Disaster early warning information issuing method and system based on cooperation of multi-modal product and intelligent agent

Abstract

The invention provides a disaster early warning information issuing method and system based on cooperation of a multi-modal product and an intelligent agent. The method and system are used for improving comprehensiveness and accuracy of disaster early warning and timeliness and effectiveness of emergency response. The method comprises the following steps: acquiring a multi-source monitoring data set comprising a satellite cloud picture data unit and a ground observation data unit, wherein the multi-source monitoring data set is acquired and generated by monitoring nodes distributed at different spatial positions in real time; joint feature analysis processing is performed on the multi-source monitoring data set through a multi-agent cooperative processing network, disaster risk trend features are generated, and the disaster risk trend features reflect a time evolution rule and a spatial diffusion mode of a disaster occurrence probability; calling a reinforcement learning algorithm to carry out early warning strategy optimization processing on the disaster risk trend characteristics, and generating an optimized early warning strategy containing early warning content types and delivery opportunities; and executing early warning content delivery operation according to the optimized early warning strategy, and linking an emergency response system to complete emergency response cooperative operation.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of data analysis, and particularly relates to a disaster early warning information publishing method and system based on multi-modal product and agent cooperation. BACKGROUND

[0002] In the field of agents and disaster early warning technology, with the development of science and technology, people pay more and more attention to the accuracy and timeliness of disaster early warning. In the early stage, disaster early warning mainly relies on a single data source, which is difficult to comprehensively grasp disaster-related information. Later, although attempts have been made to use multi-source data, the data fusion and analysis capability is limited. In the aspect of agent application, a simple agent model can only handle a single type of task and cannot cooperate with multi-source data for comprehensive analysis.

[0003] However, the existing multi-source data acquisition technology lacks systematicness, and different data sources cannot be effectively integrated, resulting in missing or conflicting data, which affects the accuracy of early warning. In addition, in the analysis of disaster risk and the formulation of early warning strategies, traditional methods are difficult to accurately capture the time evolution and spatial diffusion pattern of the probability of disaster occurrence, cannot adjust the early warning strategy according to the real-time risk, and lack effective linkage between early warning and emergency response, often resulting in problems such as untimely information transmission and unreasonable resource allocation. SUMMARY

[0004] The application provides a disaster early warning information publishing method and system based on multi-modal product and agent cooperation, to improve the comprehensiveness, accuracy of disaster early warning, and timeliness and effectiveness of emergency response.

[0005] In a first aspect, the application embodiment provides a disaster early warning information publishing method based on multi-modal product and agent cooperation, applied to a disaster early warning information publishing system, the method comprising: acquiring a multi-source monitoring data set containing a satellite cloud image data unit and a ground observation data unit, the multi-source monitoring data set being generated by monitoring nodes distributed in different spatial positions in real time; performing joint feature analysis processing on the multi-source monitoring data set through a multi-agent cooperative processing network to generate disaster risk trend features, the disaster risk trend features reflecting the time evolution law and spatial diffusion pattern of the probability of disaster occurrence; calling a reinforcement learning algorithm to perform early warning strategy optimization processing on the disaster risk trend features to generate an optimized early warning strategy containing early warning content types and delivery timing; performing early warning content delivery operations according to the optimized early warning strategy, and linking an emergency response system to complete emergency response cooperation operations.

[0006] In a second aspect, the application embodiment provides a disaster early warning information publishing system, which comprises a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the above method.

[0007] In a third aspect, the embodiments of the present application provide a computer readable storage medium comprising a computer program, when the computer program runs on a disaster early warning information publishing system, the computer program is used to make the disaster early warning information publishing system execute the steps of the above method.

[0008] In the embodiments of the present application, first, a multi-source monitoring data set containing satellite cloud image data units and ground observation data units is acquired, which is collected in real time by monitoring nodes distributed in different spatial positions, can comprehensively and timely collect multi-dimensional environmental data, provide rich and accurate data basis for disaster early warning, and make the data more complete and timely. Then, the multi-source monitoring data set is jointly analyzed and processed by a multi-agent collaborative processing network to generate disaster risk trend features reflecting the time evolution law and spatial diffusion mode of disaster occurrence probability, which can accurately analyze disaster risk from time and space dimensions, mine complex potential relationships between data, and improve the ability to grasp the development trend of disaster. Then, a reinforcement learning algorithm is called to optimize the disaster risk trend features to generate an optimized early warning strategy, which can accurately match the early warning content type and delivery time according to different disaster risks, and improve the pertinence and effectiveness of early warning. Finally, the early warning content is delivered according to the optimized early warning strategy, and the emergency response system is linked, which can realize efficient collaboration, ensure timely transmission of early warning information, and quickly launch emergency response, effectively reducing the loss caused by disasters.

[0009] As can be seen, the embodiments of the present application innovatively jointly analyze multi-source monitoring data through a multi-agent collaborative processing network, optimize early warning strategies by using a reinforcement learning algorithm, and realize efficient collaboration of early warning and emergency response, effectively making up for the shortcomings of the prior art, and improving the comprehensiveness, accuracy of disaster early warning, and timeliness and effectiveness of emergency response. BRIEF DESCRIPTION OF DRAWINGS

[0010] Figure 1 A flowchart of a disaster early warning information publishing method based on multi-modal product and agent collaboration provided by the embodiments of the present application.

[0011] Figure 2 A structural diagram of a disaster early warning information publishing system provided by the embodiments of the present application. DETAILED DESCRIPTION

[0012] Reference Figure 1 It is a disaster early warning information publishing method based on multi-modal product and agent collaboration provided by the embodiments of the present application, which can be applied to a disaster early warning information publishing system, and the specific steps are as steps 110-140.

[0013] Step 110: Obtain a multi-source monitoring data set containing satellite cloud map data units and ground observation data units, which is generated by monitoring nodes distributed in different spatial positions in real time.

[0014] In the embodiments of the present application, monitoring nodes in different spatial positions undertake the task of collecting data in real time. These nodes are widely distributed and targeted, and collect data from different levels such as satellite platforms and ground monitoring sites.

[0015] Taking a large-scale area as a target monitoring area as an example, in the monitoring system of the area, remote sensing monitoring equipment is deployed on the satellite platform, and multiple ground monitoring sites are set on the ground. The satellite platform can observe a large area of the target area due to its high altitude advantage, and the remote sensing monitoring equipment carried by the satellite platform collects data at a preset time interval, thereby generating satellite cloud map data units. The ground monitoring sites are closely related to the ground environment and can accurately obtain the real-time conditions of the ground, and synchronously collect ground observation data units with timestamps aligned with the satellite cloud map data units. These data units reflect the environmental characteristics of the target area from different angles, providing comprehensive and detailed data support for subsequent disaster warning analysis.

[0016] As an implementation manner, the obtaining of the multi-source monitoring data set containing the satellite cloud map data units and the ground observation data units comprises: Step 111: Start the remote sensing monitoring equipment deployed on the satellite platform, and collect satellite cloud map data units covering the target area at a preset time interval, wherein the satellite cloud map data units contain cloud system structure features and water vapor content distribution information.

[0017] In this step, the satellite platform is a key carrier for high-altitude data collection, and the remote sensing monitoring equipment deployed thereon is started. The equipment scans and collects data on the target area according to a pre-set time interval. The pre-set time interval is determined based on a comprehensive consideration of factors such as the meteorological change characteristics of the target area and the data collection accuracy requirements.

[0018] For example, for a region with frequent weather changes, a shorter preset time interval, such as every 15 minutes, can be set to capture subtle changes in cloud structure and water vapor content; for a region with relatively stable weather conditions, the preset time interval can be appropriately extended, for example, every 30 minutes. The collected satellite cloud image data unit carries rich information, and the cloud structure features can be reflected by the shape and distribution range of the cloud system. Different cloud shapes and distributions often indicate different weather trends. The water vapor content distribution information reflects the content and distribution of water vapor in the atmosphere. Water vapor, as one of the key factors in the formation of many meteorological disasters, its distribution is of great significance for disaster warning.

[0019] Step 112: Start the meteorological observation equipment deployed at the ground monitoring station to synchronously collect ground observation data units aligned with the timestamp of the satellite cloud image data unit, which contains ground temperature distribution features, precipitation intensity distribution features, and wind speed and direction change features.

[0020] In this step, the meteorological observation equipment deployed at each ground monitoring station is started. These devices closely monitor the ground environment and ensure that the collected ground observation data units are strictly aligned with the timestamp of the satellite cloud image data unit.

[0021] Taking a network composed of multiple ground monitoring stations as an example, each station is equipped with instruments specifically designed to measure ground temperature, precipitation intensity, and wind speed and direction. These instruments continuously collect data to reflect the real-time state of the ground environment. Ground temperature distribution features can show temperature differences in different regions, and abnormal temperature rise or fall in some regions may indicate potential weather changes. Precipitation intensity distribution features record precipitation conditions at different locations. By monitoring precipitation intensity, it can be determined whether there are weather phenomena such as heavy rain that may trigger disasters. Wind speed and direction change features help understand the flow of the atmosphere. Abnormal wind speed and direction changes may be a signal that a disaster is imminent. By synchronously collecting these data and aligning them with the timestamp of the satellite cloud image data unit, comprehensive and synchronous environmental monitoring from high altitude to the ground can be achieved, providing accurate and matched data basis for subsequent joint analysis.

[0022] Step 113: Perform radiation correction processing on the satellite cloud image data unit to eliminate brightness deviation caused by observation angle differences of remote sensing equipment, and generate a radiation-corrected satellite cloud image data unit.

[0023] Due to the difference in observation angle of remote sensing equipment when collecting satellite cloud image data units, the image brightness inevitably deviates, which affects the accurate analysis of cloud system structure and water vapor content and other information. Therefore, the satellite cloud image data units need to be processed for radiation correction.

[0024] In actual operation, the satellite cloud image data units are processed by using a radiation correction algorithm. The algorithm adjusts the brightness value of each pixel in the image by considering multiple factors such as the observation angle of the remote sensing equipment, the characteristics of the sensor, and the reflectivity of the target area. For example, for the area with low brightness due to the tilt of the observation angle, the algorithm calculates the appropriate correction factor according to the relevant parameters to improve the pixel brightness value of the area; on the contrary, for the area with high brightness, the algorithm will perform the corresponding reduction processing. After the above processing, the brightness deviation caused by the difference in observation angle can be effectively eliminated, so that the satellite cloud image data units can more accurately reflect the real situation of the target area, and generate the satellite cloud image data units after radiation correction, providing a reliable data basis for subsequent feature extraction and analysis.

