Full-airspace sudden fire aerosol source analysis algorithm using deep learning

Abstract

The invention belongs to the technical field of environmental monitoring and artificial intelligence, and discloses a full-airspace sudden fire aerosol source analysis algorithm using deep learning. According to the invention, the real-time data of a stationary orbit satellite, a long-endurance unmanned aerial vehicle and a mobile monitoring vehicle are integrated by constructing a'sky-air-ground 'whole-airspace integrated emergency monitoring network; a dynamically updated deep learning-PMF hybrid model is designed, a space-time diagram neural network and a variational auto-encoder are used to analyze fire pollution features, and an analysis result is used as a constraint condition to optimize a PMF model to realize accurate analysis of a fire pollution source; a full-process quick response mechanism is established, and automatic processing from fire monitoring to source tracing analysis is achieved; meanwhile, model parameters are continuously optimized through an online platform, and the tracing accuracy of the sudden fire disaster is improved. According to the algorithm, through combination of deep learning and a PMF algorithm, massive complex data in the whole airspace is effectively processed, and accurate and dynamic analysis of an emergent fire aerosol source is realized.

Description

Technical Field

[0001] The present invention belongs to the field of environmental monitoring and artificial intelligence technology, and relates to an algorithm for analyzing the aerosol sources of sudden fires in the entire airspace using deep learning. Background Art

[0002] Sudden fires (such as forest fires and industrial fires) are unpredictable and develop rapidly. The large amount of aerosol particles they release can have a serious impact on regional air quality and public health in a short period of time. Therefore, achieving accurate tracing and quantitative analysis of fire aerosols is of great significance to pollution prevention and control and disaster assessment. Traditional aerosol source analysis technology mainly targets conventional pollution sources and has obvious shortcomings when responding to sudden fires: First, the monitoring system has temporal and spatial limitations. The commonly used PMF (orthogonal matrix decomposition) method mainly relies on ground-based fixed monitoring data, which makes 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 changeable. Conventional source spectral libraries cannot cover its dynamic changing characteristics, which can easily lead to misjudgment of source contributions. (Ulevicius V, Bycenkiene S, Bozzetti C, et al. Fossiland non-fossil source contributions to atmospheric carbonaceous aerosols during extreme spring grassland fires in Eastern Europe[J]. AtmosphericChemistry and Physics, 2016, 16(9): 5513-5529.)

[0003] In response to the above-mentioned technical limitations, the present invention proposes an innovative improvement plan: by constructing an integrated, full-airspace, three-dimensional monitoring network, the temporal and spatial limitations of traditional single monitoring methods are overcome; further adopting a hybrid modeling method that integrates deep learning and PMF to effectively address the adaptability of traditional methods to dynamic emission characteristics; and at the same time, establishing an intelligent analysis system that includes an online learning mechanism to automatically update model parameters and source spectrum libraries by continuously absorbing new observation data, ensuring the ability to continuously and accurately analyze sudden fires. This series of technological innovations enables the present invention to achieve full-process optimization from minute-level response to precise source tracing, providing reliable technical support for emergency decision-making on fire environmental pollution. Summary of the Invention

[0004] The purpose of the present invention is to provide a full-airspace sudden fire aerosol source parsing algorithm using deep learning.

[0005] The technical solution of the present invention:

[0006] A deep learning-based algorithm for all-airspace fire aerosol source resolution. The steps are as follows:

[0007] (1) Build an integrated emergency monitoring network covering the entire airspace, covering “sky, air, and ground”;

[0008] Space-based: Obtain 10-minute fire point brightness temperature data from geostationary orbit satellites (such as FY-4A). 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.

[0009] Airborne: Deploy a swarm of long-endurance drones. After an early warning is triggered, the swarm arrives at the fire area to conduct intensified observations, generating a vertical profile of near-ground pollutant concentrations every 10 minutes.

[0010] Ground: dispatch mobile monitoring vehicles to collect ground pollutants, including PM2.5, PM10, black carbon (BC), carbon monoxide (OC), volatile organic compounds (VOCs), nitrogen oxides (NO x The mobile monitoring vehicle is equipped with a gas chromatography-mass spectrometer to determine the content of benzopyrene (BaP) in pollutants; meteorological parameters (wind speed and direction) are monitored simultaneously to provide support for the analysis of pollutant transmission paths;

[0011] Data fusion and three-dimensional pollution field construction: Edge computing equipment is deployed at drone 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 burning layer), 200-500 meters (canopy burning layer), and 500-1000 meters (smoke diffusion layer). Kriging interpolation is used to generate a continuous concentration distribution. The resulting three-dimensional pollution field is continuously generated in time and space, stored in NetCDF format, and supports real-time access.

