A Deep Learning-Based Method for Source Analysis of Sudden Fires Across the Entire Airspace

Abstract

This invention belongs to the field of environmental monitoring and artificial intelligence technology, and discloses a method for analyzing the aerosol sources of sudden fires across the entire airspace using deep learning. This invention constructs an integrated "space-air-ground" emergency monitoring network, integrating real-time data from geostationary satellites, long-endurance UAVs, and mobile monitoring vehicles; designs a dynamically updated deep learning-PMF hybrid model, utilizing spatiotemporal graph neural networks and variational autoencoders to analyze fire pollution characteristics, using the analysis results as constraints to optimize the PMF model for accurate analysis of fire pollution sources; establishes a rapid response mechanism throughout the entire process, achieving automated processing from fire monitoring to source tracing analysis; and continuously optimizes model parameters through an online platform to improve the accuracy of tracing the sources of sudden fires. This algorithm, through the combination of deep learning and the PMF algorithm, effectively processes massive and complex data across the entire airspace, achieving accurate and dynamic analysis of the aerosol sources of sudden fires.

Description

Technical Field

[0001] This invention belongs to the field of environmental monitoring and artificial intelligence technology, and relates to a method for analyzing the source of aerosols in sudden fires across the entire airspace using deep learning. Background Technology

[0002] Sudden fires (such as forest wildfires and industrial fires) are characterized by their unpredictability and rapid development. The large amounts of aerosol particles they release can severely impact regional air quality and public health in a short period. Therefore, accurate source tracing and quantitative analysis of fire aerosols are crucial for pollution control and disaster assessment. Traditional aerosol source apportionment technologies mainly target conventional pollution sources and have significant shortcomings in responding to sudden fires: First, monitoring systems have spatiotemporal limitations; commonly used PMF (orthogonal matrix factorization) methods rely primarily on fixed ground-based monitoring data, making it difficult to capture the vertical emission characteristics and diffusion trends of fires in a timely manner. Second, the aerosol composition of sudden fires is complex and variable, and conventional source libraries cannot cover its dynamic changes, easily leading to misjudgments of source contributions. (Ulevicius V, Byčenkienė S, Bozzetti C, et al.Fossil and non-fossil source contributions to atmospheric carbonaceousaerosols during extreme spring grassland fires in Eastern Europe[J].Atmospheric Chemistry and Physics, 2016, 16(9): 5513-5529.)

[0003] To address the aforementioned technical limitations, this invention proposes an innovative improvement scheme: by constructing an integrated air-space-ground three-dimensional monitoring network, it overcomes the spatiotemporal limitations of traditional single monitoring methods; further, it employs a hybrid modeling method that integrates deep learning and PMF (Polarization-Based Function) to effectively solve the adaptability problem of traditional methods to dynamic emission characteristics; simultaneously, it establishes an intelligent analysis system with an online learning mechanism, which automatically updates model parameters and source spectrum libraries by continuously absorbing new observation data, ensuring continuous and accurate analysis capabilities for sudden fires. This series of technological innovations enables this invention to achieve end-to-end optimization from minute-level response to precise source tracing, providing reliable technical support for emergency decision-making regarding fire-related environmental pollution. Summary of the Invention

[0004] The purpose of this invention is to provide a method for analyzing the source of aerosols in sudden fires across the entire airspace using deep learning.

[0005] The technical solution of this invention:

[0006] A method for analyzing the aerosol sources of sudden fires across the entire airspace using deep learning, comprising the following steps:

[0007] (1) Construct an integrated emergency monitoring network covering the entire airspace, from space to ground;

[0008] Space-based: Acquire 10-minute-level fire point brightness temperature data from geostationary orbit satellites (such as FY-4A), and trigger an early warning when the thermal anomaly intensity (LST) exceeds 300K, while simultaneously outputting the latitude and longitude coordinates of the fire point;

[0009] Airborne: Deploy a swarm of long-endurance UAVs; after the warning is triggered, the swarm of long-endurance UAVs will arrive at the fire zone to conduct intensive observations, generating a vertical profile of near-ground pollutant concentrations every 10 minutes;

[0010] Ground-based: Dispatch mobile monitoring vehicles to collect ground pollutant samples, including PM2.5, PM10, black carbon (BC), carbon monoxide (OC), volatile organic compounds (VOCs), and nitrogen oxides (NOx). x The mobile monitoring vehicle is equipped with a gas chromatography-mass spectrometry (GC-MS) instrument to determine the content of benzo[a]pyrene (BaP) in pollutants; it also monitors meteorological parameters (wind speed, wind direction) simultaneously to provide support for the analysis of pollutant transport paths.

