensing features, topographic features, visual features, and tourist distribution features of each landscape node are extracted based on the processed multi-source data. Based on the meteorological features, remote sensing features, topographic features, visual features, and tourist distribution features, feature vectors for each landscape node are formed, and a multidimensional feature tensor for the target scenic area is constructed based on the feature vectors.

[0010] In a preferred embodiment, the meteorological and topographic field of the target scenic area during the forecast period is constructed based on the multidimensional feature tensor, including: A three-dimensional topographic feature field and a three-dimensional meteorological distribution field of the target scenic area are constructed based on the multidimensional feature tensor. The terrain slope, the angle between wind direction and slope aspect, and the lifting condensation height are obtained based on multi-source data. The terrain lifting operator is calculated based on the terrain slope, the angle between wind direction and slope aspect, and the lifting condensation height. The terrain height and the thickness of the incoming air layer are obtained based on multi-source data, and the terrain blocking operator is calculated based on the terrain height and the thickness of the incoming air layer. The angle between the valley direction and the wind direction, as well as the height difference of the valley terrain, are obtained based on multi-source data. The narrowing effect operator is calculated based on the angle and the height difference of the valley terrain. The angle relationship between slope aspect and solar radiation is obtained based on multi-source data, and a thermodynamic modulation operator is calculated based on the angle relationship and solar radiation intensity. A physical coupling operator is constructed based on the topographic lifting operator, topographic blocking operator, funnel effect operator, and thermal modulation operator. The three-dimensional topographic feature field and the three-dimensional meteorological distribution field are physically coupled based on the physical coupling operator to obtain the meteorological topographic field.

[0011] In a preferred embodiment, a landscape suitability zone heat map for each landscape type during the forecast period is generated based on the suitable distribution area, including: Construct a landscape formation condition model, which includes the landscape formation conditions corresponding to each landscape type; Based on the landscape formation conditions and the meteorological characteristics of each spatial grid point in the meteorological and topographic field, the suitable distribution area of ​​the landscape type is identified, and the probability value of the spatial grid point in the suitable distribution area meeting the landscape formation conditions during the forecast period is calculated. Based on the probability values ​​and suitable distribution areas, a heat map of suitable landscape areas for each landscape type is generated during the forecast period.

[0012] In a preferred embodiment, the extraction of multidimensional feature tensors and three-dimensional encoded vectors from landscape suitability zone heatmaps based on a multi-agent collaborative framework includes: The multi-agent collaborative framework includes a meteorological analysis agent, a terrain coupling agent, a landscape identification agent, and a fusion decision-making agent. A meteorological analysis agent is constructed based on a temporal coding network, and the meteorological feature sequence of landscape nodes is used as the input of the meteorological analysis agent. The meteorological condition coding vector is output based on the meteorological analysis agent. A terrain-coupled agent is constructed based on a convolutional neural network. The meteorological topographic field and the heat map of suitable landscape areas are used as inputs to the terrain-coupled agent. The terrain-landscape coupled encoding vector is output based on the terrain-coupled agent. A landscape identification agent is constructed based on a residual network. The visual features of the landscape nodes are used as the input of the landscape identification agent, and the visual state encoding vector is output based on the landscape identification agent. The three-dimensional coding vector is constructed based on the meteorological condition coding vector, the terrain landscape coupling coding vector, and the visual state coding vector.

[0013] In a preferred embodiment, fusing the three-dimensional encoded vectors to obtain the landscape occurrence probability and confidence interval of the landscape node includes: The fusion decision agent is based on a cross-attention fusion mechanism, which deeply fuses the meteorological condition encoding vector, the terrain landscape coupled encoding vector, and the visual state encoding vector to obtain fused features. The landscape prediction probability is calculated based on the fusion features, and the landscape occurrence probability is obtained by performing a preset number of random forward calculations on the landscape prediction probability. The radius of the confidence interval is obtained based on the standard deviation of the landscape occurrence probability corresponding to the preset number of calculations.

[0014] In a preferred embodiment, identifying anomalous clustering events based on the multi-source data includes: Tourist location data is obtained from multi-source data. Similarity clustering is performed on the tourist location data to obtain the clustering results of each landscape node. Based on the clustering results, a time series curve of tourist density of each landscape node in the target scenic area is constructed. Based on the historical visitor distribution data of each scenic spot within the target scenic area, a baseline visitor density curve for each scenic spot is constructed, and the threshold exceeding the limit is calculated based on the historical visitor distribution data. The time-series curve of tourist density is compared with the benchmark tourist density curve. If the time-series curve of tourist density of a landscape node exceeds the benchmark tourist density curve and the magnitude of the exceedance is greater than the threshold, it is determined that an abnormal clustering event has occurred at the landscape node.

[0015] In a preferred embodiment, determining the implicit label of each landscape node based on anomaly clustering events includes: Set up a multi-source auxiliary signal that includes time consistency verification signal, spatial consistency verification signal and visual verification signal, and set the signal weight of each verification signal in the multi-source auxiliary signal; Cross-validation of anomalous clustering events is performed based on multi-source auxiliary signals to obtain signal validation results. The implicit labels are then obtained by weighting the signal validation results according to the signal weights.

[0016] In a preferred embodiment, updating the multi-agent cooperative framework based on forecast bias includes: Calculate the confidence weight based on the signal verification results and the number of verification signals participating in the verification among the multi-source auxiliary signals; Based on the confidence weights, implicit labels, and landscape occurrence probabilities, the prediction bias for each landscape node is calculated. Based on the prediction bias, online incremental learning is performed on each agent in the multi-agent cooperation framework to update the multi-agent cooperation framework.

[0017] A second aspect of the present invention provides a multi-source fusion-based intelligent tourism meteorological landscape probability forecasting system, comprising: The data acquisition module is used to acquire multi-source data of each landscape node in the target scenic area and construct multi-dimensional feature tensors of the landscape nodes based on the multi-source data. The thermal distribution module is used to construct the meteorological and topographic field of the target scenic area during the forecast period based on the multidimensional feature tensor, identify the suitable distribution area of ​​the landscape type based on the meteorological and topographic field, and generate a heat map of the suitable area of ​​each landscape type during the forecast period based on the suitable distribution area. The landscape occurrence probability module is used to construct a multi-agent collaborative framework. Based on the multi-agent collaborative framework, multi-dimensional feature tensors and three-dimensional encoding vectors from the landscape suitability zone heatmap are extracted. The three-dimensional encoding vectors are fused to obtain the landscape occurrence probability and confidence interval of the landscape node. The forecast update module is used to identify abnormal clustering events based on the multi-source data, determine the implicit label of each landscape node based on the abnormal clustering events, calculate the forecast deviation of each landscape node based on the implicit label and the probability of landscape occurrence, and update the multi-agent cooperation framework based on the forecast deviation. The forecast results module is used to generate a visualized forecast results interface based on the landscape suitability zone heat map, the probability of landscape occurrence at each landscape node, and the confidence interval.

[0018] The beneficial effects of this invention are: 1. This invention constructs a multidimensional feature tensor based on multi-source data of each landscape node in the target scenic area, and constructs a meteorological and topographic field of the target scenic area during the forecast period based on the multidimensional feature tensor. Based on the meteorological and topographic field, it identifies suitable distribution areas for landscape types, and generates a landscape suitable area heat map for each landscape type during the forecast period based on the suitable distribution areas. The landscape suitable area heat map improves the granularity of the forecast and further enhances the spatial refinement capability of the forecast.