[0025] Step 114: Perform outlier detection processing on the ground observation data unit, identify and eliminate abnormal observation values that exceed the statistical range of historical data, and generate the ground observation data unit after outlier elimination.

[0026] Optionally, during the collection of the ground observation data unit, it may be disturbed by various factors, resulting in some abnormal observation values. If these abnormal values are not processed, they will mislead the subsequent data analysis and disaster warning results. Therefore, the ground observation data unit needs to be processed for outlier detection.

[0027] The embodiment of the present application processes the ground observation data unit by using an outlier detection method based on statistical analysis. First, collect the same type of observation data in a long period of time in this area to establish a historical data statistical model. The model can reflect the distribution range and variation law of various observation data under normal circumstances. Then, compare each observation value in the currently collected ground observation data unit with the historical data statistical model. For example, for the ground temperature data, if an observation value is significantly higher or lower than the temperature value in the historical data statistical range, it is identified as an outlier; the precipitation intensity and wind speed and direction data are also judged by using similar methods. Once the outliers are identified, they are eliminated from the ground observation data unit, thereby generating the ground observation data unit after outlier elimination, ensuring the accuracy and reliability of the data, and providing high-quality data support for subsequent analysis.

[0028] Step 115: Align the radiation-corrected satellite cloud image data unit and the outlier-removed ground observation data unit by timestamp to generate a multi-source monitoring data set with time synchronization relationship.

[0029] In order to realize effective integration and joint analysis of high-altitude and ground data, the processed satellite cloud image data unit and ground observation data unit need to be aligned.

[0030] In specific operation, the timestamp is used as the key alignment basis. Since the time synchronization of the satellite cloud image data unit and the ground observation data unit has been ensured, in this step, the radiation-corrected satellite cloud image data unit and the outlier-removed ground observation data unit are matched by a data alignment algorithm. The algorithm traverses the timestamp information in the two data units and arranges the data with the same timestamp correspondingly. For example, the cloud system structure and water vapor content information in the satellite cloud image data unit collected at a certain time are associated with the ground temperature, precipitation intensity, and wind speed and direction information in the ground observation data unit at the same time to form a complete multi-source monitoring data set with time synchronization relationship. The above data set can organically combine data from different sources and provide a good data basis for subsequent joint feature analysis by a multi-agent collaborative processing network.

[0031] Step 120: Perform joint feature analysis on the multi-source monitoring data set by a multi-agent collaborative processing network to generate disaster risk trend features reflecting the time evolution law and spatial diffusion pattern of disaster occurrence probability.

[0032] After obtaining the multi-source monitoring data set with time synchronization relationship, these data are analyzed in depth by a multi-agent collaborative processing network to mine the disaster risk information contained therein and generate disaster risk trend features. The multi-agent collaborative processing network is composed of multiple agents, which have clear division of labor and mutual cooperation to complete the processing of the multi-source monitoring data set.

[0033] Taking a disaster monitoring scenario as an example, each agent in the multi-agent collaborative processing network is responsible for analyzing different types of data. Among them, for satellite cloud image data units and ground observation data units, agents will be assigned to perform feature extraction. Through the processing of data, various disaster-related features are extracted. Then, through the information exchange channel in the network, the features extracted by different agents are exchanged and integrated. On this basis, a corresponding model is constructed to analyze the relationship between these features, and then the disaster risk trend features reflecting the time evolution law and spatial diffusion mode of the disaster occurrence probability are generated. This process can make full use of the information in the multi-source monitoring data set to provide comprehensive and accurate risk assessment basis for disaster warning.

[0034] In an optional embodiment, the joint feature analysis and processing of the multi-source monitoring data set through the multi-agent collaborative processing network to generate disaster risk trend features comprises: Step 121: Assigning the satellite cloud image data unit in the multi-source monitoring data set to the first agent for cloud system evolution feature extraction processing to obtain cloud system structure change features and water vapor transport path features.

[0035] In the multi-agent collaborative processing network, the first agent is assigned the important task of feature extraction for satellite cloud image data units. The agent uses algorithms and models to process the characteristics of satellite cloud image data units to extract key features in the evolution of cloud systems.

[0036] Specifically, the first agent first conducts in-depth analysis on the satellite cloud image data unit, extracts cloud system structure change features from the aspects of cloud system morphology, distribution, etc. through a series of image processing and data analysis techniques. At the same time, the water vapor transport path features are determined by using tracking algorithms and models to analyze the movement trajectory of water vapor in the cloud system. For example, by comparing and analyzing satellite cloud image data units at consecutive time points, the changes in cloud system shape, size, position, etc. are observed to determine the changes in cloud system structure. By comprehensively analyzing the dynamic changes of water vapor content in the cloud system and the moving direction of the cloud system, a water vapor transport path model is constructed to obtain water vapor transport path features. These features are of great significance for understanding the evolution trend of the meteorological system and the potential disaster risk.

[0037] As a preferred implementation, the satellite cloud image data unit in the multi-source monitoring data set is assigned to the first agent for cloud system evolution feature extraction processing to obtain cloud system structure change features and water vapor transport path features, comprising: Step 1211: Perform image segmentation processing on the satellite cloud image data unit to divide the boundaries of different cloud system types and generate cloud system region segmentation results.

[0038] In this step, the first agent processes the satellite cloud map data unit using an image segmentation algorithm. The purpose of the image segmentation algorithm is to distinguish different cloud system types in the satellite cloud map and clearly define the boundaries of each cloud system region.

[0039] In specific operation, the algorithm classifies each pixel point in the satellite cloud map data unit based on the gray value, texture, and other characteristic information of the cloud system. For example, different types of cloud systems such as cumulus and stratus have obvious differences in gray value and texture. The algorithm identifies and analyzes these differences, classifies pixel points belonging to the same cloud system type into a region, and thus divides the boundaries of different cloud system types to generate a cloud system region segmentation result. The segmentation result clearly presents the distribution of different cloud systems and provides basic data for subsequent extraction of cloud system structure change features.

[0040] Step 1212: Perform morphological feature extraction processing on the cloud system region segmentation result, calculate the area change rate, shape complexity parameter, and boundary smoothness parameter of each cloud system region, and generate cloud system structure change features.

[0041] Based on the cloud system region segmentation result obtained in the previous step, the first agent further performs morphological feature extraction processing on each cloud system region. By calculating parameters to describe the morphological changes of the cloud system region, cloud system structure change features are generated.

[0042] Specifically, for each cloud system region, its area change rate is calculated. The area change rate is obtained by comparing the area sizes of the cloud system region at different time points using a corresponding calculation formula (such as area change rate = (current area - last time area) / last time area). The shape complexity parameter reflects the complexity of the shape of the cloud system region, which can be determined by calculating the relationship between the boundary length and the area. The boundary smoothness parameter is used to measure the smoothness of the cloud system region boundary, which can be calculated by analyzing the curvature and other indicators of the boundary. Through the calculation and analysis of these parameters, the changes of the cloud system structure at different times can be comprehensively described, and cloud system structure change features are generated, providing a strong basis for judging the meteorological change trend.

[0043] Step 1213: Perform optical flow field analysis processing on the satellite cloud map data unit to extract displacement vector distribution information of adjacent time-stamped cloud system regions.

[0044] In order to obtain the motion information of the cloud system in the time dimension, the first agent processes the satellite cloud map data unit using an optical flow field analysis algorithm. Based on the satellite cloud map data unit of adjacent time stamps, the optical flow field analysis algorithm analyzes the motion of pixel points in the image to extract the displacement vector distribution information of the cloud system region.

[0045] In practice, the algorithm compares satellite cloud images with adjacent timestamps to observe the positional changes of each pixel within the cloud region. Based on the correspondence between the brightness, color, and other characteristics of the pixels at different times, it calculates the displacement vector of each pixel. These displacement vectors constitute the displacement vector distribution information of the cloud region. For example, for a given cloud region, optical flow field analysis can reveal the direction and magnitude of motion of various parts within that region. This information can intuitively reflect the movement trend of the cloud system, providing important clues for subsequent tracking of water vapor transport paths.

[0046] Step 1214: Construct a water vapor transport path tracking model based on the displacement vector distribution information, calculate the migration direction and migration rate parameters of water vapor between different cloud regions through the water vapor transport path tracking model, and generate water vapor transport path features.

[0047] After obtaining the displacement vector distribution information of the cloud system region, the first intelligent agent uses this information to construct a water vapor transport path tracking model. This model combines the movement of the cloud system and the distribution changes of water vapor content to calculate the migration direction and migration rate parameters of water vapor between different cloud system regions.

[0048] When constructing the model, factors such as the direction and speed of cloud movement, as well as the gradient of water vapor content, are considered. For example, if a cloud system has a high water vapor content and is moving towards another cloud system, the model will calculate the direction and rate of water vapor migration from the high-content area to the low-content area based on the displacement vector distribution information and the changes in water vapor content. Through the calculations of the above model, water vapor transport path characteristics that accurately reflect the migration of water vapor between different cloud system areas are generated, providing key information for understanding the dynamic changes of water vapor in meteorological systems.

[0049] Step 122: Assign the ground observation data units in the multi-source monitoring data set to the second intelligent agent for surface environment feature extraction processing to obtain surface temperature change trend features, precipitation intensity accumulation features, and wind speed and direction fluctuation features.

[0050] In the multi-agent collaborative processing network, the second agent is responsible for feature extraction from ground observation data units. For different types of data contained in the ground observation data units, such as surface temperature, precipitation intensity, and wind speed and direction, the second agent uses appropriate analysis methods and models to process the data and extract key features that reflect changes in the surface environment.

[0051] Taking an actual monitoring scene as an example, the second intelligent agent determines the temperature change trend through analysis of the ground temperature data, performs statistics and analysis on the precipitation intensity data to obtain the cumulative characteristics of the precipitation intensity, and performs fluctuation analysis on the wind speed and wind direction data to obtain the fluctuation characteristics of the wind speed and wind direction. These characteristics can help understand the real-time change of the ground environment and potential disaster risk factors.

[0052] As another preferred implementation, the allocation of the ground observation data unit in the multi-source monitoring data set to the second intelligent agent for ground environment feature extraction processing to obtain the ground temperature change trend feature, the precipitation intensity cumulative feature, and the wind speed and wind direction fluctuation feature includes: Step 1221: performing time series analysis processing on the ground temperature data in the ground observation data unit to calculate temperature difference parameters and temperature change slope parameters at consecutive time points, and generating a ground temperature change trend feature.