[0011] Data fusion and 3D pollution field construction: Edge computing devices are deployed at UAV ground stations and mobile monitoring vehicle terminals to unify multi-source data to a horizontal spatial resolution of 500 meters and a 10-minute interval; vertically, the data is divided into three layers: 0-200 meters (ground combustion layer), 200-500 meters (canopy combustion layer), and 500-1000 meters (smoke diffusion layer), and continuous concentration distribution is generated through Kriging interpolation; finally, a spatiotemporally continuous 3D pollution field is formed, stored in NetCDF format, and supports real-time retrieval.

[0012] (2) Design a dynamically updated deep learning-PMF hybrid model. The spatiotemporal graph neural network is used to analyze the pollutant transmission path in the fire zone and the variational autoencoder is used to extract the dynamic source spectrum features of the fire. The probability matrix of the pollutant transmission path output by the spatiotemporal graph neural network and the dynamic source spectrum features of the fire generated by the variational autoencoder are used as constraints to constrain the PMF model and achieve accurate analysis of fire pollution sources.

[0013] Construct a spatiotemporal graph neural network: Nodes include satellite pixels, UAV observation points and ground monitoring data. Each node contains nine feature dimensions, including pollutant concentration, observation time and three-dimensional coordinates of latitude and longitude. The edge weights are dynamically calculated by the real-time wind field, and the output is a probability matrix of pollutant transport paths.

[0014] Training the variational autoencoder: Input real-time observation data from geostationary satellites, long-endurance UAV swarms, and mobile monitoring vehicles, including fire point brightness temperature, vertical profile concentration, and pollutant concentration; train the variational autoencoder using the Adam optimizer (learning rate 0.001), with a batch size of 64. The training set contains real-time observation data of fire events from the past 5 years. Training stops when the accuracy of the validation set reaches 92%; the variational autoencoder outputs the dynamic source spectrum characteristics of the fire and its uncertainties, including fire feature fingerprints, contribution weights at each altitude layer, and temporal variation curves of source intensity.

[0015] The probability matrix of pollutant transport paths output by the spatiotemporal graph neural network is used as a spatial constraint, and the dynamic source spectrum features of the fire generated by the variational autoencoder are used as a component constraint input to the PMF model. The alternating least squares method is used to solve the problem. The number of preset factors in the PMF model ranges from 5 to 12. The PMF model with different preset factor numbers is run, and the preset factor contribution matrix and preset factor loading matrix are recorded for each output. The variance explained rate of the factor loading matrix is ​​calculated based on the preset factor contribution matrix and preset factor loading matrix. The minimum number of preset factors that meets the threshold of ≥85% variance explained rate of the factor loading matrix is ​​selected. The number of iterations is set to 200. When the convergence threshold 1e-6 is reached, the results are output, including the spatial contribution distribution and component contribution ratio of each pollutant.

[0016] (3) Establish a rapid response mechanism for the entire process to achieve automated processing from fire monitoring to source tracing and analysis;

[0017] Set thresholds for multiple early warning levels and dispatch monitoring equipment according to the warning level: Level 1 warning (routine monitoring), satellite thermal anomaly intensity 330K≤LST<340K, triggering 1 UAV and 1 mobile monitoring vehicle to conduct basic observations along a preset route; Level 2 warning (enhanced monitoring), 340K≤LST<350K, triggering 3 UAVs and 2 mobile monitoring vehicles; Level 3 warning (emergency monitoring), LST≥350K, triggering 6 or more UAVs and 3 or more mobile monitoring vehicles.

[0018] Develop a resource scheduling algorithm to optimize the deployment paths of drones and mobile monitoring vehicles: adopt an improved Dijkstra algorithm, integrate road traffic status (real-time traffic data), drone endurance (forced return when remaining battery ≥ 30%), and wind direction in the fire zone (avoiding downwind danger zones) to generate the optimal deployment path, with a scheduling response time ≤ 3 minutes;

[0019] An automated processing pipeline is constructed to realize the entire process of unattended operation from data acquisition in step (1), the probability matrix of pollutant transmission path output by the spatiotemporal graph neural network in step (2) and the dynamic source spectrum features of the fire generated by the variational autoencoder to the fire pollution source, and the source tracing results are updated every 10 minutes.