[0019] 2. This invention constructs a multi-agent collaborative framework and extracts multi-dimensional feature tensors and three-dimensional encoding vectors from landscape suitability zone heatmaps based on this framework. The three-dimensional encoding vectors are fused to obtain the landscape occurrence probability and confidence interval of landscape nodes. Abnormal clustering events are identified based on multi-source data, and the implicit labels of each landscape node are determined based on these events. The forecast deviation of each landscape node is calculated based on the implicit labels and landscape occurrence probability. The multi-agent collaborative framework is updated based on the forecast deviation. By using multi-source heterogeneous data to identify abnormal clustering events and calculating the forecast deviation based on the implicit labels obtained from these events, the accuracy and reliability of the forecast results can be further improved through the multi-agent collaborative framework updated with the forecast deviation. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the probabilistic forecasting method for landscape based on multi-source fusion of intelligent tourism meteorological agents. Figure 2This embodiment provides a flowchart illustrating the method for constructing the meteorological and topographic field of a target scenic area during a forecast period based on a multidimensional feature tensor. Figure 3 This is a schematic diagram of the method for extracting multidimensional feature tensors and three-dimensional encoded vectors from landscape suitability zone heatmaps based on a multi-agent collaborative framework provided in this embodiment. Figure 4 This is a schematic diagram of the structure of a tourism meteorological intelligent body landscape probability forecasting system based on multi-source fusion. Detailed Implementation

[0021] 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0022] Example 1, please refer to Figure 1 , Figure 1 This invention presents a multi-source fusion-based intelligent agent-based probabilistic forecasting method for tourism meteorology, comprising the following steps: S1. Obtain multi-source data of each landscape node in the target scenic area, and construct a multi-dimensional feature tensor of the landscape node based on the multi-source data.

[0023] Among them, landscape nodes refer to the pre-set discrete landscape viewing points within the target scenic area, such as various viewing platforms set up within the target scenic area.

[0024] In this embodiment, step S1 includes: S11. Multi-source data includes meteorological observation data, numerical weather prediction data, satellite remote sensing data, topographic elevation data, scenic area monitoring images, and tourist location data.

[0025] Meteorological observation data is collected in real time by automatic weather stations deployed in and around the target scenic area. The data includes meteorological elements such as temperature (°C), relative humidity (%), wind speed (m / s), wind direction (°), precipitation (mm), air pressure (hPa), visibility (m), and dew point temperature (°C). The sampling frequency can be once every 10 minutes.

[0026] Among them, numerical weather forecast data refers to the numerical weather forecast grid data of the target scenic area for the next 72 hours, including meteorological elements such as temperature, relative humidity, wind speed, wind direction, total cloud cover, low cloud cover, and precipitation. The original spatial resolution is no less than 9km×9km, and the temporal resolution is 3 hours.

[0027] Among them, satellite remote sensing data refers to geostationary meteorological satellite observation data within the target scenic area, including infrared cloud images (channel brightness and temperature), visible light cloud images (albedo), cloud top height (m), cloud top temperature (K), and cloud cover (%), with a time resolution of 15 minutes to 1 hour.

[0028] Among them, the terrain elevation data refers to the high-precision digital elevation model (DEM) of the target scenic area, with a horizontal resolution of no less than 30m×30m, covering three-dimensional terrain features such as altitude (m), slope (°), aspect (°), terrain relief (m), and terrain roughness index.

[0029] Among them, the scenic area monitoring images are real-time image data collected by video monitoring equipment deployed at various scenic nodes, with a resolution of no less than 1920×1080 pixels and a frame rate of 1 frame per second.

[0030] Among them, tourist location data includes mobile terminal signaling data of tourists within the target scenic area, which is obtained by cooperating with telecommunications operators to obtain anonymized group aggregation information; scenic area WiFi probe data, which is obtained by deploying probes with a density of no less than 20 probes per square kilometer; and electronic ticketing gate data, which is obtained by recording the time of tourists entering and leaving the park. The data sampling interval for tourist location data can be set to 5 minutes.

[0031] S12. Perform time alignment and standardization on the multi-source data, and extract meteorological features, remote sensing features, topographic features, visual features, and tourist distribution features of each landscape node based on the processed multi-source data.

[0032] Time alignment of multi-source data refers to interpolating all multi-source data into a time grid with a resolution of 10 minutes to achieve time alignment. Standardization of the time-aligned multi-source data involves using the Z-score standardization method, which is well-known in existing technology, for all types of multi-source data. For example, for temperature in meteorological observation data, the mean temperature and standard deviation of temperature in the multi-source data within the forecast period are calculated. The mean temperature is subtracted from the real-time temperature value, and the difference is divided by the standard deviation of temperature to complete the temperature standardization.

[0033] S13. Based on meteorological features, remote sensing features, topographic features, visual features, and tourist distribution features, feature vectors are formed for each landscape node, and a multidimensional feature tensor of the target scenic area is constructed based on the feature vectors.

[0034] The feature vector of each landscape node can be specifically represented as follows: ,in, As a meteorological feature, As a topographic feature, For remote sensing features, As a visual feature, Characteristics of tourist distribution; The process of extracting meteorological features is as follows: from the standardized meteorological observation data and numerical weather prediction data, spatiotemporal interpolation is performed according to the node position of the landscape node, and the linear fitting slope of the past 6 hours is calculated. The meteorological features may specifically include data such as temperature, relative humidity, wind speed, wind direction sine value, wind direction cosine value, precipitation, air pressure, visibility, dew point temperature, total cloud cover, and low cloud cover.

[0035] The process of extracting remote sensing features is as follows: extract the corresponding pixel values ​​from the satellite remote sensing data according to the location of the landscape node to obtain the remote sensing features. The remote sensing features may specifically include data such as infrared cloud image brightness and temperature, visible light cloud image albedo, cloud top height, cloud top temperature, and cloud cover.

[0036] The process of extracting terrain features is as follows: based on terrain elevation data, the slope and aspect are calculated using a third-order unweighted difference algorithm known to those skilled in the art, and the terrain undulation is calculated using a 3×3 window movement. The terrain features may specifically include data such as altitude, slope, aspect sine, aspect cosine, terrain undulation, terrain roughness index, and valley-ridge index.

[0037] The process of visual feature extraction is as follows: the scenic area monitoring images are scaled up to 224×224 pixels and input into the ResNet-50 network to extract the visual features.

[0038] The process of extracting tourist distribution characteristics is as follows: based on spatiotemporal interpolation of WiFi probe and signaling data, it can specifically include the current tourist density (people / 100 square meters), the tourist density change rate in the past 1 hour, and the historical average density at the same time in the past 24 hours.

[0039] Specifically, the construction of a multidimensional feature tensor for the target scenic area based on feature vectors involves using the number of landscape nodes, the number of time steps within the forecast period, and the feature dimension as the dimensions of the multidimensional feature tensor. The multidimensional feature tensor for the target scenic area is then constructed based on the feature vectors of each landscape node. The target feature tensor can be specifically represented as follows: ,in, As a meteorological feature, As a topographic feature, For remote sensing features, As a visual feature, The distribution characteristics of tourists are represented by n, which is the number of the landscape node.

[0040] S2. Construct the meteorological and topographic field of the target scenic area during the forecast period based on the multidimensional feature tensor, identify the suitable distribution area of ​​the landscape type based on the meteorological and topographic field, and generate a heat map of the suitable area of ​​each landscape type during the forecast period based on the suitable distribution area.