[0053] The second intelligent agent processes the ground temperature data in the ground observation data unit by using a time series analysis method. The time series analysis method analyzes a series of ground temperature data arranged in time sequence to mine the change law of temperature over time.

[0054] In specific operation, the temperature difference parameters at consecutive time points are calculated, that is, the temperature at a later time is subtracted from the temperature at a previous time to reflect the change amount of the temperature in a short time. At the same time, the temperature change slope parameter is calculated, which is obtained by fitting the overall trend of the temperature change in a period of time to obtain the rate of change of the temperature over time. For example, by analyzing the ground temperature data collected every hour in a day, the temperature difference and the temperature change slope at each time point are calculated, which can clearly show the change trend of the ground temperature, and generate the ground temperature change trend feature to provide a basis for judging the ground temperature abnormality and the possible disaster.

[0055] Step 1222: performing cumulative amount calculation processing on the precipitation intensity data in the ground observation data unit to calculate total precipitation parameters and precipitation intensity peak parameters in a preset time window, and generating a precipitation intensity cumulative feature.

[0056] For the precipitation intensity data in the ground observation data unit, the second intelligent agent performs cumulative amount calculation processing. In a preset time window, the precipitation intensity data is statistically analyzed to obtain the total precipitation parameter and the precipitation intensity peak parameter.

[0057] The setting of the preset time window is determined according to actual needs and meteorological characteristics, for example, one hour, one day or one week, etc. In this time window, the precipitation intensity data collected at each time point is accumulated to obtain the total precipitation parameter. At the same time, the maximum value of the precipitation intensity in this time window, that is, the precipitation intensity peak value parameter, is found. Through the statistics and analysis of these parameters, the precipitation intensity accumulation characteristics are generated, which can intuitively reflect the precipitation situation in a period of time, and have important significance for judging whether there is a disaster risk related to precipitation such as flood.

[0058] Step 1223: performing fluctuation analysis processing on the wind speed and direction data in the ground observation data unit, calculating the wind speed difference value parameter and the wind direction deflection angle parameter of adjacent time points, and generating the wind speed and direction fluctuation characteristics.

[0059] The second intelligent agent processes the wind speed and direction data in the ground observation data unit by using the fluctuation analysis method. By calculating the wind speed difference value parameter and the wind direction deflection angle parameter of adjacent time points, the change of wind speed and direction is described.

[0060] In specific calculation, for wind speed, the wind speed difference value of adjacent time points is calculated, that is, the wind speed at the next time minus the wind speed at the previous time, to reflect the change amount of wind speed. For wind direction, the deflection angle of wind direction of adjacent time points is calculated, and the change degree of wind direction is measured by angle calculation. For example, wind speed and direction data is collected every ten minutes, and the wind speed difference value and wind direction deflection angle between each collection time point are calculated. These parameters can reflect the fluctuation of wind speed and direction, generate wind speed and direction fluctuation characteristics, and help to understand the instability degree of atmosphere and the disaster risk that may be caused, such as strong wind, storm, etc.

[0061] Step 123: through the information interaction channel of the multi-agent collaborative processing network, the cloud system structure change characteristics, the water vapor transport path characteristics, and the ground surface temperature change trend characteristics, the precipitation intensity accumulation characteristics and the wind speed and direction fluctuation characteristics are cross-correlated and analyzed, and the spatio-temporal coupling relationship between different characteristics is identified.

[0062] In the multi-agent collaborative processing network, the information interaction channel can realize the exchange and fusion of the features extracted by different agents. In this step, the cloud system structure change characteristics and the water vapor transport path characteristics extracted by the first intelligent agent are cross-correlated and analyzed with the ground surface temperature change trend characteristics, the precipitation intensity accumulation characteristics and the wind speed and direction fluctuation characteristics extracted by the second intelligent agent through the information interaction channel.

[0063] Taking a real meteorological monitoring scenario as an example, the change of cloud system structure may affect the ground temperature and precipitation. For example, when the cloud system structure becomes dense and moves slowly, it may block solar radiation, causing the surface temperature to drop; at the same time, the change of water vapor transport path may also change the precipitation intensity in some areas. The change of surface temperature will affect the convective motion of the atmosphere, and then affect the wind speed and direction.

[0064] By correlation analysis algorithm, these different types of features are comprehensively considered, and the algorithm analyzes the corresponding relationship of different features in time and space dimensions. In the time dimension, the correlation between the changes of cloud system structure, water vapor transport and the changes of ground temperature, precipitation intensity, wind speed and direction at the same time or different times is observed; in the spatial dimension, the relationship between the cloud system features of different regions and the environmental features of the corresponding regions on the ground is determined. For example, while analyzing the change of cloud system structure above a certain region, the changes of ground temperature, precipitation intensity and wind speed and direction in the region and surrounding areas are observed, so as to identify the spatio-temporal coupling relationship between different features. This spatio-temporal coupling relationship can reveal the internal relationship between meteorological elements, and provide an important basis for subsequent construction of disaster risk correlation model.

[0065] Step 124: Based on the spatio-temporal coupling relationship, a disaster risk correlation model is constructed, and the evolution parameters of disaster occurrence probability in different time windows are calculated through the disaster risk correlation model, to generate disaster probability time series features reflecting the time evolution law.

[0066] After identifying the spatio-temporal coupling relationship between different features, a disaster risk correlation model is constructed based on this, which aims to integrate various feature information related to disasters, and establish the relationship between these features and disaster occurrence probability through mathematical modeling and data analysis method.

[0067] Taking flood disaster as an example, cloud system structure change, water vapor transport path, ground temperature, precipitation intensity and wind speed and direction are closely related to the occurrence of flood. The change of cloud system structure may indicate the coming of heavy rainfall, and the accumulation of precipitation intensity and the influence of ground temperature and wind speed and direction on precipitation distribution and runoff may increase the probability of flood occurrence. The disaster risk correlation model quantitatively analyzes the spatio-temporal coupling relationship and comprehensively calculates different feature parameters.

[0068] In calculating the evolution parameters of disaster occurrence probability in different time windows, the model considers the weights of different features at different times and their interactions according to the set algorithm. For example, in a shorter time window, a sudden increase in precipitation intensity may have a greater impact on disaster occurrence probability; while in a longer time window, the continuous change of cloud system structure and the long-term trend of surface temperature may be more critical. Through dynamic analysis and calculation of these factors, the evolution parameters of disaster occurrence probability over time are obtained.

[0069] The above evolution parameters are arranged in chronological order to form disaster probability time series features reflecting the time evolution law. For example, record the evolution parameters of disaster occurrence probability at each time point in units of hours, days or weeks, to clearly show the change trend of disaster occurrence probability at different times, providing risk assessment basis in time dimension for disaster warning.

[0070] Step 125: According to the disaster probability time series features and the spatial diffusion direction parameters in the spatio-temporal coupling relationship, a disaster spatial diffusion model is constructed, and the expansion rate parameters of disaster influence range are calculated through the disaster spatial diffusion model to generate disaster influence area expansion features reflecting the spatial diffusion pattern.

[0071] Based on the disaster probability time series features generated in the previous step and the spatial diffusion direction parameters contained in the spatio-temporal coupling relationship, a disaster spatial diffusion model is constructed, which focuses on the propagation and expansion law of disaster in space, and predicts the change of disaster influence range through analysis of existing features.

[0072] Disaster probability time series features reflect the likelihood of disaster occurrence at different times, while spatial diffusion direction parameters in the spatio-temporal coupling relationship indicate the possible direction of disaster diffusion. For example, in meteorological disasters, the spatial diffusion direction is determined according to the moving direction of cloud system, water vapor transport path and terrain, etc. The disaster spatial diffusion model combines these information, considering the influence of geographical environment, population distribution and other factors on disaster diffusion in different regions.

[0073] In calculating the expansion rate parameters of disaster influence range, the model analyzes the changes of disaster occurrence probability and spatial diffusion direction at different time points, and calculates the expansion speed of disaster influence range in unit time combined with geographical information data. For example, for a rainfall event that may trigger a flood, according to the previous meteorological features and geographical information, the model calculates that the flood may spread to the surrounding area at a certain kilometer per hour.

[0074] Through the above calculation, disaster impact area expansion features reflecting the spatial diffusion pattern are generated, which are presented in the form of multi-dimensional data, including expansion rates in different directions, boundary ranges of expansion, and other information, comprehensively describing the spatial diffusion trend of disasters, providing risk assessment basis in the spatial dimension for disaster early warning, and helping to take preventive measures in different regions in advance.

[0075] Step 126: Perform feature fusion processing on the disaster probability time series features and the disaster impact area expansion features to generate disaster risk trend features.

[0076] In order to more comprehensively and accurately reflect the overall trend of disaster risk, it is necessary to perform feature fusion processing on the disaster probability time series features and the disaster impact area expansion features. These two features describe the changes of disaster risk from the time and spatial dimensions respectively, and through fusion, a comprehensive disaster risk trend feature can be formed.

[0077] For example, a corresponding feature fusion algorithm is used to process these two features, which takes into account the importance and correlation of the two features and assigns appropriate weights to features of different dimensions. For example, in some cases, the time evolution of disaster occurrence probability may be more critical to the overall risk trend, in which case a higher weight will be assigned to the disaster probability time series features. In other cases, the expansion of the disaster impact area may be more important, and accordingly the weight of the disaster impact area expansion features will be increased.

[0078] In the fusion process, the data of the two features are combined. For example, the probability value at each time point in the disaster probability time series is associated with the spatial expansion information in the corresponding disaster impact area expansion feature. In this way, a comprehensive feature set containing time and space information, i.e. disaster risk trend features, is generated, which can clearly show the evolution and diffusion trend of disaster risk in time and space, providing comprehensive and accurate basic data for subsequent optimization of early warning strategies and issuance of early warning information.

[0079] Step 130: Call the reinforcement learning algorithm to optimize the disaster risk trend features for early warning strategy, generating an optimized early warning strategy containing early warning content types and delivery timing.

[0080] After obtaining the disaster risk trend features, in order to develop more effective early warning strategies, a reinforcement learning algorithm is called to optimize them. Reinforcement learning algorithm adjusts its strategy according to the reward signals feedback by the agent and the environment, in order to maximize the target.

[0081] In the embodiments of the present application, disaster risk trend characteristics are used as input information for a reinforcement learning algorithm, which will explore the optimal early warning strategy based on these information. Taking a practical disaster early warning scenario as an example, different disaster risk trends may require different types of early warning content and delivery timing. For example, for a high-risk and rapidly developing disaster, intuitive and strong early warning information such as images and voice may need to be delivered immediately to attract the public's attention; while for a relatively low-risk and slowly developing disaster, text early warning information and delayed delivery strategy may be more appropriate.