[0020] Edge computing devices are used to summarize the source tracing results in real time and standardize the format to generate a real-time source tracing report, including a pollution contribution rate matrix and confidence interval for each fire point, as well as emergency prevention and control priority recommendations.

[0021] (4) Continuously optimize the parameters of the deep learning-PMF hybrid model through the online platform to improve the accuracy of tracing the source of sudden fires;

[0022] Establish a fire case database to record complete observation data and source tracing results for each emergency, including raw data from multi-platform monitoring, dynamic source spectrum characteristics, and final contribution rate matrix;

[0023] Design an incremental learning algorithm to continuously update the parameters of the deep learning-PMF hybrid model using new fire case data, with a focus on optimizing the ability to identify new types of combustibles;

[0024] Develop a performance evaluation module to regularly test the ability of the deep learning-PMF hybrid model to backtrack and analyze historical emergencies, and verify key indicators: source contribution rate error, accuracy of impact range prediction, and fire feature fingerprint matching degree.

[0025] The beneficial effects of this invention are as follows: This invention constructs an integrated emergency monitoring network covering the entire airspace (space, air, and ground), integrating real-time data from geostationary satellites, long-endurance UAVs, and rapid-response mobile monitoring vehicles; it designs a dynamically updated deep learning-PMF hybrid model, utilizing deep learning algorithms such as spatiotemporal graph neural networks and variational autoencoders to analyze fire pollution characteristics, using the analysis results as constraints to optimize the PMF model for accurate analysis of fire pollution sources; it further establishes a full-process rapid response mechanism to automate the process from fire monitoring to source tracing analysis; and it continuously optimizes model parameters through an online platform to improve the accuracy of tracing the source of sudden fires. This algorithm, through the organic combination of deep learning and the PMF algorithm, can effectively process massive and complex data across the entire airspace, achieving accurate and dynamic analysis of the aerosol sources of sudden fires. Detailed Implementation

[0026] The specific embodiments of the present invention will be further described below in conjunction with the technical solution.

[0027] Understandably, traditional aerosol source apportionment technologies are mainly designed for conventional pollution sources and have significant shortcomings in dealing with sudden fires: First, the monitoring system has spatiotemporal limitations, and the commonly used PMF (orthogonal matrix factorization) method mainly relies on ground-based fixed monitoring data, making it difficult to capture the vertical emission characteristics and diffusion trends of fires in a timely manner; Second, the aerosol composition of sudden fires is complex and variable, and conventional source spectral libraries cannot cover its dynamic changes, which can easily lead to misjudgment of source contributions.

[0028] To address the aforementioned issues, this invention constructs an integrated "space-air-ground" emergency monitoring network, consolidating real-time data from geostationary satellites, long-endurance UAVs, and rapid-response mobile monitoring vehicles. A dynamically updated deep learning-PMF hybrid model is designed, utilizing deep learning algorithms such as spatiotemporal graph neural networks and variational autoencoders to analyze fire pollution characteristics. The analysis results are used as constraints to optimize the PMF model, achieving accurate analysis of fire pollution sources. Furthermore, a full-process rapid response mechanism is established to automate the process from fire monitoring to source tracing. Simultaneously, model parameters are continuously optimized through an online platform to improve the accuracy of tracing the source of sudden fires. This algorithm, through the organic combination of deep learning and the PMF algorithm, can effectively process massive and complex data across the entire airspace, achieving accurate and dynamic analysis of the aerosol sources of sudden fires.

[0029] Specifically, in a preferred embodiment of the present invention, a method for analyzing the source of aerosols in sudden fires across the entire airspace using deep learning is provided, the method comprising the following steps:

[0030] Step S101: Construct an integrated emergency monitoring network covering the entire airspace, from space to ground.