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

[0013] Construct a spatiotemporal graph neural network: Nodes contain satellite pixels, drone observation points, and ground monitoring data. Each node contains nine feature dimensions: pollutant concentration, observation time, and longitude and latitude coordinates. Edge weights are dynamically calculated based on real-time wind field conditions, and a probability matrix of pollutant transmission paths is output.

[0014] Training the variational autoencoder: Real-time observation data from geostationary satellites, long-endurance drone swarms, and mobile monitoring vehicles was used as input, including fire point temperature, vertical profile concentration, and pollutant concentration. The variational autoencoder was trained using the Adam optimizer (learning rate 0.001) with a batch size of 64. The training set consisted of real-time observation data from fire events over the past five years. Training was terminated when the validation set accuracy reached 92%. The variational autoencoder outputs the dynamic source spectrum characteristics and their uncertainties of the fire, including fire characteristic fingerprints, contribution weights of each altitude layer, and time-series curves of source intensity changes.

[0015] The probability matrix of the pollutant transmission path output by the spatiotemporal graph neural network is used as a spatial constraint, and the dynamic source spectrum characteristics of the fire generated by the variational autoencoder are used as component constraints to input into the PMF model, and the solution is solved by the alternating least squares method. The preset number of factors of the PMF model ranges from 5 to 12. The PMF model with different preset numbers of factors is run, and the preset factor contribution matrix and preset factor load matrix output each time are recorded. The variance explanation rate of the factor load matrix is ​​calculated based on the preset factor contribution matrix and the preset factor load matrix, and the minimum preset number of factors that meets the factor load matrix variance explanation rate ≥ 85% threshold is selected. The number of iterations is set to 200, and the results are output when the convergence threshold of 1e-6 is reached, including the spatial contribution distribution and component contribution ratio of each pollutant.

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

[0017] Set thresholds for multiple warning levels and dispatch monitoring equipment according to the warning level: Level 1 warning (routine monitoring), when the satellite thermal anomaly intensity is 330K ≤ LST < 340K, triggers one drone and one mobile monitoring vehicle to conduct basic observations along the preset route; Level 2 warning (enhanced monitoring), when 340K ≤ LST < 350K, triggers three drones and two mobile monitoring vehicles; Level 3 warning (emergency monitoring), when LST ≥ 350K, triggers six or more drones and three or more mobile monitoring vehicles;

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

[0019] Build an automated processing pipeline to achieve unmanned operation from data collection in step (1), the probability matrix of pollutant transmission paths output by the spatiotemporal graph neural network in step (2), and the dynamic source spectrum characteristics of the fire generated by the variational autoencoder to the source of the fire pollution, and update the traceability results every 10 minutes;

[0020] Use edge computing devices to aggregate and standardize traceability results in real time, generating a real-time traceability report that includes a pollution contribution matrix and confidence intervals 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 the complete observation data and traceability results of each emergency, including multi-platform monitoring raw data, 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, focusing on optimizing the ability to identify new types of combustion objects;

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

[0025] Beneficial effects of the present invention: The present invention constructs an integrated emergency monitoring network for the entire airspace of "sky-air-ground", integrating real-time data from geostationary satellites, long-endurance drones, and fast-response mobile monitoring vehicles; designs a dynamically updated deep learning-PMF hybrid model, utilizes deep learning algorithms such as spatiotemporal graph neural networks and variational autoencoders to analyze fire pollution characteristics, uses the analysis results as constraints, and optimizes the PMF model to achieve accurate analysis of fire pollution sources; further establishes a full-process rapid response mechanism to achieve automated processing from fire monitoring to source tracing and analysis; at the same time, continuously optimizes model parameters through an online platform to improve the accuracy of source tracing of sudden fires. Through the organic combination of deep learning and PMF algorithms, the algorithm can effectively process massive and complex data in the entire airspace and achieve accurate and dynamic analysis of the aerosol sources of sudden fires. DETAILED DESCRIPTION

[0026] The specific implementation of the present invention is further described below in conjunction with the technical solution.

[0027] It is understandable that traditional aerosol source analysis technology mainly targets conventional pollution sources and has obvious shortcomings when responding to sudden fires: first, the monitoring system has temporal and spatial limitations, and the commonly used PMF (orthogonal matrix decomposition) method mainly relies on ground-based fixed monitoring data, which makes 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 changeable, and conventional source spectrum libraries cannot cover their dynamic change characteristics, which can easily lead to misjudgment of source contributions.