[0082] The reinforcement learning algorithm tries different combinations of early warning content types and delivery timing, and evaluates the effects of these combinations according to the reward signals from the environment. The reward signals are set based on the disaster impact reduction effects of different action combinations in historical early warning events. Through multiple iterations and learning, the algorithm gradually finds the strategy that maximizes the reward function value, i.e. generates an optimized early warning strategy containing optimal early warning content types and optimal delivery timing. This optimized early warning strategy can accurately select the appropriate early warning method and release time according to different disaster risk trends, thereby improving the effectiveness and timeliness of early warning.

[0083] In one embodiment, the reinforcement learning algorithm is called to perform early warning strategy optimization processing on the disaster risk trend characteristics, generating an optimized early warning strategy containing early warning content types and delivery timing, including: Step 131: Use the disaster probability time series feature and disaster impact area expansion feature in the disaster risk trend characteristics as state input for the reinforcement learning algorithm.

[0084] The operation of the reinforcement learning algorithm depends on the accurate perception of the environment state. In this step, the disaster probability time series feature and disaster impact area expansion feature in the disaster risk trend characteristics are used as state input for the algorithm, which comprehensively describes the current disaster risk situation from the time and space dimensions, providing rich decision-making basis for the algorithm.

[0085] Taking a storm disaster as an example, the disaster probability time series feature can show the changes of the storm probability at different times, such as gradually increasing in the next few hours and then decreasing. The disaster impact area expansion feature can indicate the range of areas that may be affected by the storm and the expansion trend of the range over time, for example, the storm may first affect a certain area and then gradually spread to the surrounding areas.

[0086] It can be understood that the reinforcement learning algorithm can understand the dynamic changes of the current disaster risk after receiving these state inputs, and thus make more reasonable early warning strategy decisions based thereon. Through analysis of these features, the algorithm can determine the severity, development speed and possible impact range of the disaster, and thus provide a basis for selecting appropriate early warning content types and delivery times.

[0087] Step 132: Set a combination of early warning content types and delivery times as the action space of the reinforcement learning algorithm, the early warning content types including text early warning information, image early warning information and voice early warning information, and the delivery times including immediate delivery, delayed delivery and staged delivery.

[0088] In order to enable the reinforcement learning algorithm to explore different early warning strategies, the action space of the algorithm needs to be defined. In the embodiments of the present application, various combinations of early warning content types and delivery times are used as the action space.

[0089] Early warning content types cover a variety of forms. Text early warning information can describe disaster information, possible impacts and countermeasures in a concise and clear manner; image early warning information can show the impact range and dangerous areas of a disaster through intuitive graphics, enhancing public awareness of the disaster; voice early warning information can convey important information in a timely manner, especially for people who cannot view text or images in time.

[0090] In terms of delivery times, immediate delivery can convey early warning information to target audiences in the first time, which is suitable for emergency and dangerous situations; delayed delivery can release early warnings after obtaining more information or waiting for the right time to avoid unnecessary panic; staged delivery can gradually release early warning information of different degrees according to the development process of the disaster, guiding the public to take appropriate countermeasures.

[0091] For example, when facing an upcoming earthquake disaster, according to the predicted intensity and possible impact range of the earthquake, image and voice early warning information can be selected for immediate delivery to enable the public to quickly understand the danger; for some weather disasters that may last for a long time, such as rainstorm early warning, a staged delivery approach can be used, with preliminary text early warning information released first, and image and voice early warning information released in a timely manner as the disaster develops, to better guide the public to respond to the disaster. These different combinations constitute a rich action space for the reinforcement learning algorithm to explore and optimize.

[0092] Step 133: Construct a reward function based on the disaster impact reduction effects of different action combinations in historical early warning events.

[0093] It can be understood that the reward function is used in the reinforcement learning algorithm to evaluate the effect of different action combinations (i.e., the combination of early warning content types and delivery timing), and provides a feedback signal for the algorithm, so that the algorithm learns the optimal strategy. In this step, the construction of the reward function is based on the disaster impact reduction effect of different action combinations in historical early warning events.

[0094] By collecting and analyzing the data of historical early warning events, the early warning content types, delivery timing and the final disaster impact degree adopted in each event are recorded. For example, in a past flood early warning, the combination of instant image and voice early warning information was adopted, so that most residents were timely notified and took preventive measures, and the personnel casualties and property losses caused by the disaster were significantly reduced; while in another early warning, due to improper selection of early warning content types and delayed delivery timing, the disaster impact was greater.

[0095] Based on these historical data, different action combinations are evaluated, and the disaster impact reduction amplitude corresponding to each combination is calculated. The disaster impact reduction amplitude is quantitatively processed, for example, a normalized method can be used to convert the reduction amplitude into a value within a preset range. Then, according to these quantified values, an association relationship between historical early warning content types, historical delivery timing and the quantified indicators is established, and a reward value mapping table is formed.

[0096] Based on the reward value mapping table, a reward function is generated, and the output value of the reward function is the historical average reward value corresponding to the current action combination. For example, for a new early warning situation, if a similar action combination to a successful early warning in history is adopted, the reward function will output a relatively high reward value, encouraging the reinforcement learning algorithm to choose this combination more in the future; on the contrary, if the action combination is similar to a combination with poor effect in history, the reward function will output a lower reward value, guiding the algorithm to avoid selecting such combination. Through the above reward function configuration, the reinforcement learning algorithm can continuously learn more effective early warning strategies.

[0097] In an implementation manner, the constructing the reward function comprises: Step 1330: obtaining a record of implemented early warning strategies in historical disaster events, the early warning strategy record containing historical early warning content types, historical delivery timing and corresponding disaster impact degree data; performing normalization processing on the disaster impact degree data to generate a quantified indicator reflecting the disaster impact reduction amplitude; establishing an association relationship between historical early warning content types, historical delivery timing and the quantified indicator to generate a reward value mapping table; generating a reward function based on the reward value mapping table, and the output value of the reward function is the historical average reward value corresponding to the current action combination.

[0098] First, historical disaster event records of implemented early warning strategies are comprehensively collected, which details the types of early warning content used in each disaster early warning process, such as whether it is a text, image or voice warning; the timing of delivery, whether it is immediate delivery, delayed delivery or staged delivery; and the final disaster impact data, including the number of casualties, property damage amount, and disaster area range.

[0099] Next, the disaster impact data is normalized. Since different disaster events differ in scale and nature, directly comparing disaster impact data may not be accurate. Therefore, a normalization algorithm is used to convert the disaster impact data of different events into quantifiable indicators that are comparable. For example, according to the maximum and minimum values of historical data, each disaster impact data can be mapped to a fixed interval, such as [0, 1], with smaller values indicating a greater reduction in disaster impact.

[0100] Then, according to the normalized quantifiable indicators, the association between historical early warning content types, historical delivery timing and these indicators is established. For example, it is recorded that in the corresponding disaster scenario, the combination of immediate delivery of image early warning information corresponds to a quantifiable indicator of 0.2, indicating that this combination has achieved good results in reducing disaster impact; while the combination of delayed delivery of text early warning information corresponds to a quantifiable indicator of 0.8, indicating that the effect of this combination is relatively poor. By organizing these associations, a reward value mapping table is generated.

[0101] Finally, a reward function is generated based on the reward value mapping table. When the reward function receives the current action combination (i.e. the combination of early warning content type and delivery timing), it will search the reward value mapping table for a similar historical combination and output the historical average reward value corresponding to that combination. This reward value can reflect the performance of the current action combination in history, providing a clear optimization direction for the reinforcement learning algorithm, prompting the algorithm to continuously search for action combinations with higher reward values, thereby optimizing the early warning strategy.

[0102] Step 134: Through the policy gradient optimization process of the reinforcement learning algorithm, a mapping relationship between the state input and the action space is established, generating a policy network that maximizes the reward function value.

[0103] It can be understood that the reinforcement learning algorithm learns the optimal policy through the policy gradient optimization process. In this step, based on the previously set state input (disaster probability time series features and disaster impact area expansion features) and action space (combination of early warning content type and delivery timing), as well as the constructed reward function, the algorithm begins to optimize the policy.

[0104] The policy gradient optimization process is an iterative process. In each iteration, the algorithm selects an action (i.e., a combination of early warning strategies) from the action space based on the current state input and executes the action. Then, based on the reward value (calculated by the reward function) feedback from the environment, the algorithm calculates the policy gradient. The policy gradient represents the gradient of the reward function with respect to the policy parameters, indicating the direction in which the policy parameters should be adjusted to increase the reward function value.

[0105] By continuously adjusting the policy parameters, the algorithm gradually establishes a mapping relationship between the state input and the action space. For example, when the algorithm observes that the disaster probability time series feature shows that the disaster occurrence probability is rapidly rising and the disaster impact area is expanding rapidly, it will adjust to a strategy that is more inclined to immediately launch image and voice early warning information based on the direction of the policy gradient. As the number of iterations increases, the algorithm continuously optimizes this mapping relationship and gradually finds a policy network that maximizes the reward function value. This policy network can accurately select the optimal early warning content type and launch time (action) based on different disaster risk trend features (state input) to maximize the reduction of disaster impact. For example, for different types of disasters such as earthquakes, floods, and typhoons, and different disaster development stages, the policy network can provide the corresponding optimal early warning strategy, thereby achieving intelligent optimization of early warning strategies.

[0106] Step 135: Perform action selection processing on the current state input using the policy network to generate an optimized early warning strategy containing the optimal early warning content type and the optimal launch time.

[0107] After generating a policy network that maximizes the reward function value, the network is used to perform action selection processing on the current disaster risk trend features (i.e., the current state input), thereby obtaining an optimized early warning strategy containing the optimal early warning content type and the optimal launch time.

[0108] It can be understood that the policy network has learned the optimal early warning strategy to be taken under different disaster risk states. When receiving the current disaster probability time series feature and the disaster impact area expansion feature, the policy network will make a decision based on the mapping relationship established inside it.

[0109] For example, if the current disaster probability time series feature shows that the disaster occurrence probability will peak in the next few hours, and the disaster impact area expansion feature indicates that the disaster will quickly spread to a large area, the policy network may choose a combination of immediate image and voice early warning information as the optimal early warning strategy to ensure that the public can quickly obtain accurate and intuitive disaster information and take effective preventive measures.