[0041] In this embodiment, step S2 includes: S21. Before constructing the meteorological and topographic field of the target scenic area during the forecast period based on the multidimensional feature tensor, construct the three-dimensional topographic feature field and the three-dimensional meteorological distribution field of the target scenic area based on the multidimensional feature tensor.

[0042] Specifically, step S21 includes: S211. Construct a three-dimensional topographic feature field of the target scenic area based on the topographic features in the multidimensional feature tensor. The three-dimensional topographic feature field includes the elevation gradient field, slope field, aspect field, topographic roughness index, and valley-ridge orientation field.

[0043] Among them, the elevation gradient field can be obtained directly from the topographic elevation data; the slope field and aspect field are calculated by using the well-known third-order unweighted difference algorithm; the topographic roughness index is calculated based on the elevation standard deviation within a 3×3 window; and the valley-ridge orientation field can be extracted based on topographic curvature analysis.

[0044] S212. Downscale the meteorological features in the multidimensional feature tensor to a three-dimensional spatial grid of a preset scale to obtain the temperature stratification field, humidity field, wind field and cloud water content field of each spatial grid point. Based on the temperature stratification field, humidity field, wind field and cloud water content field of each spatial grid point, construct the meteorological three-dimensional distribution field of the target scenic area.

[0045] The process of downscaling meteorological features in a multidimensional feature tensor to a three-dimensional spatial grid of a preset scale refers to using a bilinear interpolation method known in existing technology to downscale meteorological features in a multidimensional feature tensor to a three-dimensional spatial grid of a preset scale. The preset scale parameters of the three-dimensional spatial network are: horizontal resolution of 100m×100m; vertical layering from the ground surface to an altitude of 3000m; and one layer every 100m.

[0046] It should be noted that the meteorological three-dimensional distribution field is specifically presented in the form of a three-dimensional spatial grid. That is, through the meteorological three-dimensional distribution field, each location of the target scenic area is represented as a discrete grid point in a three-dimensional spatial grid. A grid point is each intersection point on a continuous regular grid formed after the target scenic area space is regularly discretized in the horizontal and vertical directions. In the horizontal direction, the horizontal space (east-west × north-south) of the scenic area is divided into regular small squares. In this scheme, the preset size of each square is 100m × 100m. In the vertical direction, at each horizontal grid point, a vertical layer is set every 100m from the ground surface to an altitude of 3000m. Therefore, each grid point is a three-dimensional spatial point, and its coordinate form is... ,in The coordinates are in the horizontal longitude direction. The coordinates are in the horizontal latitude direction. The vertical height coordinates represent the coordinates of a specific location within the target scenic area, i.e., the horizontal coordinates and altitude.

[0047] S22. Construct the meteorological and topographic field of the target scenic area during the forecast period based on the multidimensional feature tensor.

[0048] Please see Figure 2 Step S22 includes: S221. Obtain the terrain slope, the angle between wind direction and slope aspect, and the lifting condensation height based on multi-source data, and calculate the terrain lifting operator based on the terrain slope, the angle between wind direction and slope aspect, and the lifting condensation height.

[0049] Specifically, the absolute value of the difference between the wind direction angle and the slope aspect angle is used as the angle between the wind direction and the slope aspect. The horizontal wind speed is obtained, and the lifting speed of the air mass along the slope is calculated based on the horizontal wind speed, the terrain slope, and the angle between the wind direction and the slope aspect. The specific calculation process is as follows: multiply the horizontal wind speed by the sine of the terrain slope, and then multiply it by the cosine of the angle between the wind direction and the slope aspect to obtain the lifting speed of the air mass along the slope. The base lifting condensation height is obtained, and the terrain lifting operator is calculated based on the base lifting condensation height and the lifting speed of the air mass along the slope. The specific calculation process is as follows: calculate the definite integral of the lifting speed function with respect to time from the lower limit of integration 0 to the upper limit of integration t, and add the base lifting condensation height to the definite integral result to obtain the terrain lifting operator.

[0050] S222. Obtain the terrain height and the thickness of the incoming air layer based on multi-source data, and calculate the terrain blocking operator based on the terrain height and the thickness of the incoming air layer.

[0051] The terrain blocking operator is calculated based on the ratio of terrain height to the thickness of the incoming air layer.

[0052] S223. Based on multi-source data, obtain the angle between the valley direction and the wind direction, as well as the height difference of the valley terrain, and calculate the septum effect operator based on the angle and the height difference of the valley terrain.

[0053] Specifically, when the angle between the wind direction and the valley direction is greater than 30°, the value of the funnel effect operator is 1; When the angle between the wind direction and the valley orientation is less than 30°, the calculation process of the narrowing effect operator is as follows: divide the valley topographic height difference by the valley bottom width and add 1 to obtain the narrowing effect operator.

[0054] S224. Obtain the angular relationship between slope aspect and solar radiation based on multi-source data, and calculate the thermal modulation operator based on the angular relationship and solar radiation intensity.

[0055] Among them, the angular relationship between slope aspect and solar radiation includes the angle of incidence of sunlight on the slope and the reference angle of incidence of the sun under flat terrain conditions.

[0056] The specific calculation process of the thermal modulation operator is as follows: subtract the cosine value of the solar reference incident angle from the cosine value of the solar incident angle on the slope, and then multiply the difference by the solar radiation intensity to obtain the thermal modulation operator.

[0057] S225. Based on the terrain lifting operator, terrain blocking operator, septum effect operator and thermal modulation operator, a physical coupling operator is constructed, and the three-dimensional terrain feature field and the three-dimensional meteorological distribution field are physically coupled based on the physical coupling operator to obtain the meteorological terrain field.

[0058] Specifically, the terrain lifting operator, terrain blocking operator, funneling effect operator, and thermal modulation operator are input into the physical coupling operator. That is, the physical coupling operator includes these four operators: terrain lifting operator, terrain blocking operator, funneling effect operator, and thermal modulation operator. The physical coupling operator can be specifically represented as follows: ; In the formula, For meteorological and topographic fields, It is a three-dimensional terrain feature field. For the three-dimensional distribution field of meteorology, It is a physical coupling operator that includes terrain lifting operator, terrain blocking operator, funneling effect operator, and thermal modulation operator. The coordinates are in the horizontal longitude direction. The coordinates are in the horizontal latitude direction. The vertical height coordinates are... The units are all meters; This refers to time, expressed in hours.

[0059] It should be noted that since the numerical weather forecast data acquired by S1 is for large areas, it assumes that the surface is flat or that the terrain is described by only a coarse grid. However, in scenic areas with obvious mountainous terrain, the temperature and wind speed changes by 100 meters of elevation gain are enough to completely change the landscape conditions. Therefore, by physically coupling the three-dimensional terrain feature field and the three-dimensional meteorological distribution field through a physical coupling operator, the three-dimensional meteorological distribution field can be transformed into meteorological data of the three-dimensional terrain feature field corresponding to the target scenic area. That is, the meteorological topographic field obtained by physical coupling represents the meteorological data of each grid point in the target scenic area. In the process of physical coupling, the terrain is independently corrected for each grid point (100m×100m) in the target scenic area.

[0060] S23. Generate a heat map of suitable landscape areas for each landscape type during the forecast period based on the suitable distribution area.

[0061] Specifically, step S23 includes: S231. Construct a landscape formation condition model, which includes the landscape formation conditions corresponding to each landscape type.