[0028] To address the above-mentioned issues, the embodiments of the present invention construct an integrated "sky-air-ground" full-airspace emergency monitoring network, integrating real-time data from geostationary satellites, long-endurance drones, 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 to achieve accurate analysis of fire pollution sources. A full-process rapid response mechanism is further established to achieve automated processing from fire monitoring to source tracing and analysis. Simultaneously, model parameters are continuously optimized through an online platform to improve the accuracy of tracing the source of sudden fires. Through the organic combination of deep learning and PMF algorithms, this algorithm can effectively process massive amounts of complex data across the entire airspace, achieving accurate and dynamic analysis of the aerosol sources of sudden fires.

[0029] Specifically, in a preferred embodiment provided by the present invention, a deep learning-based full-airspace fire aerosol source parsing algorithm is provided, the method comprising the following steps:

[0030] Step S101: Build an integrated emergency monitoring network for the entire airspace of "sky, air, and ground".

[0031] In the embodiment of the present invention, 10-minute fire point temperature data from geostationary satellites (such as FY-4A) are obtained. 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 synchronously. A long-endurance drone group is configured. After the early warning is triggered, the long-endurance drone group arrives at the fire area to conduct encrypted observations, generating a vertical profile concentration of near-ground pollutants every 10 minutes. Mobile monitoring vehicles are dispatched to collect ground pollutants, including PM2.5, PM10, black carbon (BC), carbon monoxide (OC), volatile organic compounds (VOCs), and other pollutants. The system monitors volatile organic compounds (VOCs), nitrogen oxides (NOx) and benzopyrene (BaP), and simultaneously monitors meteorological parameters (wind speed and direction); deploys edge computing equipment at the UAV ground station and the mobile monitoring vehicle terminal to unify multi-source data to a horizontal spatial resolution of 500 meters and a 10-minute interval; vertically, it is divided into three layers: 0-200 meters (ground burning layer), 200-500 meters (canopy burning layer), and 500-1000 meters (smoke diffusion layer); and generates a continuous concentration distribution through Kriging interpolation, ultimately forming a continuous three-dimensional pollution field in time and space.

[0032] Among them, in the preferred embodiment provided by the present invention, the construction of the "sky-air-ground" full airspace integrated emergency monitoring network specifically includes the following steps:

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

[0034] Step S1012: Configure a swarm of long-flight drones. After the warning is triggered, the swarm arrives at the fire area to conduct intensified observations, generating a vertical profile of near-ground pollutant concentrations every 10 minutes.

[0035] Step S1013: dispatch a mobile monitoring vehicle to collect ground pollutants, including PM2.5, PM10, black carbon (BC), carbon monoxide (OC), volatile organic compounds (VOCs), nitrogen oxides (NO x ); The mobile monitoring vehicle is equipped with a gas chromatography-mass spectrometry instrument to determine the content of benzopyrene (BaP) in pollutants; and simultaneously monitors meteorological parameters (wind speed and wind direction) to provide support for the analysis of pollutant transmission paths.

[0036] In step S1014, edge computing equipment is deployed at the drone ground station 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 burning layer), 200-500 meters (canopy burning layer), and 500-1000 meters (smoke diffusion layer). Kriging interpolation is used to generate a continuous concentration distribution. Ultimately, a spatiotemporally continuous three-dimensional pollution field is formed, stored in NetCDF format, and supports real-time access.

[0037] Furthermore, the aerosol source parsing algorithm for sudden fires in the entire airspace using deep learning also includes the following steps:

[0038] In step S102, a dynamically updated deep learning-PMF hybrid model is designed. The pollutant transmission path in the fire area is analyzed through the spatiotemporal graph neural network, and the dynamic source spectrum characteristics of the fire are extracted by the variational autoencoder. The probability matrix of the pollutant transmission path output by the spatiotemporal graph neural network and the dynamic source spectrum characteristics of the fire generated by the variational autoencoder are used as constraints to constrain the PMF model and achieve accurate analysis of the fire pollution source.