[0110] In this way, the generated optimized early warning strategy can closely integrate the current disaster risk trends, select the most appropriate early warning content type and delivery time, and improve the relevance and effectiveness of early warning, providing scientific guidance for subsequent early warning content delivery and emergency response.

[0111] Step 140: Perform early warning content delivery operation according to the optimized early warning strategy, and coordinate emergency response system to complete emergency response operation.

[0112] After obtaining the optimized early warning strategy, the next key step is to accurately and timely deliver early warning information to the target audience and coordinate with the emergency response system to minimize disaster losses.

[0113] Taking a city facing possible rainstorm flood disaster as an example, the optimized early warning strategy specifies the early warning content type and delivery time. The early warning content type may be image and voice early warning information containing detailed disaster information, danger area prompt, and response suggestions, and the delivery time is immediate delivery.

[0114] According to this strategy, the system first calls the corresponding content generation module to generate early warning information that meets the requirements. The early warning information not only accurately conveys key information such as disaster occurrence probability and impact area, but also provides specific defense suggestions, such as advising residents to avoid low-lying areas and prepare emergency supplies in advance. The generated early warning information will be disseminated through the preset information delivery channel.

[0115] In an alternative embodiment, the early warning content delivery operation according to the optimized early warning strategy, and the coordination of the emergency response system to complete the emergency response operation, includes: Step 141: Extract the optimal early warning content type from the optimized early warning strategy, call the corresponding content generation module to generate early warning information that meets the optimal early warning content type, and the early warning information contains disaster occurrence probability, impact area and defense suggestions.

[0116] In this step, the system first accurately extracts the optimal early warning content type from the optimized early warning strategy. For example, if the optimal early warning content type is image early warning information, the system will quickly call the image generation module. The image generation module generates intuitive and clear visual images of disaster impact range based on relevant information in the disaster risk trend characteristics, such as the affected range shown by the disaster impact area expansion characteristic, and the assessment of the disaster occurrence probability. In the image, key information such as high-risk areas and possibly affected areas will be clearly identified, and text explanations will be added, including the approximate value of the disaster occurrence probability and defense suggestions for different areas, such as reminding residents in high-risk areas to evacuate to safe areas as soon as possible, and prompting to take protective measures in potentially affected areas.

[0117] If the optimal warning content type is text warning information, the system calls the text generation module. The text generation module generates structured text information based on the disaster occurrence probability, affected area, and defense suggestions in the disaster risk trend characteristics. The text content elaborates on the type of disaster, specific description of the occurrence probability (such as "there is a 70% chance of heavy rain and flood disaster in the next 6 hours"), the scope of the affected area (such as "it is expected to affect XX district, XX street, and multiple areas"), and specific defense suggestions (such as "residents should close doors and windows in time, move valuable items to high places, and avoid going out").

[0118] If the optimal warning content type is voice warning information, the system first calls the text generation module to generate structured text information, and then calls the speech synthesis module to convert the structured text information into natural speech signals to generate voice warning information. The voice warning information conveys disaster-related information in clear and concise language to ensure that residents can accurately receive and understand it.

[0119] Optionally, the calling of the corresponding content generation module to generate warning information conforming to the optimal warning content type comprises: Step 1410: When the optimal warning content type is text warning information, call the text generation module to generate structured text information based on the disaster occurrence probability, affected area, and defense suggestions in the disaster risk trend characteristics; when the optimal warning content type is image warning information, call the image generation module to generate a disaster impact range visualization image based on the disaster impact area expansion characteristics in the disaster risk trend characteristics; when the optimal warning content type is voice warning information, call the speech synthesis module to convert the structured text information into natural speech signals to generate voice warning information.

[0120] When the optimal warning content type is text warning information, the text generation module will organize and organize each item of information in the disaster risk trend characteristics. In detail, the text generation module will first introduce the basic situation of the disaster, i.e. the probability of disaster occurrence, using accurate language and data to describe the probability size, so that residents have a clear understanding of the possibility of disaster occurrence. Then, it specifies the affected area, clearly indicating which geographical areas are involved, making it easy for residents to determine whether they are within the affected area. Finally, specific defense suggestions are given for different situations, which are targeted and operable, for example, for heavy rain and flood disasters, residents are advised to prepare emergency lighting equipment, store a certain amount of water and food for daily use, etc.

[0121] When the optimal warning content type is image warning information, the image generation module generates a disaster impact range visualization image based on the disaster impact area expansion characteristics, using professional graphics rendering algorithms and geographic information data. The image is presented in an intuitive map form, with different colors or symbols distinguishing different levels of impact areas, such as red for high-risk areas and yellow for potentially affected areas. At the same time, necessary textual annotations are added to the image, indicating disaster type, occurrence probability and other key information, so that residents can understand the general situation of the disaster and the risk level of their own area at a glance.

[0122] When the optimal warning content type is voice warning information, the text generation module first generates structured text information, which is carefully edited and has simple and clear language. Then, the speech synthesis module converts the text into natural speech signals. During the conversion process, the rhythm, volume and timbre of the speech are adjusted according to different language styles and tone requirements to ensure that the speech is clear, smooth and easy to understand, and can accurately convey disaster-related information.

[0123] Step 142: Extract the optimal delivery time from the optimized warning strategy, and send the warning information to the preset information delivery channel according to the optimal delivery time. The information delivery channel includes the mobile phone SMS platform, social media platform and community radio system.

[0124] In this step, the system extracts the optimal delivery time from the optimized warning strategy. For example, if the optimal delivery time is immediate delivery, the system will immediately send the generated warning information to the preset information delivery channel.

[0125] For the mobile phone SMS platform, the system will organize the warning information according to the requirements of the SMS format, ensuring that the content is concise and the key information is prominent. Then, through the interface with the SMS service provider, the SMS is sent to the mobile phones of residents in the target area. In addition to containing disaster occurrence probability, impact area and defense suggestions, the SMS content also adds emergency prompt symbols such as "Emergency Warning" to attract the attention of residents.

[0126] For the social media platform, the system will adapt the warning information according to the characteristics and format requirements of different platforms. For example, on the microblog platform, the information will be published in the form of images and text, with the generated disaster impact range visualization image, and a vivid and concise text description will be written to attract user attention and guide them to view detailed information. On the WeChat public number, an article containing voice warning information will be published, allowing users to click and listen, and detailed disaster-related content will be elaborated in the article.

[0127] For the community broadcasting system, the system will send voice warning information to the broadcasting equipment in each community. The broadcasting equipment will play the warning information in the community according to the set frequency and volume, ensuring that residents in the community can hear it. The broadcast content is simple and loud, covering every corner of the community.

[0128] Step 143: Send a coordination request message containing disaster impact area expansion features to the emergency response system, which is used to instruct the emergency response system to start the resource allocation process.

[0129] At the same time as the warning information is put out, the system sends a coordination request message to the emergency response system, which mainly contains disaster impact area expansion features, which is an important basis for the emergency response system to allocate resources.

[0130] Taking fire disaster as an example, the disaster impact area expansion features will describe in detail the direction, speed and expected impact range of the fire spread. After receiving this information, the emergency response system can accurately determine the type and quantity of resources needed. For example, if the fire is likely to spread to a densely populated area, the emergency response system will prioritize the allocation of fire vehicles, rescue personnel and evacuation guide equipment and other resources to the corresponding area.

[0131] The coordination request message also contains basic information about the disaster, such as the type of disaster, the time of occurrence, etc., so that the emergency response system can fully understand the situation and develop a more reasonable resource allocation plan. After receiving the coordination request message, the emergency response system will quickly start the resource allocation process and organize all parties to participate in emergency preparedness and rescue work.

[0132] Step 144: Receive resource allocation status information returned by the emergency response system, which contains rescue material reserves, rescue team location and arrival time prediction.

[0133] After starting the resource allocation process, the emergency response system will feed back the resource allocation status information to the disaster warning system in real time, which is very important for the disaster warning system to understand the progress of the emergency response.

[0134] Rescue material reserve information can let the disaster warning system know whether the current material for disaster response is sufficient, such as whether there are enough fire extinguishers, rescue ropes, medical supplies, etc. If the reserves are insufficient, the disaster warning system can consider taking other measures, such as coordinating material reinforcements.

[0135] The rescue team position information enables the disaster warning system to clearly understand the distribution of rescue forces, which areas have already arrived by rescue teams, and which areas need further strengthening. Through this information, the disaster warning system can better convey the progress of rescue to the public, and stabilize public emotions.

[0136] The arrival time prediction information enables the disaster warning system to estimate the time of the rescue team arriving at the disaster area, which is very important for informing residents in advance about the time arrangement of rescue, for example, through warning information to inform residents that "the fire brigade is expected to arrive in 30 minutes, please everyone to remain calm and wait for rescue", so that residents have a clear expectation of rescue and enhance their confidence in coping with disasters.

[0137] Step 145: integrate the resource allocation status information with the warning information to generate comprehensive warning information containing emergency response progress, and re-release the comprehensive warning information through the information release channel.

[0138] In order to let the residents more comprehensively understand the disaster situation and the progress of emergency response, the system will integrate the resource allocation status information with the warning information.

[0139] First, extract the key content in the resource allocation status information, such as whether the rescue material reserve is sufficient, the position of the rescue team and the predicted arrival time, etc. Then, fuse these information with the disaster occurrence probability, affected area and defense suggestion in the warning information.

[0140] For example, the generated comprehensive warning information may be expressed as: "In the next 6 hours, there is a 70% possibility of rainstorm and flood disaster in this area, which is expected to affect multiple areas such as XX district, XX street, etc. Please residents to close the windows and doors in time, and move valuable items to high places to avoid going out. At present, the rescue material reserve is sufficient, and the fire brigade is on the way, and is expected to arrive at the disaster area in 20 minutes."

[0141] Then, the generated comprehensive warning information is re-released through the information release channel used before, that is, the mobile phone short message platform, social media platform and community broadcasting system. Through the re-release, residents can obtain the latest disaster and rescue information, better prepare for the disaster, and at the same time, feel the positive action of the emergency response system, and enhance their confidence in coping with disasters.

[0142] Preferably, the integration of the resource allocation status information with the warning information to generate comprehensive warning information containing emergency response progress comprises: Step 1451: extract the rescue material reserve, rescue team position and arrival time prediction in the resource allocation status information to generate emergency response progress description information.

[0143] At this step, the system accurately extracts key information such as the amount of rescue supplies, the location of rescue teams, and the predicted arrival time from the resource allocation status information returned by the emergency response system.