[0031] In this embodiment of the invention, 10-minute-level fire point brightness temperature data from geostationary satellites (such as FY-4A) are acquired. When the thermal anomaly intensity (LST) exceeds 300K, an early warning is triggered, and the latitude and longitude coordinates of the fire point are output simultaneously. A long-endurance UAV swarm is configured. After the early warning is triggered, the long-endurance UAV swarm arrives at the fire area to conduct intensive observation, generating a near-ground pollutant vertical profile concentration every 10 minutes. A mobile monitoring vehicle is dispatched to collect ground pollutants, including PM2.5, PM10, black carbon (BC), carbon monoxide (OC), and volatile organic compounds. The system simultaneously monitors meteorological parameters (wind speed and direction) for organic matter (VOCs), nitrogen oxides (NOx), and benzo[a]pyrene (BaP). Edge computing devices are deployed at UAV ground stations and mobile monitoring vehicle terminals to unify multi-source data to a horizontal spatial resolution of 500 meters and a 10-minute interval. Vertically, the system is divided into three layers: 0-200 meters (ground combustion layer), 200-500 meters (canopy combustion layer), and 500-1000 meters (smoke diffusion layer). Continuous concentration distribution is generated through Kriging interpolation, ultimately forming a spatiotemporally continuous three-dimensional pollution field.

[0032] In a preferred embodiment of the present invention, the construction of an integrated emergency monitoring network covering the entire airspace (space, air, and ground) specifically includes the following steps:

[0033] Step S1011: Obtain 10-minute fire point brightness temperature data from geostationary satellites (such as FY-4A). When the thermal anomaly intensity (LST) exceeds 300K, trigger an early warning and simultaneously output the latitude and longitude coordinates of the fire point.

[0034] Step S1012: Configure a long-endurance UAV swarm; after the warning is triggered, the long-endurance UAV swarm arrives at the fire zone to carry out intensive observation, generating a near-ground pollutant vertical profile concentration every 10 minutes.

[0035] Step S1013: Dispatch the mobile monitoring vehicle to collect ground pollutant data, including PM2.5, PM10, black carbon (BC), carbon monoxide (OC), volatile organic compounds (VOCs), and nitrogen oxides (NOx). x The mobile monitoring vehicle is equipped with a gas chromatography-mass spectrometry (GC-MS) instrument to determine the content of benzo[a]pyrene (BaP) in pollutants; it also monitors meteorological parameters (wind speed, wind direction) simultaneously to provide support for the analysis of pollutant transport paths.

[0036] Step S1014: Deploy edge computing devices on the UAV ground station and mobile monitoring vehicle terminal to unify multi-source data to a horizontal spatial resolution of 500 meters and a 10-minute interval; vertically divide the data into three layers: 0-200 meters (ground combustion layer), 200-500 meters (tree canopy combustion layer), and 500-1000 meters (smoke diffusion layer), and generate a continuous concentration distribution through Kriging interpolation; finally, a spatiotemporally continuous three-dimensional pollution field is formed, stored in NetCDF format, and supports real-time access.

[0037] Furthermore, the method for analyzing the aerosol sources of sudden fires across the entire airspace using deep learning also includes the following steps:

[0038] Step S102: Design a dynamically updated deep learning-PMF hybrid model. The model uses a spatiotemporal graph neural network to analyze the pollutant transport path in the fire zone and a variational autoencoder to extract the dynamic source spectrum features of the fire. The probability matrix of the pollutant transport path output by the spatiotemporal graph neural network and the dynamic source spectrum features of the fire generated by the variational autoencoder are used as constraints to constrain the PMF model, thereby achieving accurate analysis of fire pollution sources.

[0039] In this embodiment of the invention, a spatiotemporal graph neural network is constructed. Each node contains satellite pixels, UAV observation points, and ground monitoring data. Each node contains nine feature dimensions: pollutant concentration, observation time, and three-dimensional coordinates (latitude and longitude). Edge weights are dynamically calculated based on real-time wind fields, and the output is a probability matrix of pollutant transport paths. Simultaneously, a variational autoencoder is trained, inputting real-time observation data from geostationary satellites, long-endurance UAV swarms, and mobile monitoring vehicles, including fire point brightness temperature, vertical profile concentration, and pollutant concentration. The variational autoencoder is trained using an Adam optimizer (learning rate 0.001) with a batch size of 64. The training set contains real-time observation data of fire events from the past five years, and training stops when the validation set accuracy reaches 92%. The variational autoencoder outputs the dynamic source spectrum features of the fire and its... The deterministic approach includes fire feature fingerprints, contribution weights at each altitude level, and temporal variation curves of source intensity. Furthermore, the probability matrix of pollutant transport paths output by the spatiotemporal graph neural network is used as a spatial constraint, and the dynamic source spectrum features of the fire generated by the variational autoencoder are used as component constraints input to the PMF model. Alternating least squares is used to solve the model. The preset number of factors is 5-12. PMF models with different numbers of factors are run, and the factor contribution matrix and factor loading matrix output are recorded each time. The variance explanation rate of the factor loading matrix is ​​calculated based on these two matrices, and the minimum number of factors that satisfies the factor loading matrix variance explanation rate ≥ 85% is selected. The number of iterations is set to 200, and the results are output when the convergence threshold 1e-6 is reached, including the spatial contribution distribution and component contribution ratio of each pollution source.