[0062] The landscape formation condition model can be preset based on the actual landscape scene of the target scenic area, and it includes at least landscape types such as sea of ​​clouds, sunrise, and starry sky. Specifically, the formation conditions for sea of ​​clouds can be to meet the requirements of inversion layer strength. 0.5℃ / 100m and low cloud cover The area has 60% humidity and a coupled humidity field RH ≥ 85%; the specific conditions for the formation of sunrise landscapes can be that the terrain shading angle in the direction of the eastern skyline is met. 3° and total cloud cover at upper altitude 40% of the area; the specific conditions for the formation of starry sky landscapes can be defined as meeting the total cloud cover in the upper atmosphere. 30% ground visibility 15km, altitude The local average altitude is 200m higher and the light pollution index is higher. For Level 2 areas, the light pollution index is a fixed parameter for the scenic area, obtained using the Bortle dark sky classification method, which is known in existing technology. Level 1 is complete darkness, and Level 9 is strong urban light pollution.

[0063] It should be noted that by comparing the meteorological data of each grid point within the target scenic area in the meteorological and topographic field with the landscape formation conditions of each landscape type in the landscape formation condition model, it is determined which grid points within the target scenic area meet the landscape formation conditions, thus obtaining the landscape type of each grid point. Furthermore, a grid point is only considered to have that landscape type if its meteorological data meets all the landscape formation conditions for that type. For example, if the meteorological data of a grid point all meet the inversion layer intensity... 0.5℃ / 100m and low cloud cover 60% and the coupled humidity field RH If the success rate is 85%, then it can be determined that a sea of ​​clouds landscape will appear at that grid point.

[0064] S232. Based on the landscape formation conditions and the meteorological characteristics of each spatial grid point in the meteorological and topographic field, identify the suitable distribution area of ​​the landscape type, and calculate the probability value of the spatial grid point in the suitable distribution area meeting the landscape formation conditions during the forecast period.

[0065] Specifically, for each spatial grid point in the meteorological and topographic field, the determination of whether the spatial grid point meets the landscape formation conditions is based on the landscape formation conditions of each type of landscape in the landscape formation condition model. The specific determination process is to identify each spatial grid point one by one. If a spatial grid point meets all the landscape formation conditions of a landscape type, then the spatial grid point is marked as a suitable distribution area of ​​that landscape type.

[0066] Specifically, for each spatial grid point belonging to the landscape suitability zone, the probability value of meeting the landscape formation conditions during the forecast period is calculated as follows: (1) Calculate the degree of satisfaction of each judgment condition in the landscape formation conditions for the spatial grid points. Since each judgment condition is defined by a set of thresholds, the actual values ​​of the spatial grid points are mapped to the interval [0, 1] based on the thresholds to obtain the degree of satisfaction of each judgment condition in the landscape formation conditions for the spatial grid points. For example, the landscape formation condition of the sea of ​​clouds is the inversion layer intensity. 0.5℃ / 100m and low cloud cover ≥60% and coupled humidity field RH 85%. If the actual values ​​of the meteorological data in the spatial grid point at the current moment are: inversion layer intensity of 0.7℃ / 100m, low cloud cover of 70%, and RH of the coupled humidity field of 88%, then the degree to which the spatial grid point satisfies each judgment condition can be obtained through the following transformation standard, which is specifically expressed as follows: ; In the formula, For the degree of satisfaction of a single item, These are the actual values ​​from the meteorological data. Using a threshold value, the degree of satisfaction of each individual condition in the landscape formation criteria is determined based on the transformation standard. For example, for the condition of inversion layer intensity, the threshold is 0.5℃ / 100m. If the actual value of the inversion layer intensity in the meteorological data at the current moment is 0.7℃ / 100m, then the degree of satisfaction of this condition at the current moment is... That is, the degree of satisfaction of a single condition is 1, and the degree of satisfaction of a single condition is a dimensionless specific value.

[0067] (2) The probability value corresponding to each judgment condition in the landscape formation conditions is obtained by weighted summation of the degree of satisfaction of each individual condition. The weight value of each judgment condition can be determined based on the sensitivity of the landscape occurrence to historical observation data. For example, for the cloud sea landscape, according to historical observation data, humidity is the direct material basis for the formation of cloud sea and is usually the most critical formation factor. Low cloud cover is the direct visual manifestation of cloud sea formation and is second only to humidity in importance. Temperature inversion intensity is an important dynamic and thermal condition for the formation of cloud sea and provides a stable environment for the maintenance of cloud layer. It can be regarded as a secondary triggering / maintenance condition. Therefore, the weight range of the degree of satisfaction of each judgment condition for cloud sea landscape can be set as follows: humidity field is Low cloud cover The inversion intensity is The optimal weight combination for the degree of satisfaction of each judgment condition for the sea of ​​clouds landscape can be as follows: humidity field is 0.50, low cloud cover is 0.30, and inversion intensity is 0.20. It should be noted that the sum of the weight values ​​of each judgment condition is 1.

[0068] It should be noted that the probability value means the degree to which the spatial grid point meets the landscape formation conditions corresponding to the landscape type at the current forecast time. The larger the probability value, the higher the degree to which the spatial grid point meets the landscape formation conditions corresponding to the landscape type, that is, the higher the probability that the grid point will have that landscape type. The smaller the probability value, the lower the degree to which the spatial grid point meets the landscape formation conditions corresponding to the landscape type, that is, the lower the probability that the grid point will have that landscape type.

[0069] S233. Generate a heat map of suitable landscape areas for each landscape type during the forecast period based on probability values ​​and suitable distribution areas.

[0070] If a spatial grid point satisfies all the landscape formation conditions of a landscape type, then the spatial grid point is marked as a suitable distribution area of ​​that landscape type. A landscape suitable area heat map is formed based on the probability values ​​of the landscape formation conditions of the spatial grid points in the landscape suitable area. Each grid point on the landscape suitable area heat map corresponds to a probability value between 0 and 1.

[0071] It should be noted that a spatial grid point may contain a probability value for at least one landscape type. The probability value of a spatial grid point for different landscape types is specifically obtained based on the identification and judgment results between the actual values ​​of meteorological data in the spatial grid point and the corresponding landscape formation conditions of each landscape type. Moreover, the probability value of a spatial grid point for different landscape types is represented independently. That is, the probability value of a spatial grid point for a sea of ​​clouds landscape may be 1, and the probability value for a sunrise landscape may be 0.8.

[0072] It should be noted that existing technologies can only determine whether a landscape type will appear, such as whether a sea of ​​clouds will appear. This step, however, can determine the location of the landscape type, such as which grid points in the target scenic area the sea of ​​clouds will appear. Furthermore, existing technologies output a probability value for the forecast result, such as a 70% probability of a sea of ​​clouds tomorrow. This step, on the other hand, outputs a heat map of suitable landscape areas. This spatial distribution map represents which locations in the target scenic area are likely to have a sea of ​​clouds and their probability of occurrence. In addition, existing technologies only use fixed-point forecasts, such as forecasts based on the locations of two weather stations. This step, however, makes point-by-point judgments on all spatial grid points in the target scenic area, resulting in a more accurate and comprehensive landscape forecast.

[0073] S3. Construct a multi-agent collaborative framework, extract multi-dimensional feature tensors and three-dimensional encoding vectors from the landscape suitability zone heatmap based on the multi-agent collaborative framework, and fuse the three-dimensional encoding vectors to obtain the landscape occurrence probability and confidence interval of the landscape node.