[0144] For the amount of rescue supplies, the quantity and type of each type of material are recorded in detail, such as "currently reserves 500 sandbags, 200 sets of life jackets, 100 boxes of drinking water, and other rescue supplies". For the location of rescue teams, the specific location of each team is clearly stated, such as "Fire Brigade One is located at the intersection of XX Road and XX Street, and is rapidly advancing towards the disaster area; the medical rescue team has arrived at XX Community and is ready to carry out rescue work at any time". The predicted arrival time is accurate to the minute, such as "Fire Brigade Two is expected to arrive at the core disaster area in 15 minutes".

[0145] Organize and organize these information to generate emergency response progress description information, which can clearly show the current progress of emergency response work and provide a basis for subsequent integration with warning information.

[0146] Step 1452: The emergency response progress description information is integrated with the disaster occurrence probability, impact area and defense suggestions in the warning information to generate comprehensive text information containing disaster risk information and emergency response information.

[0147] Next, the emergency response progress description information is integrated with the key content in the warning information. During the integration process, the language is organized according to a certain logical order, so that the comprehensive text information contains both disaster risk information and accurately conveys the emergency response situation.

[0148] For example, the comprehensive text information may be expressed as: "After monitoring, there is an 80% probability of a mudslide disaster in this area within the next 3 hours, and the predicted impact area includes multiple villages in XX Town. Residents living in the area are advised to immediately stop outdoor activities and seek shelter in nearby safe places. At present, rescue supplies are fully prepared, and a large number of tents, food and medical supplies have been allocated. Rescue teams are on their way, with one team already arriving near XX Village and expected to arrive at the most severely affected area in 10 minutes, and another team expected to arrive in 20 minutes. Please remain calm, follow instructions and actively cooperate with rescue efforts."

[0149] Through the above integration process, the comprehensive text information generated can enable residents to fully understand the risk level, impact area and progress of emergency response of the disaster, providing more comprehensive guidance for them to cope with the disaster.

[0150] Step 1453: Spatially superimpose the disaster impact range visualization image with the rescue team location information in the emergency response progress description information to generate a comprehensive visualization image containing rescue force deployment.

[0151] To more intuitively show the relationship between the deployment of rescue forces and the disaster impact range, the system will spatially superimpose the disaster impact range visualization image and the rescue team location information in the emergency response progress description information.

[0152] Based on the disaster impact range visualization image, the location information of the rescue team is added to the image in the form of intuitive graphics or icons. For example, different colored icons represent different types of rescue teams, with red icons representing fire teams and blue icons representing medical rescue teams. The name and estimated arrival time of each team will be marked next to the icon.

[0153] Through this spatial superimposition, the generated comprehensive visualization image can clearly show the distribution of each rescue team within the disaster-affected area, as well as their distance from the affected area and estimated arrival time. Residents can intuitively understand the deployment of rescue forces by viewing this image, gain a clearer understanding of their own safety situation, and better cooperate with rescue efforts.

[0154] Step 1454: Convert the comprehensive text information into natural speech signals and superimpose voice prompt information for the estimated arrival time of the rescue team to generate comprehensive voice information containing the progress of the emergency response.

[0155] To meet the needs of different residents to obtain information, especially to facilitate those who cannot timely view text or image information, the system will convert the comprehensive text information into natural speech signals and superimpose voice prompt information for the estimated arrival time of the rescue team.

[0156] First, use voice synthesis technology to accurately and clearly convert the comprehensive text information into natural speech signals. During the conversion process, the tone, speed, and volume of the voice will be adjusted according to the emotional and tonal requirements of the text content, making the voice more appealing and infectious.

[0157] Then, for the information on the estimated arrival time of the rescue team, generate separate voice prompt information, such as "Please note that the fire team is expected to arrive in 15 minutes, please wait patiently." Superimpose these voice prompt information into the comprehensive voice information to ensure that residents can hear key rescue time information when listening to the voice information.

[0158] The generated comprehensive voice information containing the emergency response progress is spread through community broadcasting systems, mobile devices, etc. so that residents can obtain the latest disaster and rescue information at any time, ensuring comprehensive coverage and effective communication of information.

[0159] Step 1455: Selecting the corresponding comprehensive warning information form according to the optimal warning content type, and generating comprehensive warning information containing the emergency response progress.

[0160] Finally, according to the optimal warning content type determined in the optimized warning strategy, the corresponding form is selected from the generated comprehensive text information, comprehensive visual image and comprehensive voice information to generate the final comprehensive warning information containing the emergency response progress.

[0161] If the optimal warning content type is text warning information, the comprehensive text information is directly used as the comprehensive warning information for secondary distribution; if it is image warning information, the comprehensive visual image is used as the main form, and the necessary text description is used as the comprehensive warning information; if it is voice warning information, the comprehensive voice information is used as the comprehensive warning information for dissemination.

[0162] In this way, it is ensured that the warning information can be communicated to residents in the most suitable form, maximizing the needs of residents to obtain information and improving the effectiveness of the warning and the coordination efficiency of the emergency response.

[0163] As a non-limiting example, after the operation of distributing the warning content according to the optimized warning strategy and the operation of coordinating the emergency response system, the following operations are further included: Obtaining multi-channel feedback data, the feedback data including user reception status data returned by the information distribution channel, resource allocation execution progress data returned by the emergency response system and disaster actual evolution data collected by sensors in the target area; Performing consistency checking on the feedback data to eliminate abnormal feedback data with unmatched timestamps or untrustworthy sources; Based on the feedback data after consistency checking, calculating the warning information coverage completion rate, the emergency response resource on-time rate and the deviation rate of the actual disaster evolution and the predicted trend, and generating a set of warning effect evaluation indexes; Comparing and analyzing the set of warning effect evaluation indexes with the preset effect evaluation threshold to generate a warning effect evaluation report containing a warning effectiveness judgment result.

[0164] After the operation of distributing the warning information and coordinating the emergency response, in order to evaluate the effect of the entire warning process, multi-channel feedback data needs to be obtained.

[0165] It can be understood that the user receiving state data returned by the information delivery channel can reflect whether the early warning information is successfully delivered to the target user. For example, the mobile phone message platform can return the success rate of sending messages, whether the user reads, and other information; the social media platform can provide data such as the number of views and the number of forwards of the information, through which the user's reception of the early warning information can be understood.

[0166] The resource allocation execution progress data returned by the emergency response system shows the actual progress of the emergency response work. For example, whether the rescue materials are delivered to the designated location on time and in the right amount, whether the rescue team arrives at the disaster area according to the plan and carries out the work, etc.

[0167] The disaster actual evolution data collected by the sensors in the target area is a direct reflection of the true situation of the disaster. For example, for meteorological disasters, sensors can monitor changes in wind speed, rainfall, temperature, etc. in real time; for geological disasters, sensors can monitor ground displacement, crack expansion, etc.

[0168] After obtaining these feedback data, consistency verification processing is first performed to ensure the accuracy and reliability of the data, and abnormal feedback data with inconsistent timestamps or questionable sources is excluded. For example, if the data collected by a sensor is significantly inconsistent with other data, or the data returned by an information delivery channel has questionable sources, these data will be excluded.

[0169] Based on the feedback data that has passed consistency verification, the early warning effect evaluation indexes are calculated. The early warning information coverage completion rate is calculated by the ratio of the actual number of users delivered to the target number of users, reflecting the propagation range of the early warning information. The emergency response resource on-time rate is determined according to the comparison of the actual arrival time of rescue materials and teams with the planned arrival time, reflecting the efficiency of emergency response. The deviation rate of disaster actual evolution and predicted trend is calculated by comparing the actual occurrence of the disaster with the previously predicted trend, measuring the accuracy of disaster prediction.

[0170] The evaluation indexes calculated are compared and analyzed with the preset effect evaluation threshold. The preset effect evaluation threshold is set according to historical experience, industry standards, etc., to judge whether the early warning effect meets the expectation. For example, if the early warning information coverage completion rate is above 90%, the emergency response resource on-time rate is above 85%, and the deviation rate of disaster actual evolution and predicted trend is within 10%, the early warning effect is considered good.

[0171] Through comparison and analysis, an early warning effect evaluation report containing the early warning effectiveness judgment result is generated. The report will list the actual values of each evaluation index, the preset threshold, and the final judgment result in detail, providing an important basis for subsequent improvement of the early warning system and emergency response process.

[0172] As a non-limiting example, after the pre-warning content delivery operation is performed according to the optimized pre-warning strategy, and the emergency response system is linked to complete the emergency response coordination operation, the method further includes: extracting the disaster impact area expansion feature in the disaster risk trend feature and the rescue force deployment area information in the resource allocation state information returned by the emergency response system; based on the disaster impact area expansion feature and the rescue force deployment area information, identifying the spatial overlap degree between the current monitoring coverage and the disaster high-risk area and the rescue key area; According to the spatial overlap degree, the scanning range adjustment parameter of the satellite remote sensing monitoring device and the sampling frequency adjustment parameter of the ground observation device are calculated, and the monitoring parameter adjustment instruction is generated; The monitoring parameter adjustment instruction is sent to the corresponding satellite platform remote sensing monitoring device and ground monitoring station meteorological observation device, and the satellite platform remote sensing monitoring device and ground monitoring station meteorological observation device are driven to perform dynamic monitoring operation according to the adjusted scanning range and sampling frequency.

[0173] In this step, first, the disaster impact area expansion feature is accurately extracted from the disaster risk trend feature, and the rescue force deployment area information is obtained from the resource allocation state information returned by the emergency response system. These two types of information are crucial for subsequent analysis and operation.

[0174] For example, in a possible forest fire scenario, the disaster impact area expansion feature describes in detail the direction and range of possible fire spread, and as time goes on, the fire is expected to spread to several surrounding forest areas; while the rescue force deployment area information clearly shows the distribution of fire brigade, rescue material reserve point, etc. in the disaster area and surrounding areas.

[0175] Based on these two types of information, the spatial overlap degree between the current monitoring coverage and the disaster high-risk area and the rescue key area is identified. Through geographic information system (GIS) technology and spatial analysis algorithm, the area covered by the current monitoring device is compared with the disaster high-risk area and the rescue key area in space. For example, by calculating the intersection area of the two areas and the proportion of the area of the disaster high-risk area or the rescue key area, the spatial overlap degree is determined. If the overlap degree between the monitoring coverage and the disaster high-risk area is high, it means that the monitoring of the high-risk area by the current monitoring is more comprehensive; on the contrary, if the overlap degree is low, it means that there may be some high-risk areas that have not been effectively monitored.