[0040] In a preferred embodiment of the present invention, the design of a dynamically updated deep learning-PMF hybrid model, which analyzes the pollutant transport path in the fire zone through a spatiotemporal graph neural network and extracts the dynamic source spectrum features of the fire through a variational autoencoder, and uses the probability matrix of the pollutant transport path output by the spatiotemporal graph neural network and the dynamic source spectrum features of the fire generated by the variational autoencoder as constraints to constrain the PMF model, thereby achieving accurate analysis of fire pollution sources, specifically includes the following steps:

[0041] Step S1021: Construct a spatiotemporal graph neural network. The nodes include satellite pixels, UAV observation points and ground monitoring data. Each node contains nine feature dimensions, including pollutant concentration, observation time and three-dimensional coordinates of latitude and longitude. The edge weights are dynamically calculated by the real-time wind field. Output the probability matrix of pollutant transport paths.

[0042] Step S1022: Train the variational autoencoder by inputting real-time observation data from geostationary satellites, long-endurance UAV swarms, and mobile monitoring vehicles, including fire point brightness temperature, vertical profile concentration, and pollutant concentration; train the variational autoencoder using the Adam optimizer (learning rate 0.001), with a batch size of 64. The training set contains real-time observation data of fire events over the past 5 years, and training stops when the accuracy of the validation set reaches 92%; the variational autoencoder outputs the dynamic source spectrum characteristics of the fire and its uncertainties, including fire feature fingerprints, contribution weights at each altitude layer, and temporal variation curves of source intensity.

[0043] Step S1023: The probability matrix of pollutant transport paths output by the spatiotemporal graph neural network is used as a spatial constraint, and the dynamic source spectrum features of the fire generated by the variational autoencoder are used as component constraints. The PMF model is then solved using the alternating least squares method. The preset number of factors is 5-12. The PMF model with different numbers of factors is run, and the factor contribution matrix and factor loading matrix output each time are recorded. The variance explanation rate of the factor loading matrix is ​​calculated based on these two matrices, and the minimum number of factors that meets the threshold of variance explanation rate of factor loading matrix ≥ 85% is selected. The number of iterations is set to 200. When the convergence threshold 1e-6 is reached, the results are output, including the spatial contribution distribution and component contribution ratio of each pollution source.

[0044] Furthermore, the method for analyzing the aerosol sources of sudden fires across the entire airspace using deep learning also includes the following steps:

[0045] Step S103: Establish a rapid response mechanism for the entire process to achieve automated processing from fire monitoring to source tracing and analysis.

[0046] In this embodiment of the invention, multi-level early warning thresholds are set, and different response levels are automatically triggered based on the intensity of satellite thermal anomalies. At the same time, a resource scheduling algorithm is developed to optimize the deployment paths of drones and mobile monitoring vehicles. Furthermore, an automated processing pipeline is constructed to achieve unattended operation of the entire process from data collection and feature extraction to source analysis, with the analysis results updated every 10 minutes. After the response, a real-time source tracing report is generated, including the pollution contribution rate matrix and confidence interval of each fire point, as well as emergency prevention and control priority suggestions.

[0047] In a preferred embodiment of the present invention, the establishment of a full-process rapid response mechanism to automate the process from fire monitoring to source tracing and analysis specifically includes the following steps:

[0048] Step S1031: Set thresholds for multiple warning levels and dispatch monitoring equipment according to the warning level: Level 1 warning (routine monitoring): satellite thermal anomaly intensity 330K≤LST<340K, trigger 1 UAV and 1 mobile monitoring vehicle to conduct basic observations along a preset route; Level 2 warning (enhanced monitoring): 340K≤LST<350K, trigger 3 UAVs and 2 mobile monitoring vehicles; Level 3 warning (emergency monitoring): LST≥350K, trigger 6 or more UAVs and 3 or more mobile monitoring vehicles.