[0074] In this embodiment, step S3 includes: S31. Extract multidimensional feature tensors and three-dimensional encoding vectors from landscape suitability heatmaps based on a multi-agent collaborative framework.

[0075] Please see Figure 3 Step S31 includes: S311, the multi-agent collaboration framework includes a meteorological analysis agent, a terrain coupling agent, a landscape identification agent, and a fusion decision-making agent.

[0076] S312. Construct a meteorological analysis agent based on a time-series coding network, and use the meteorological feature sequence of the landscape node as the input of the meteorological analysis agent, and output the meteorological condition coding vector according to the meteorological analysis agent.

[0077] The meteorological analysis agent adopts a Transformer-based temporal coding network, which includes a 6-layer Transformer encoder, each layer containing 8 attention heads, a hidden layer dimension of 512 dimensions, and a feedforward network dimension of 2048 dimensions. The input of the meteorological analysis agent is the temporal meteorological feature sequence of each landscape node in the past 72 hours. After supplementing the temporal meteorological feature sequence with temporal location information, it is input into the Transformer encoder and outputs a meteorological condition coding vector. The specific training method for the meteorological analysis agent is as follows: historical landscape observation records are used as training data to train the meteorological analysis agent. These historical landscape observation records can be obtained from the professional observation records of scenic area staff. Furthermore, the optimizer for the meteorological analysis agent is the Adam optimizer, with a learning rate initialized to 1×10⁻⁶. -4The cosine annealing decay strategy is adopted, with a training batch size of 32 and a training round size of 100. An early stopping strategy is adopted, which terminates the training early if the loss on the validation set does not decrease for 10 consecutive rounds.

[0078] S313. Construct a terrain-coupled intelligent agent based on a convolutional neural network, using the meteorological topographic field and the heat map of suitable landscape areas as inputs to the terrain-coupled intelligent agent, and outputting a terrain-landscape coupled encoding vector based on the terrain-coupled intelligent agent.

[0079] The terrain-coupled agent is constructed based on a convolutional neural network. The parameters of the convolutional neural network are three convolutional layers, each with a kernel size of 3×3 and 64, 128, and 256 channels, respectively. It is followed by a global average pooling layer and a fully connected layer to encode the meteorological and topographic field into a terrain-landscape coupled coding vector. The terrain-landscape coupled coding vector is used to characterize the real-time modulation effect of terrain on landscape formation conditions under the current meteorological conditions.

[0080] S314. Construct a landscape identification agent based on a residual network, using the visual features of landscape nodes as the input of the landscape identification agent, and outputting a visual state encoding vector based on the landscape identification agent. Specifically, the visual features of the landscape nodes are the keyframes of the current time and the past 6 hours, with a sampling interval of 10 minutes. The landscape identification agent adopts a residual network with a ResNet-50 architecture. After removing the classification layer, it is connected to a global average pooling layer and a fully connected layer to output a visual state encoding vector after inputting visual features.

[0081] The training method for the residual network in the landscape identification agent is as follows: after pre-training on ImageNet, the residual network is fine-tuned on a scenic area monitoring image dataset with a learning rate of 1×10. -5 The training batches are 16, and the training rounds are 20.

[0082] S315. A three-dimensional coding vector is constructed based on the meteorological condition coding vector, the terrain landscape coupled coding vector, and the visual state coding vector.

[0083] The three-dimensional encoding vector includes meteorological condition encoding vector, topographic landscape coupling encoding vector, and visual state encoding vector.

[0084] S32. The three-dimensional encoding vectors are fused to obtain the landscape occurrence probability and confidence interval of the landscape nodes.

[0085] Specifically, step S32 includes: S321. The fusion decision-making agent, based on the cross-attention fusion mechanism, performs deep fusion of multi-source features, including meteorological condition encoding vector, terrain landscape coupled encoding vector, and visual state encoding vector, to obtain fused features.

[0086] Since meteorology is the fundamental driving force for landscape formation, the core variable for the probability of landscape occurrence is the spatiotemporal evolution of meteorological elements (humidity, cloud cover, wind speed, etc.), while topographic conditions and visual state are auxiliary conditions. That is, topographic conditions determine where the landscape will appear if meteorological conditions are met, and visual state determines whether the current landscape has actually appeared and continues to exist. Therefore, the meteorological condition encoding vector is used as the query for the cross-attention fusion mechanism, and the topographic landscape coupling encoding vector and the visual state encoding vector are concatenated and merged. The concatenated topographic landscape coupling encoding vector and the visual state encoding vector are used as the index labels of the information to be retrieved by the cross-attention fusion mechanism and the final extraction of actual feature data. The correlation weight between meteorology and topography / visual data is calculated through the cross-attention mechanism to complete the feature fusion of three types of data: meteorology, topography, and remote sensing images.

[0087] The specific fusion process of the fusion features is as follows: (1) The concatenated vectors of meteorological condition coding vector, topographic landscape coupling coding vector, and visual state coding vector are mapped to a unified 256-dimensional space through linear projection matrices. The product of the weather condition encoding vector and the feature projection yields the weather matrix. ,based on Characterize current meteorological conditions; through a linear projection matrix The product of the concatenated vectors of the topographic landscape coupling encoding vector and the visual state encoding vector yields the concatenated matrix after feature projection. ,based on Representing noteworthy information in terms of terrain and visual aspects; through linear projection matrices The product of the concatenated vectors of the topographic landscape coupling encoding vector and the visual state encoding vector yields the concatenated matrix after feature projection. ,based on The specific content that represents information worthy of attention; (2) Calculate the meteorological matrix after feature projection With all spliced ​​matrices The similarity between them was normalized using a scaling factor, specifically by: [calculating the similarity between the meteorological matrices]. Dot product concatenation matrix The transpose of the result is divided by the square root of 256, and a softmax operation is performed on the result to obtain the attention score matrix. The attention score matrix is ​​a 256-dimensional attention weight distribution matrix, representing the degree of attention meteorological conditions place on each component of topographic and visual features. The square root of 256 is the scaling factor, and the softmax operation normalizes each row of the matrix, outputting a probability distribution that sums to 1. Specifically, the attention score matrix means that if topographic conditions play a decisive role in the formation of landscape type in a forecast, the dimension corresponding to the topographic feature in the attention score matrix receives a higher weight; if the visual state clearly indicates that a landscape has appeared, the dimension corresponding to the visual feature in the attention score matrix receives a higher weight.

[0088] (3) The attention score matrix is ​​used to perform a weighted summation of the concatenation matrix to obtain the fused features. Specifically, the attention score matrix and the concatenation matrix are summed. Performing matrix dot product yields the fused features, which integrate information from three sources: meteorology, topography, and remote sensing, with meteorological features serving as the primary basis for weight allocation.

[0089] It should be noted that the linear projection matrix , , Output weight matrix Bias terms It is obtained in the following ways: (1) Obtain the Xavier uniform initialization method known in the prior art. , , , , The initialization parameters are: a uniform distribution U initialized by Xavier uniform initialization; the sampling interval is from the square root of 6 divided by the input dimension plus the output dimension to the square root of 6 divided by the input dimension plus the output dimension; and the bias term is given. The initialization parameter is 0; (2) Based on the loss function of the fusion decision agent, historical landscape observation records are used as supervision signals, and the Adam optimizer is used for end-to-end training. During training, the learning rate is set to 5%. 10 -5 The batch size is 64. In each iteration, the gradient of the loss function with respect to each parameter is calculated using the backpropagation algorithm, and the parameters are updated using the gradient descent method. After multiple rounds of iterative training, the loss function value of the fused decision agent converges. The values ​​of each parameter at this point are saved, which is the training result. , , , , The final parameters.