[0176] According to the obtained spatial overlap degree, the scanning range adjustment parameters of the satellite remote sensing monitoring equipment and the sampling frequency adjustment parameters of the ground observation equipment are calculated. For the satellite remote sensing monitoring equipment, if it is found that the monitoring coverage range and the disaster high-risk area overlap degree is insufficient, the scanning range may need to be increased. By analyzing the specific value of the spatial overlap degree and the geographical features of the target area and other factors, the scanning angle, range and other parameters that need to be adjusted are calculated. For example, if the current satellite scanning range only covers part of the disaster high-risk area, the scanning angle needs to be increased by several degrees to ensure that the high-risk area can be fully covered.

[0177] For the ground observation equipment, whether the sampling frequency needs to be adjusted is determined according to the spatial overlap degree. If a certain area is a rescue focus area but the current ground observation equipment has a low sampling frequency, it may not be able to capture the environmental changes in that area in time, and the sampling frequency needs to be increased. According to the spatial overlap degree and the importance of the area and other factors, the appropriate sampling frequency adjustment value is calculated, such as adjusting the original sampling once every hour to sampling once every half hour.

[0178] The monitoring parameter adjustment instructions calculated are sent to the corresponding satellite platform remote sensing monitoring equipment and ground monitoring station meteorological observation equipment. After receiving the instructions, the satellite platform adjusts its observation mode according to the new scanning range parameters to ensure that the disaster-affected area can be more comprehensively and accurately monitored. The meteorological observation equipment of the ground monitoring station increases or reduces the number of data collection according to the adjusted sampling frequency to obtain more intensive or more appropriate data, so as to better reflect the real-time changes of the ground environment.

[0179] Through the adjustment of such dynamic monitoring operation, the monitoring system can better adapt to the development of disasters and the needs of emergency response, improve the pertinence and effectiveness of monitoring, and provide more accurate and timely data support for subsequent disaster warning and emergency decision-making.

[0180] As a non-limiting embodiment, after the operation of delivering the warning content according to the optimized warning strategy is performed, and the emergency response system is linked to complete the collaborative operation of emergency response, it further includes: Collecting disaster warning full-process data of this round, the full-process data including a multi-source monitoring data set, disaster risk trend characteristics, an optimized warning strategy, warning content delivery records and emergency response collaboration records; Performing feature extraction processing on the full-process data to extract key knowledge elements of multi-source data correlation mode, disaster trend evolution law, strategy optimization effect and response collaboration efficiency; Performing similarity matching processing on the key knowledge elements and the existing knowledge elements in the historical disaster warning knowledge base to identify knowledge items that need to be updated or supplemented; Based on the knowledge items that need to be updated or supplemented, the disaster risk correlation model parameters in the multi-agent collaborative processing network and the reward function mapping relationship of the reinforcement learning algorithm are iteratively updated to generate an optimized disaster risk correlation model and reward function.

[0181] After completing a round of disaster warning and emergency response collaborative operation, in order to continuously improve the performance and accuracy of the system, the data of the entire process needs to be collected and analyzed.

[0182] First, collect the data of the entire disaster warning process. The multi-source monitoring data set contains various information collected from satellite cloud data units and ground observation data units, which records the environmental characteristics at different levels before and during the disaster. Disaster risk trend characteristics are the results of joint feature analysis of multi-source monitoring data, reflecting the time evolution law and spatial diffusion pattern of disaster occurrence probability. The optimized warning strategy records the best warning content type and delivery time for the current disaster risk. The warning content delivery record details how the warning information is sent to the target audience through various channels, including delivery time, channel, and audience feedback. The emergency response collaboration record covers the resource allocation and action progress of the emergency response system during the entire process.

[0183] Next, feature extraction is performed on these whole-process data. For multi-source data correlation patterns, data analysis techniques are used to uncover the internal relationships between satellite cloud data and ground observation data, such as the correlation between cloud system structure changes and surface temperature, and precipitation intensity. For disaster trend evolution laws, analyze the change trend of disaster occurrence probability in different time periods and the expansion law of the affected area. For strategy optimization effect, evaluate the effectiveness of the optimized warning strategy in actual application, such as whether the warning information is effectively conveyed and whether it reduces the impact of the disaster. Response collaboration efficiency focuses on the coordination between the emergency response system and the warning system, including the timeliness of resource allocation, the matching degree of rescue actions and warning information, etc. Key knowledge elements are extracted from these aspects, which will serve as an important basis for subsequent knowledge updating and model optimization.

[0184] Then, the extracted key knowledge elements are matched with the existing knowledge elements in the historical disaster warning knowledge base. The historical disaster warning knowledge base stores the experience and knowledge from past disaster warning processes. Through similarity matching algorithms, the current key knowledge elements are compared with the existing knowledge items in the knowledge base to find similarities and differences. For example, if it is found that the cloud system structure changes in the current disaster are similar to those in a certain historical disaster, but there are differences in the expansion of disaster impact, further analysis of these differences is needed.

[0185] Based on the results of similarity matching, identify the knowledge items that need to be updated or supplemented. For those new knowledge elements with low similarity to historical knowledge but important value, or parts of historical knowledge that cannot accurately reflect the current situation, determine them as content that needs to be updated or supplemented. For example, if a new multi-source data association mode is found in the current disaster warning, but there is no relevant record in the historical knowledge base, then this new mode needs to be added as a supplementary knowledge item.

[0186] Finally, according to the knowledge items that need to be updated or supplemented, the disaster risk association model parameters in the multi-agent collaborative processing network and the reward function mapping relationship of the reinforcement learning algorithm are iteratively updated. For the disaster risk association model, according to the newly discovered disaster trend evolution law and multi-source data association mode, the parameters in the model are adjusted so that the model can more accurately reflect the disaster risk. For example, if a new cloud system structure is found to have a stronger correlation with precipitation intensity, the correlation parameter between the two factors in the model is adjusted accordingly.

[0187] For the reward function mapping relationship of the reinforcement learning algorithm, the reward value mapping table is updated according to new knowledge in terms of policy optimization effect and response collaboration efficiency. For example, if a new warning strategy achieves better disaster impact reduction effect in actual application, the reward value of this strategy in the reward function mapping relationship is increased so that the reinforcement learning algorithm can choose this strategy more in the future. Through the above iterative update processing, the optimized disaster risk association model and reward function are generated, thereby improving the performance and adaptability of the entire disaster warning system, so that it can provide more accurate warning and more effective coping strategies when facing similar disasters in the future.

[0188] In specific applications, further optimization can also be based on existing meteorological monitoring and machine learning techniques: for satellite cloud image radiation correction, mature atmospheric correction models such as 6S or MODTRAN are used to eliminate observation angle differences. In the anomaly detection section, the threshold range of the statistical model is supplemented by box plots or Z-score standardization combined with historical data distribution; the light flow field analysis can reuse classic algorithms such as Horn-Schunck or Lucas-Kanade to extract displacement vectors.

[0189] In the embodiments of the present application, the probability calculation of the disaster risk association model can refer to interpretable machine learning methods such as Logistic regression or random forest to quantify the spatio-temporal coupling relationship, thereby improving the ambiguity of the probability calculation logic. The reward function mapping relationship of reinforcement learning can achieve policy optimization through public Q-learning or SARSA algorithms, and historical data normalization processing uses conventional Min-Max normalization to ensure dimensional consistency.

[0190] Exemplarily, the collaborative analysis of the multiple agents can realize feature interaction based on an existing federated learning framework, and disclose the weight distribution logic of feature fusion through an attention mechanism. In addition, all physical quantities (such as migration rate and area change rate) in the embodiments of the present application are measured in international standard units (km / h, %), and the observation values of temperature, precipitation and the like in the preprocessing link adopt a unified dimension to ensure the effectiveness of data fusion.

[0191] The embodiments of the present application first acquire a multi-source monitoring data set containing satellite cloud image data units and ground observation data units, which are collected in real time by monitoring nodes distributed at different spatial positions, can comprehensively and timely collect multi-dimensional environmental data, provide rich and accurate data basis for disaster warning, and make the data more complete and timely. Then, the multi-source monitoring data set is jointly analyzed and processed by the multi-agent collaborative processing network to generate disaster risk trend features reflecting the time evolution law and spatial diffusion mode of disaster occurrence probability, which can accurately analyze disaster risk from time and space dimensions, mine complex potential relationships between data, and improve the ability to grasp the development trend of disaster. Then, the disaster risk trend features are processed by calling the reinforcement learning algorithm to generate an optimized warning strategy, which can accurately match the warning content type and delivery time according to different disaster risks, and improve the pertinence and effectiveness of the warning. Finally, the warning content is delivered according to the optimized warning strategy and the emergency response system is linked, which can realize efficient cooperation, ensure that the warning information is timely conveyed, the emergency response is quickly launched, and the loss caused by the disaster is effectively reduced.

[0192] As can be seen, the embodiments of the present application innovatively jointly analyze multi-source monitoring data through a multi-agent collaborative processing network, optimize the warning strategy by using a reinforcement learning algorithm, and realize efficient cooperation of warning and emergency response, effectively making up for the deficiencies of the prior art and improving the comprehensiveness, accuracy of disaster warning and timeliness and effectiveness of emergency response.

[0193] Based on the same inventive concept, the embodiments of the present application also provide a disaster warning information publishing system. Referring to Figure 2 Fig. 1 shows a possible structure of a disaster warning information publishing system according to an embodiment of the present application, Figure 2 In the embodiments of the present application, the disaster warning information publishing system 200 includes a processor 210 and a memory 220. The memory 220 stores a computer program executable by the processor 210, and the processor 210 can execute the steps of the above-mentioned disaster warning information publishing method based on multi-modal product and agent collaboration by executing the instructions stored in the memory 220.

[0194] Based on the same inventive concept, the embodiments of the present application provide a computer readable storage medium comprising a computer program, when the computer program is run on a disaster early warning information publishing system, the computer program is used to make the disaster early warning information publishing system execute the steps of the disaster early warning information publishing method based on the cooperation of multi-modal products and agents described above. In some possible implementation manners, various aspects of the disaster early warning information publishing method based on the cooperation of multi-modal products and agents provided by the present application can also be implemented in the form of a program product, which comprises a computer program, when the program product is run on a disaster early warning information publishing system, the computer program is used to make the disaster early warning information publishing system execute the steps in the disaster early warning information publishing method based on the cooperation of multi-modal products and agents described above, for example, the disaster early warning information publishing system can execute the steps as shown in Figure 1 FIG. 6.