[0049] Step S1032: Develop a resource scheduling algorithm to optimize the deployment paths of UAVs and mobile monitoring vehicles: adopt an improved Dijkstra algorithm, integrate road traffic status (real-time traffic data), UAV endurance (forced return if remaining battery ≥ 30%), and wind direction in the fire zone (avoiding downwind danger zones) to generate the optimal deployment path, with a scheduling response time ≤ 3 minutes.

[0050] Step S1033: Construct an automated processing pipeline to achieve unattended operation of the entire process from data collection and feature extraction to fire pollution source, and update the source tracing results every 10 minutes.

[0051] Step S1034: Use edge computing devices to summarize the source tracing results in real time and standardize the format to generate a real-time source tracing report, including the pollution contribution rate matrix and confidence interval of each fire point, as well as emergency prevention and control priority suggestions.

[0052] Furthermore, the method for analyzing the aerosol sources of sudden fires across the entire airspace using deep learning also includes the following steps:

[0053] Step S104: Continuously optimize model parameters through an online platform to improve the accuracy of tracing the source of sudden fires.

[0054] In this embodiment of the invention, it is necessary to establish a fire case database to record complete observation data and source tracing results for each emergency, including: original data from multi-platform monitoring, dynamic source spectrum features, and final contribution rate matrix; design an incremental learning algorithm to continuously update model parameters using new case data, focusing on optimizing the ability to identify new combustibles; and develop a performance evaluation module to periodically test the model's ability to backtrack and analyze historical emergencies, and verify key indicators such as source contribution rate error, accuracy of influence range prediction, and fire feature fingerprint matching degree.

[0055] In a preferred embodiment of the present invention, the step of continuously optimizing model parameters through an online platform to improve the accuracy of tracing the source of sudden fires specifically includes the following steps:

[0056] Step S1041: Establish a fire case database, recording complete observation data and source tracing results for each emergency, including: original data from multi-platform monitoring, dynamic source spectrum characteristics, and final contribution rate matrix.

[0057] Step S1042: Design an incremental learning algorithm to continuously update model parameters using new case data, with a focus on optimizing the ability to identify new types of combustibles.

[0058] Step S1043: Develop a performance evaluation module to periodically test the model's ability to backtrack and analyze historical emergencies, and verify key indicators such as source contribution rate error, accuracy of impact range prediction, and fire feature fingerprint matching degree.

Claims

1. A method for analyzing the aerosol sources of sudden fires across the entire airspace using deep learning, characterized in that, Here are the steps: (1) Construct an integrated emergency monitoring network covering the entire airspace, including space, air, and ground; (2) Design a dynamically updated deep learning-PMF hybrid model. The spatiotemporal graph neural network is used to analyze the pollutant transmission path in the fire zone and the variational autoencoder is used to extract the dynamic source spectrum features of the fire. The probability matrix of the pollutant transmission path output by the spatiotemporal graph neural network and the dynamic source spectrum features of the fire generated by the variational autoencoder are used as constraints to constrain the PMF model and achieve accurate analysis of fire pollution sources. (3) Establish a rapid response mechanism for the entire process to achieve automated processing from fire monitoring to source tracing and analysis; (4) Continuously optimize the parameters of the deep learning-PMF hybrid model through the online platform to improve the accuracy of tracing the source of sudden fires.

2. The method for analyzing the source of aerosols in sudden fires across the entire airspace using deep learning as described in claim 1, characterized in that, The specific implementation process of step (1) is as follows: Space-based: Acquires 10-minute-level fire point brightness temperature data from geostationary satellites; when the thermal anomaly intensity exceeds 300K, triggers an early warning and simultaneously outputs the latitude and longitude coordinates of the fire point; Airborne: Deploy a swarm of long-endurance UAVs; after the warning is triggered, the swarm of long-endurance UAVs will arrive at the fire zone to conduct intensive observations, generating a vertical profile of near-ground pollutant concentrations every 10 minutes; Ground-based: Dispatch mobile monitoring vehicles to collect ground pollutants, including PM2.5, PM10, black carbon, carbon monoxide, volatile organic compounds, and nitrogen oxides; the mobile monitoring vehicles are equipped with gas chromatography-mass spectrometry (GC-MS) instruments to determine the content of benzo[a]pyrene in pollutants; and simultaneously monitor meteorological parameters: wind speed and wind direction. Data fusion and 3D pollution field construction: Edge computing devices are deployed at UAV ground stations and mobile monitoring vehicle terminals to unify multi-source data to a horizontal spatial resolution of 500 meters and a 10-minute interval; Vertically, it consists of three layers: a ground combustion layer (0-200 meters), a canopy combustion layer (200-500 meters), and a smoke diffusion layer (500-1000 meters). A continuous concentration distribution is generated through Kriging interpolation. Ultimately, a spatiotemporally continuous three-dimensional pollution field is formed, stored in NetCDF format, and can be accessed in real time.