[0090] The loss function of the fusion decision agent consists of the sum of the cross-entropy loss term and the regularization loss term. Specifically: The cross-entropy loss term is obtained by multiplying the natural logarithm of the landscape prediction probability by the true label value, subtracting the natural logarithm of the landscape prediction probability from 1 by subtracting the true label value from 1, summing the two products and taking the negative. A true label value of 1 indicates that the target landscape actually appeared at the corresponding time; a true label value of 0 indicates that the target landscape did not actually appear at the corresponding time. The true label value can be obtained from the professional observation records of the scenic area staff. The regularization loss term is applied by fusing all trainable parameters in the decision agent (i.e. , , , , The regularization loss term is obtained by multiplying the square of the Frobenius norm of the corresponding weight matrix parameters by a preset constant, where the default value of the preset constant is 1. 10 -5 .

[0091] The gradient descent method specifically involves dividing the gradient descent into two groups. , , , , The old values ​​of the parameters are subtracted from the learning rate multiplied by the partial derivative of the loss function with respect to the old values ​​of each parameter to obtain the new values ​​of the parameters. After multiple rounds of iterative training, the loss function value of the fused decision agent converges, and the values ​​of each parameter at this time are saved.

[0092] S322. Calculate the landscape prediction probability based on the fusion features, and perform a preset number of random forward calculations on the landscape prediction probability to obtain the landscape occurrence probability. Obtain the interval radius of the confidence interval based on the standard deviation of the landscape occurrence probability corresponding to the preset number of calculations.

[0093] The calculation of landscape prediction probability based on fused features refers to mapping the fused features to the final landscape prediction probability through a fully connected layer and a sigmoid activation function. Specifically, this involves multiplying the fused features by the output weight matrix. Add the bias term to the product result. The calculation results are obtained, and the sigmoid function is applied to the results to obtain the landscape prediction probability. The output weight matrix is ​​shown below. and bias terms It can be obtained through S321.

[0094] When using the Monte Carlo random inactivation method known in the prior art to perform a preset number of random forward propagation calculations on the landscape prediction probability, the preset number of calculations can be set to 30, that is, to perform 30 random forward propagation calculations on the landscape prediction probability, take the mean of the 30 output probabilities as the final landscape occurrence probability, and take twice the standard deviation of the 30 output probabilities as the radius of the confidence interval.

[0095] S4. Identify abnormal clustering events based on multi-source data, determine the implicit label of each landscape node based on the abnormal clustering events, calculate the prediction deviation of each landscape node based on the implicit label and the probability of landscape occurrence, and update the multi-agent collaboration framework based on the prediction deviation.

[0096] In this embodiment, step S4 includes: S41. Identify abnormal clustering events based on multi-source data.

[0097] Specifically, step S41 includes: S411. Obtain tourist location data from multi-source data, perform similarity clustering on the tourist location data to obtain the clustering results of each landscape node, and construct the time series curve of tourist density of each landscape node in the target scenic area based on the clustering results.

[0098] When acquiring tourist location data from multi-source data, the sampling interval can be 5 minutes. Specifically, the clustering results for each landscape node obtained by similarity clustering of tourist location data are as follows: (1) Use a density-based spatiotemporal clustering method to perform similarity clustering on tourist location data. The specific parameters of the spatiotemporal clustering method include spatial neighborhood radius, temporal neighborhood radius and minimum number of cluster points. The spatial neighborhood radius is preferably 50 meters, and the specific spatial neighborhood radius can be determined based on the average distance between landscape nodes in the scenic area. The temporal neighborhood radius is preferably 15 minutes, and the specific temporal neighborhood radius can be determined based on the average stay time of tourists at the landscape points. The minimum number of cluster points is preferably 30, and the specific number of cluster points can be determined based on the minimum statistical sample size requirement of the scenic area.

[0099] (2) Calculate the spatiotemporal joint distance between landscape nodes, identify two landscape nodes with a spatiotemporal joint distance less than or equal to 1 as spatiotemporal adjacency points, take the spatiotemporal adjacency point as a landscape node, and construct the time series curve of tourist density of each landscape node in the target scenic area based on each landscape node in the clustering results.

[0100] S412. Construct a baseline visitor density curve for each scenic node based on historical visitor distribution data within the target scenic area, and calculate the threshold value based on historical visitor distribution data.

[0101] Among them, the historical tourist distribution data refers to the tourist distribution data of non-scenic days in the same period of the past 30 days, that is, the tourist distribution data of ordinary days without special scenic events or large-scale activities. The median of the tourist distribution data is taken and a 3-point moving average is used to smooth and generate the benchmark tourist density curve for each scenic node. Specifically, the calculation of the threshold based on historical visitor distribution data involves multiplying the standard deviation of the historical baseline visitor density of each scenic node with an adjustable constant to calculate the threshold. The adjustable constant ranges from 1.5 to 3.0, and in this embodiment, the value of the adjustable constant is preferably 2.0.

[0102] S413. Compare the tourist density time series curve with the benchmark tourist density curve. If the tourist density time series curve of the landscape node exceeds the benchmark tourist density curve and the exceedance value is greater than the exceedance threshold, then it is determined that the landscape node has experienced an abnormal clustering event.

[0103] S42. Determine the implicit labels of each landscape node based on abnormal clustering events.

[0104] Specifically, step S42 includes: S421. Set up a multi-source auxiliary signal that includes time consistency verification signal, spatial consistency verification signal and visual verification signal, and set the signal weight of each verification signal in the multi-source auxiliary signal.

[0105] Among them, the time consistency verification signal refers to determining whether the occurrence time of the abnormal clustering event is consistent with the typical occurrence time of the target landscape. For example, the typical occurrence time of the sea of ​​clouds landscape is generally within 1 hour before and after sunrise; the typical occurrence time of the starry sky landscape is generally 2 hours after sunset to 2 hours before sunrise; and the typical occurrence time of the sunrise landscape is generally 30 minutes before and after sunrise.

[0106] Among them, the spatial consistency verification signal refers to determining whether the location of the abnormal clustering event is consistent with the area of ​​the landscape suitability zone heat map.

[0107] Among them, visual verification signal refers to the use of scenic area monitoring images to verify whether the target landscape actually appears if the landscape node is equipped with monitoring equipment. For example, for a sea of ​​clouds landscape, the cloud coverage rate is detected; for a starry sky landscape, the number of stars is detected.

[0108] The setting of signal weights for each verification signal in the multi-source auxiliary signals refers to setting signal weights based on the degree of influence of time consistency verification signal, spatial consistency verification signal and visual verification signal on the verification result. The sum of the signal weights of time consistency verification signal, spatial consistency verification signal and visual verification signal is 1. In this embodiment, the signal weights of time consistency verification signal, spatial consistency verification signal and visual verification signal are preferably 0.4, 0.3 and 0.3, respectively.

[0109] S422. Cross-validate the abnormal clustering events based on multi-source auxiliary signals to obtain signal validation results, and calculate the implicit labels by weighting the signal validation results according to the signal weights.