[0195] In the technical solutions related to the above embodiments of the present application, whether it is the comparison calculation of multi-dimensional features or the construction of composite parameters, if there are problems caused by significant differences in the number of dimensions, units of dimensions and semantic meanings of different features, those skilled in the art can fully understand that these differences need to be properly handled based on their professional knowledge and past practical experience, so that the calculation result is accurate and has comparability, and logical confusion, unclear mathematical meaning and other conditions are avoided. In detail, when facing features with different number of dimensions, in order to accurately calculate the similarity, matching degree or feature distance between different features, those skilled in the art can use various strategies, including but not limited to feature selection, feature extraction, kernel function processing. When dealing with the comparison of multi-dimensional features, in order to realize the comparability alignment of the feature space, those skilled in the art can use various existing general technical means, including but not limited to standardization preprocessing, mapping conversion, space projection. In the construction process of composite parameters (such as loss function values), different parameter items often have different dimensions, and those skilled in the art can use normalization processing or adaptive weight distribution mechanism based on distribution characteristics. The above general technical means for solving feature matching and loss balancing problems all belong to the common knowledge in the art. These technical means have been fully verified and widely used in a large number of practical applications, and those skilled in the art can skillfully and flexibly use these methods to deal with similar dimensional difference problems.

Claims

1. A method for disaster warning information release based on multi-modal product and agent cooperation, characterized in that, The method comprises: acquiring a multi-source monitoring data set containing satellite cloud image data units and ground observation data units, the multi-source monitoring data set being generated by monitoring nodes distributed in different spatial positions in real time; performing joint feature analysis processing on the multi-source monitoring data set through a multi-agent collaborative processing network to generate disaster risk trend features, the disaster risk trend features reflecting the time evolution law and spatial diffusion mode of disaster occurrence probability; calling a reinforcement learning algorithm to perform early warning strategy optimization processing on the disaster risk trend features to generate an optimized early warning strategy containing early warning content types and delivery timing; performing early warning content delivery operations according to the optimized early warning strategy and coordinating emergency response systems to complete emergency response operations. 2.The method of claim 1, wherein, The acquisition of the multi-source monitoring data set containing satellite cloud image data units and ground observation data units comprises: starting remote sensing monitoring equipment deployed on a satellite platform to collect satellite cloud image data units covering a target area at preset time intervals, the satellite cloud image data units containing cloud system structure features and water vapor content distribution information; starting meteorological observation equipment deployed at ground monitoring sites to synchronously collect ground observation data units with timestamps aligned with those of the satellite cloud image data units, the ground observation data units containing ground surface temperature distribution features, precipitation intensity distribution features, and wind speed and direction change features; performing radiation correction processing on the satellite cloud image data units to eliminate brightness deviation caused by observation angle differences of remote sensing equipment and generate radiation-corrected satellite cloud image data units; performing outlier detection processing on the ground observation data units to identify and eliminate abnormal observation values outside the statistical range of historical data and generate ground observation data units after outlier elimination; performing data alignment processing on the radiation-corrected satellite cloud image data units and the ground observation data units after outlier elimination according to timestamps to generate a multi-source monitoring data set with time synchronization relationship. 3.The method of claim 1, wherein, The joint feature analysis processing on the multi-source monitoring data set through the multi-agent collaborative processing network to generate disaster risk trend features comprises: allocating satellite cloud image data units in the multi-source monitoring data set to a first agent for cloud system evolution feature extraction processing to obtain cloud system structure change features and water vapor transport path features; allocating ground observation data units in the multi-source monitoring data set to a second agent for ground surface environment feature extraction processing to obtain ground surface temperature change trend features, precipitation intensity accumulation features, and wind speed and direction fluctuation features; performing cross-correlation analysis processing on the cloud system structure change features, water vapor transport path features, ground surface temperature change trend features, precipitation intensity accumulation features, and wind speed and direction fluctuation features through information exchange channels of the multi-agent collaborative processing network to identify spatio-temporal coupling relationships between different features; constructing a disaster risk correlation model based on the spatio-temporal coupling relationships, calculating evolution parameters of disaster occurrence probability in different time windows through the disaster risk correlation model, and generating disaster probability time series features reflecting the time evolution law. According to the disaster probability time sequence characteristics and the spatial diffusion direction parameters in the space-time coupling relationship, a disaster spatial diffusion model is constructed, and an expansion rate parameter of a disaster influence range is calculated through the disaster spatial diffusion model to generate disaster influence area expansion characteristics reflecting a spatial diffusion mode; The disaster probability time sequence characteristics and the disaster influence area expansion characteristics are subjected to feature fusion processing to generate disaster risk trend characteristics.

4. The disaster early warning information publishing method based on multi-modal product and agent cooperation according to claim 3, characterized in that, The satellite cloud image data unit in the multi-source monitoring data set is allocated to a first intelligent agent for cloud system evolution feature extraction processing to obtain cloud system structure change characteristics and water vapor transport path characteristics, including: The satellite cloud image data unit is subjected to image segmentation processing to divide the region boundaries of different cloud system types and generate a cloud system region segmentation result; The cloud system region segmentation result is subjected to morphological feature extraction processing to calculate the area change rate, shape complexity parameter and boundary smoothness parameter of each cloud system region, and generate cloud system structure change characteristics; The satellite cloud image data unit is subjected to optical flow field analysis processing to extract displacement vector distribution information of adjacent time stamp cloud system regions; A water vapor transport path tracking model is constructed based on the displacement vector distribution information, and a migration direction and migration rate parameter of water vapor between different cloud system regions are calculated through the water vapor transport path tracking model to generate water vapor transport path characteristics. 5.The method of claim 3, wherein, The ground observation data unit in the multi-source monitoring data set is allocated to a second intelligent agent for ground surface environment feature extraction processing to obtain ground surface temperature change trend characteristics, precipitation intensity accumulation characteristics and wind speed and direction fluctuation characteristics, including: The ground surface temperature data in the ground observation data unit is subjected to time sequence analysis processing to calculate temperature difference parameters and temperature change slope parameters of consecutive time points, and generate ground surface temperature change trend characteristics; The precipitation intensity data in the ground observation data unit is subjected to accumulation amount calculation processing to calculate total precipitation parameters and precipitation intensity peak parameters within a preset time window, and generate precipitation intensity accumulation characteristics; The wind speed and direction data in the ground observation data unit are subjected to fluctuation analysis processing to calculate wind speed difference parameters and wind direction deflection angle parameters of adjacent time points, and generate wind speed and direction fluctuation characteristics. 6.The method of disaster early warning information release based on multi-modal product and agent cooperation according to claim 1, characterized in that, The disaster risk trend characteristics are subjected to early warning strategy optimization processing by calling a reinforcement learning algorithm to generate an optimized early warning strategy including early warning content types and delivery timing, including: The disaster probability time sequence characteristics and the disaster influence area expansion characteristics in the disaster risk trend characteristics are taken as state inputs of the reinforcement learning algorithm; A combination of early warning content types and delivery timing is set as an action space of the reinforcement learning algorithm, the early warning content types include text early warning information, image early warning information and voice early warning information, and the delivery timing includes immediate delivery, delayed delivery and staged delivery; A reward function is constructed, the reward function is configured based on the disaster influence reduction effects of different action combinations in historical early warning events; Through a strategy gradient optimization process of the reinforcement learning algorithm, a mapping relationship between the state inputs and the action space is established to generate a strategy network maximizing the reward function value; The strategy network is used for action selection processing on a current state input, to generate an optimized early warning strategy including an optimal early warning content type and an optimal delivery time.

7. The method of claim 6, wherein the method further comprises: The construction of the reward function includes: Obtain records of implemented early warning strategies in historical disaster events, the records including historical early warning content types, historical delivery times, and corresponding disaster impact degree data; Normalize the disaster impact degree data to generate a quantitative indicator reflecting the reduction in disaster impact; Establish an association between the historical early warning content types, the historical delivery times, and the quantitative indicator to generate a reward value mapping table; Generate a reward function based on the reward value mapping table, the output value of the reward function being a historical average reward value corresponding to a current action combination. 8.The method of claim 1, wherein, The execution of the early warning content delivery operation according to the optimized early warning strategy and the completion of the emergency response coordination operation by the emergency response system include: Extract the optimal early warning content type in the optimized early warning strategy, call a corresponding content generation module to generate early warning information conforming to the optimal early warning content type, the early warning information including a disaster occurrence probability, an impact area, and a defense suggestion; Extract the optimal delivery time in the optimized early warning strategy, and send the early warning information to a preset information delivery channel according to the optimal delivery time, the information delivery channel including a mobile phone short message platform, a social media platform, and a community broadcasting system; Send coordination request information including disaster impact area expansion characteristics to the emergency response system, the coordination request information being used to instruct the emergency response system to start a resource allocation process; Receive resource allocation state information returned by the emergency response system, the resource allocation state information including a rescue material reserve amount, a rescue team position, and an arrival time prediction; Integrate the resource allocation state information and the early warning information to generate comprehensive early warning information including an emergency response progress, and perform secondary delivery of the comprehensive early warning information through the information delivery channel. 9.The method of claim 8, wherein, The calling of the corresponding content generation module to generate early warning information conforming to the optimal early warning content type includes: When the optimal early warning content type is text early warning information, call a text generation module to generate structured text information based on the disaster occurrence probability, the impact area, and the defense suggestion in the disaster risk trend characteristics; When the optimal early warning content type is image early warning information, call an image generation module to generate a disaster impact range visualization image based on the disaster impact area expansion characteristics in the disaster risk trend characteristics; When the optimal early warning content type is voice early warning information, call a speech synthesis module to convert the structured text information into natural speech signals to generate voice early warning information. The integration of the resource allocation state information and the early warning information to generate comprehensive early warning information including an emergency response progress includes: Extract the rescue material reserve amount, the rescue team position, and the arrival time prediction in the resource allocation state information to generate emergency response progress description information; The emergency response progress description information is content-fusion processed with the disaster occurrence probability, influence area and defense suggestion in the early warning information to generate comprehensive text information containing disaster risk information and emergency response information; The disaster influence range visual image is spatially superimposed processed with the rescue team position information in the emergency response progress description information to generate a comprehensive visual image containing rescue force deployment; The comprehensive text information is converted into natural speech signals, and the speech prompt information of the rescue team arrival time prediction is superimposed to generate comprehensive speech information containing emergency response progress; According to the optimal early warning content type selection, corresponding comprehensive early warning information forms are selected to generate comprehensive early warning information containing emergency response progress.

10. A disaster early warning information distribution system characterized by comprising: The device comprises a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the method in any one of claims 1-9.

Images