3. The method for analyzing the source of aerosols in sudden fires across the entire airspace using deep learning as described in claim 1, characterized in that, The specific implementation process of step (2) is as follows: Construct a spatiotemporal graph neural network: Nodes include satellite pixels, UAV observation points and ground monitoring data. Each node contains nine feature dimensions, including pollutant concentration, observation time and three-dimensional coordinates of latitude and longitude. The edge weights are dynamically calculated by the real-time wind field, and the output is a probability matrix of pollutant transport paths. Training the variational autoencoder: Input real-time observation data from geostationary satellites, long-endurance UAV swarms, and mobile monitoring vehicles, including fire point brightness temperature, vertical profile concentration, and pollutant concentration; train the variational autoencoder using the Adam optimizer with a batch size of 64. The training set contains real-time observation data of fire events from the past 5 years, and training stops when the accuracy of the validation set reaches 92%; output the dynamic source spectrum characteristics of the fire and its uncertainties, including fire feature fingerprints, contribution weights at each altitude level, and temporal variation curves of source intensity; The probability matrix of pollutant transport paths output by the spatiotemporal graph neural network is used as a spatial constraint, and the dynamic source spectrum features of the fire generated by the variational autoencoder are used as a component constraint input to the PMF model. The alternating least squares method is used to solve the problem. The number of preset factors in the PMF model ranges from 5 to 12. The PMF model with different preset factor numbers is run, and the preset factor contribution matrix and preset factor loading matrix are recorded each time. The variance explanation rate of the factor loading matrix is ​​calculated based on the preset factor contribution matrix and preset factor loading matrix. The minimum number of preset factors that meets the threshold of ≥85% variance explanation rate of the factor loading matrix is ​​selected. The number of iterations is set to 200. When the convergence threshold 1e-6 is reached, the output results are given, including the spatial contribution distribution and component contribution ratio of each pollutant.

4. The method for analyzing the source of aerosols in sudden fires across the entire airspace using deep learning as described in claim 1, characterized in that, The specific implementation process of step (3) is as follows: Set thresholds for multiple warning levels and dispatch monitoring equipment according to the warning level: Level 1 warning, satellite thermal anomaly intensity 330K≤LST<340K, triggers 1 UAV and 1 mobile monitoring vehicle to carry out basic observations along a preset route; Level 2 warning, 340K≤LST<350K, triggers 3 UAVs and 2 mobile monitoring vehicles. Level 3 warning, LST≥350K, triggers more than 6 drones and more than 3 mobile monitoring vehicles; Develop resource scheduling algorithms to optimize the deployment paths of drones and mobile monitoring vehicles: adopt an improved Dijkstra algorithm, integrate road traffic conditions, drone endurance, and wind direction in the fire area to generate the optimal deployment path, with a scheduling response time of ≤3 minutes; An automated processing pipeline is constructed to realize the entire process of unattended operation from data acquisition in step (1), the probability matrix of pollutant transmission path output by the spatiotemporal graph neural network in step (2) and the dynamic source spectrum features of the fire generated by the variational autoencoder to the fire pollution source, and the source tracing results are updated every 10 minutes. Edge computing devices are used to aggregate the source tracing results in real time and standardize the format to generate a real-time source tracing report, including a pollution contribution rate matrix and confidence interval for each fire point, as well as emergency prevention and control priority recommendations.

5. The method for analyzing the source of aerosols in sudden fires across the entire airspace using deep learning as described in claim 1, characterized in that, The specific implementation process of step (4) is as follows: Establish a fire case database to record complete observation data and source tracing results for each emergency, including raw data from multi-platform monitoring, dynamic source spectrum characteristics, and final contribution rate matrix; Design an incremental learning algorithm to continuously update the parameters of the deep learning-PMF hybrid model using new fire case data, with a focus on optimizing the ability to identify new types of combustibles; Develop a performance evaluation module to regularly test the ability of the deep learning-PMF hybrid model to backtrack and analyze historical emergencies, and verify key indicators: source contribution rate error, accuracy of impact range prediction, and fire feature fingerprint matching degree.