[0110] Specifically, cross-validation of anomalous clustering events is performed based on multi-source auxiliary signals to determine whether the anomalous clustering event satisfies any of the verification signals in the multi-source auxiliary signals. If the anomalous clustering event satisfies a certain verification signal in the multi-source auxiliary signals, the signal verification result of the anomalous clustering event for that verification signal is 1; if the anomalous clustering event does not satisfy a certain verification signal in the multi-source auxiliary signals, the signal verification result of the anomalous clustering event for that verification signal is 0. For example, if the anomalous clustering event satisfies the time consistency verification signal in the multi-source auxiliary signals, the signal verification result of the anomalous clustering event for the time consistency verification signal is 1; if the anomalous clustering event does not satisfy the time consistency verification signal in the multi-source auxiliary signals, the signal verification result of the anomalous clustering event for the time consistency verification signal is 0.

[0111] Among them, the signal weights selected based on S421, namely the signal weights of the time consistency verification signal, the spatial consistency verification signal, and the visual verification signal, are 0.4, 0.3, and 0.3, respectively. Based on this set of signal weights, the signal verification results of the abnormal clustering event for each verification signal in the multi-source auxiliary signal are weighted and calculated to obtain the implicit label of the abnormal clustering event. This implicit label is a value within the range of... The specific values ​​between.

[0112] S43. Calculate the forecast bias of each landscape node based on the implicit label and the probability of landscape occurrence, and update the multi-agent cooperation framework based on the forecast bias.

[0113] Specifically, step S43 includes: S431. Calculate the confidence weight based on the signal verification results and the number of verification signals participating in the verification among the multi-source auxiliary signals.

[0114] The signal verification results of each individual auxiliary signal in the multi-source auxiliary signal are accumulated, and the confidence weight is calculated based on the ratio of the accumulated result to the number of verification signals. It should be noted that if each individual auxiliary signal in the multi-source auxiliary signal passes the verification in the abnormal clustering event, the signal verification result of that individual auxiliary signal is 1; if each individual auxiliary signal fails the verification in the abnormal clustering event, the signal verification result of that individual auxiliary signal is 0.

[0115] S432. Calculate the forecast bias for each landscape node based on the confidence weight, implicit label, and landscape occurrence probability.

[0116] The specific calculation process for the prediction bias of landscape nodes is as follows: multiply the confidence weight by the implicit label and then by the logarithm of the probability of landscape occurrence to obtain the first product result. Take the logarithm of the difference between 1 and the probability of landscape occurrence, multiply the logarithmic difference by the difference between 1 and the implicit label, and multiply the difference by the confidence weight to obtain the second product result. Take the sum of the first product result and the second product result to obtain the prediction bias of the landscape node.

[0117] S433. Based on the prediction bias, perform online incremental learning on each agent in the multi-agent cooperation framework to update the multi-agent cooperation framework.

[0118] Among them, based on the prediction bias, the elastic weight consolidation method is used to carry out online incremental learning for each agent in the multi-agent cooperative framework. The physical meaning of the elastic weight consolidation method is to apply strong constraints to the important parameters of each agent when updating the parameters of each agent, so that the change range is as small as possible. The elastic weight consolidation method is embodied in a loss function, which is obtained by adding the forward loss and the regularization constraint. Specifically, the forward loss is the prediction bias, and the regularization loss is calculated by squaring the difference between the current parameter value and the optimal parameter value, multiplying the result by the parameter importance, summing the products of all parameters, and multiplying the sum by 0.5 times the regularization strength to obtain the regularization loss. The default value of the regularization strength is 100, and the parameter importance is obtained by taking the expected value of the square of the first-order partial derivative of the prediction bias with respect to the current parameter value. Specifically, updating the parameters of each agent involves: In the meteorological analysis agent: parameters related to key meteorological elements such as cloud cover, humidity, and temperature stratification are assigned higher importance weights. Specifically, the importance of these parameters is multiplied by 2.0 to obtain the adjusted importance weights.

[0119] In the terrain-coupled agent: a higher importance weight is set for the physical coupling operator related to the calculation of terrain uplift condensation height, specifically by multiplying the parameter importance by 2.5 to obtain the adjusted importance weight.

[0120] In the landscape recognition agent: the standard importance of the visual feature parameters is set to 1.0.

[0121] In the fusion decision agent: the standard importance of the weight matrix parameter in the cross-attention mechanism is set to 1.0.

[0122] The visual interface can be presented via a display screen or by displaying a report.

[0123] Please see Figure 4 Example 2 Figure 4 This invention presents a multi-source fusion-based intelligent tourism meteorological landscape probability forecasting system, comprising: The data acquisition module is used to acquire multi-source data of each landscape node in the target scenic area and construct multi-dimensional feature tensors of the landscape nodes based on the multi-source data. The thermal distribution module is used to construct the meteorological and topographic field of the target scenic area during the forecast period based on the multidimensional feature tensor, identify the suitable distribution area of ​​the landscape type based on the meteorological and topographic field, and generate a heat map of the suitable area of ​​each landscape type during the forecast period based on the suitable distribution area. The landscape occurrence probability module is used to construct a multi-agent collaborative framework. Based on the multi-agent collaborative framework, multi-dimensional feature tensors and three-dimensional encoding vectors from the landscape suitability zone heatmap are extracted. The three-dimensional encoding vectors are fused to obtain the landscape occurrence probability and confidence interval of the landscape node. The forecast update module is used to identify anomalous clustering events based on multi-source data, determine the implicit labels of each landscape node based on the anomalous clustering events, calculate the forecast deviation of each landscape node based on the implicit labels and the probability of landscape occurrence, and update the multi-agent collaboration framework based on the forecast deviation. The forecast results module is used to generate a visualized forecast results interface based on the landscape suitability zone heat map, the probability of landscape occurrence at each landscape node, and the confidence interval.

[0124] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0125] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0126] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0127] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0128] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

[0129] In conclusion, the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A probabilistic landscape forecasting method based on multi-source fusion for intelligent tourism meteorological agents, characterized in that, Includes the following steps: Acquire multi-source data of each landscape node in the target scenic area, and construct a multi-dimensional feature tensor of the landscape node based on the multi-source data; Based on the multidimensional feature tensor, a meteorological and topographic field of the target scenic area is constructed during the forecast period. Based on the meteorological and topographic field, suitable distribution areas of landscape types are identified, and a heat map of suitable areas for each landscape type during the forecast period is generated based on the suitable distribution areas. A multi-agent collaborative framework is constructed, and multi-dimensional feature tensors and three-dimensional encoding vectors from landscape suitability zone heatmaps are extracted based on the multi-agent collaborative framework. The three-dimensional encoding vectors are then fused to obtain the landscape occurrence probability and confidence interval of landscape nodes. Based on the multi-source data, abnormal clustering events are identified, and the implicit labels of each landscape node are determined based on the abnormal clustering events. The prediction deviation of each landscape node is calculated based on the implicit labels and the probability of landscape occurrence, and the multi-agent cooperation framework is updated based on the prediction deviation. A visualized forecast result interface is generated based on the heat map of suitable landscape areas, the probability of landscape occurrence at each landscape node, and the confidence interval.

2. The method for probabilistic landscape forecasting based on multi-source fusion of tourism meteorological intelligent agents according to claim 1, characterized in that, Constructing multidimensional feature tensors for landscape nodes based on multi-source data, including: The multi-source data includes meteorological observation data, numerical weather prediction data, satellite remote sensing data, topographic elevation data, scenic area monitoring images, and tourist location data; The multi-source data is time-aligned and standardized, and meteorological features, remote sensing features, topographic features, visual features, and tourist distribution features of each landscape node are extracted based on the processed multi-source data. Based on the meteorological features, remote sensing features, topographic features, visual features, and tourist distribution features, feature vectors for each landscape node are formed, and a multidimensional feature tensor for the target scenic area is constructed based on the feature vectors.

3. The method for probabilistic landscape forecasting based on multi-source fusion of tourism meteorological intelligent agents according to claim 1, characterized in that, The meteorological and topographic field of the target scenic area during the forecast period is constructed based on the multidimensional feature tensor, including: A three-dimensional topographic feature field and a three-dimensional meteorological distribution field of the target scenic area are constructed based on the multidimensional feature tensor. The terrain slope, the angle between wind direction and slope aspect, and the lifting condensation height are obtained based on multi-source data. The terrain lifting operator is calculated based on the terrain slope, the angle between wind direction and slope aspect, and the lifting condensation height. The terrain height and the thickness of the incoming air layer are obtained based on multi-source data, and the terrain blocking operator is calculated based on the terrain height and the thickness of the incoming air layer. The angle between the valley direction and the wind direction, as well as the height difference of the valley terrain, are obtained based on multi-source data. The narrowing effect operator is calculated based on the angle and the height difference of the valley terrain. The angle relationship between slope aspect and solar radiation is obtained based on multi-source data, and a thermodynamic modulation operator is calculated based on the angle relationship and solar radiation intensity. A physical coupling operator is constructed based on the topographic lifting operator, topographic blocking operator, funnel effect operator, and thermal modulation operator. The three-dimensional topographic feature field and the three-dimensional meteorological distribution field are physically coupled based on the physical coupling operator to obtain the meteorological topographic field.

4. The probabilistic landscape forecasting method based on multi-source fusion for tourism meteorological intelligent agents according to claim 1, characterized in that, Based on suitable distribution areas, heat maps of suitable landscape areas for each landscape type during the forecast period are generated, including: Construct a landscape formation condition model, which includes the landscape formation conditions corresponding to each landscape type; Based on the landscape formation conditions and the meteorological characteristics of each spatial grid point in the meteorological and topographic field, the suitable distribution area of ​​the landscape type is identified, and the probability value of the spatial grid point in the suitable distribution area meeting the landscape formation conditions during the forecast period is calculated. Based on the probability values ​​and suitable distribution areas, a heat map of suitable landscape areas for each landscape type is generated during the forecast period.

5. The probabilistic landscape forecasting method based on multi-source fusion for intelligent tourism meteorological agents according to claim 1, characterized in that, Based on a multi-agent collaborative framework, multi-dimensional feature tensors and three-dimensional encoded vectors from landscape suitability zone heatmaps are extracted, including: The multi-agent collaborative framework includes a meteorological analysis agent, a terrain coupling agent, a landscape identification agent, and a fusion decision-making agent. A meteorological analysis agent is constructed based on a temporal coding network, and the meteorological feature sequence of landscape nodes is used as the input of the meteorological analysis agent. The meteorological condition coding vector is output based on the meteorological analysis agent. A terrain-coupled agent is constructed based on a convolutional neural network. The meteorological topographic field and the heat map of suitable landscape areas are used as inputs to the terrain-coupled agent. The terrain-landscape coupled encoding vector is output based on the terrain-coupled agent. A landscape identification agent is constructed based on a residual network. The visual features of the landscape nodes are used as the input of the landscape identification agent, and the visual state encoding vector is output based on the landscape identification agent. The three-dimensional coding vector is constructed based on the meteorological condition coding vector, the terrain landscape coupling coding vector, and the visual state coding vector.

6. The probabilistic landscape forecasting method based on multi-source fusion for tourism meteorological intelligent agents according to claim 5, characterized in that, The landscape occurrence probability and confidence interval of the landscape node are obtained by fusing the three-dimensional encoded vectors, including: The fusion decision agent is based on a cross-attention fusion mechanism, which deeply fuses the meteorological condition encoding vector, the terrain landscape coupled encoding vector, and the visual state encoding vector to obtain fused features. The landscape prediction probability is calculated based on the fusion features, and the landscape occurrence probability is obtained by performing a preset number of random forward calculations on the landscape prediction probability. The radius of the confidence interval is obtained based on the standard deviation of the landscape occurrence probability corresponding to the preset number of calculations.

7. The method for probabilistic landscape forecasting based on multi-source fusion of tourism meteorological intelligent agents according to claim 1, characterized in that, Identifying anomalous clustering events based on the multi-source data includes: Tourist location data is obtained from multi-source data. Similarity clustering is performed on the tourist location data to obtain the clustering results of each landscape node. Based on the clustering results, a time series curve of tourist density of each landscape node in the target scenic area is constructed. Based on the historical visitor distribution data of each scenic spot within the target scenic area, a baseline visitor density curve for each scenic spot is constructed, and the threshold exceeding the limit is calculated based on the historical visitor distribution data. The time-series curve of tourist density is compared with the benchmark tourist density curve. If the time-series curve of tourist density of a landscape node exceeds the benchmark tourist density curve and the magnitude of the exceedance is greater than the threshold, it is determined that an abnormal clustering event has occurred at the landscape node.

8. The method for probabilistic landscape forecasting based on multi-source fusion of tourism meteorological intelligent agents according to claim 1, characterized in that, The implicit labels of each landscape node are determined based on anomaly clustering events, including: Set up a multi-source auxiliary signal that includes time consistency verification signal, spatial consistency verification signal and visual verification signal, and set the signal weight of each verification signal in the multi-source auxiliary signal; Cross-validation of anomalous clustering events is performed based on multi-source auxiliary signals to obtain signal validation results. The implicit labels are then obtained by weighting the signal validation results according to the signal weights.

9. The probabilistic landscape forecasting method based on multi-source fusion for tourism meteorological intelligent agents according to claim 8, characterized in that, A multi-agent cooperation framework based on forecast bias updates includes: Calculate the confidence weight based on the signal verification results and the number of verification signals participating in the verification among the multi-source auxiliary signals; Based on the confidence weights, implicit labels, and landscape occurrence probabilities, the prediction bias for each landscape node is calculated. Based on the prediction bias, online incremental learning is performed on each agent in the multi-agent cooperation framework to update the multi-agent cooperation framework.

10. A system using the multi-source fusion-based tourism meteorological intelligent agent landscape probabilistic forecasting method as described in any one of claims 1-9, comprising: The data acquisition module is used to acquire multi-source data of each landscape node in the target scenic area and construct multi-dimensional feature tensors of the landscape nodes based on the multi-source data. The thermal distribution module is used to construct the meteorological and topographic field of the target scenic area during the forecast period based on the multidimensional feature tensor, identify the suitable distribution area of ​​the landscape type based on the meteorological and topographic field, and generate a heat map of the suitable area of ​​each landscape type during the forecast period based on the suitable distribution area. The landscape occurrence probability module is used to construct a multi-agent collaborative framework. Based on the multi-agent collaborative framework, multi-dimensional feature tensors and three-dimensional encoding vectors from the landscape suitability zone heatmap are extracted. The three-dimensional encoding vectors are fused to obtain the landscape occurrence probability and confidence interval of the landscape node. The forecast update module is used to identify abnormal clustering events based on the multi-source data, determine the implicit label of each landscape node based on the abnormal clustering events, calculate the forecast deviation of each landscape node based on the implicit label and the probability of landscape occurrence, and update the multi-agent cooperation framework based on the forecast deviation. The forecast results module is used to generate a visualized forecast results interface based on the landscape suitability zone heat map, the probability of landscape occurrence at each landscape node, and the confidence interval.

Images