[0039] In an embodiment of the present invention, a spatiotemporal graph neural network is constructed, in which nodes include satellite pixels, drone observation points and ground monitoring data. Each node includes 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 probability matrix of the pollutant transmission path is output. At the same time, a variational autoencoder is trained, and real-time observation data from geostationary orbit satellites, long-endurance drone swarms and mobile monitoring vehicles are input, including fire point brightness temperature, vertical profile concentration and pollutant concentration. The Adam optimizer (learning rate 0.001) is used to train the variational autoencoder with a batch size of 64. The training set includes real-time observation data of fire events in the past five years. Training is stopped when the accuracy of the validation set reaches 92%. The variational autoencoder outputs the dynamic source spectrum characteristics of the fire and its non- Determinism, including fire characteristic fingerprints, contribution weights of each height layer, and time-series change curve of source strength; further, the probability matrix of the pollutant transmission path output by the spatiotemporal graph neural network is used as the spatial constraint and the dynamic source spectrum characteristics of the fire generated by the variational autoencoder are used as the component constraint to input into the PMF model, and the alternating least squares method is used to solve it; the preset number of factors ranges from 5 to 12, and the PMF model with different number of factors is run, and the factor contribution matrix and factor load matrix output each time are recorded. The variance explanation rate of the factor load matrix is ​​calculated based on these two matrices, and the minimum number of factors that meets the factor load matrix variance explanation rate ≥ 85% threshold is selected; the number of iterations is set to 200 times, 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] Among them, in the preferred embodiment provided by the present invention, the design of the dynamically updated deep learning-PMF hybrid model analyzes the pollutant transmission path in the fire area through the spatiotemporal graph neural network and extracts the dynamic source spectrum characteristics of the fire through the variational autoencoder. The probability matrix of the pollutant transmission path output by the spatiotemporal graph neural network and the dynamic source spectrum characteristics of the fire generated by the variational autoencoder are used as constraints to constrain the PMF model and achieve accurate analysis of the fire pollution source. Specifically, the following steps are included:

[0041] Step S1021: Construct a spatiotemporal graph neural network. The nodes contain satellite pixels, drone observation points, and ground monitoring data. Each node contains nine feature dimensions: pollutant concentration, observation time, and longitude and latitude three-dimensional coordinates. The edge weights are dynamically calculated based on the real-time wind field, and the probability matrix of the pollutant transmission path is output.

[0042] Step S1022: Train the variational autoencoder. Input real-time observation data from geostationary satellites, long-endurance drone swarms, and mobile monitoring vehicles, including fire point temperature, vertical profile concentration, and pollutant concentration, is used. The variational autoencoder is trained 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 five years. Training is terminated when the validation set accuracy reaches 92%. The variational autoencoder outputs the dynamic source spectrum characteristics and their uncertainties of the fire, including the fire characteristic fingerprint, the contribution weight of each altitude layer, and the source intensity time series change curve.

[0043] In step S1023, the probability matrix of the pollutant transmission path output by the spatiotemporal graph neural network is used as a spatial constraint, and the dynamic source spectrum characteristics of the fire generated by the variational autoencoder are used as component constraints to input into the PMF model, and the solution is obtained by alternating least squares method; the preset number of factors ranges from 5 to 12, and the PMF model with different numbers of factors is run, and the factor contribution matrix and factor load matrix output each time are recorded. The variance explanation rate of the factor load matrix is ​​calculated based on these two matrices, and the minimum number of factors that meets the factor load matrix variance explanation rate ≥ 85% threshold is selected; the number of iterations is set to 200 times, and the results are output when the convergence threshold of 1e-6 is reached, including the spatial contribution distribution and component contribution ratio of each pollution source.

[0044] Furthermore, the aerosol source parsing algorithm for sudden fires in the entire airspace using deep learning also includes the following steps:

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

[0046] In an embodiment of the present invention, a multi-level warning threshold is set to automatically trigger different response levels according to 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. An automated processing pipeline is further constructed to achieve unmanned operation of the entire process from data collection, feature extraction to source analysis, and the analysis results are updated every 10 minutes. After the response, a real-time traceability report is generated, including the pollution contribution rate matrix and confidence interval of each fire point, as well as emergency prevention and control priority recommendations.

[0047] Among them, in the preferred embodiment provided by the present invention, the establishment of a full-process rapid response mechanism to achieve automated processing from fire monitoring to source tracing and analysis specifically includes the following steps:

[0048] Step S1031 sets the thresholds for multiple warning levels and dispatches monitoring equipment according to the warning levels: Level 1 warning (routine monitoring), when the satellite thermal anomaly intensity is 330K ≤ LST < 340K, triggers one UAV and one mobile monitoring vehicle to conduct basic observations along the preset route; Level 2 warning (enhanced monitoring), when 340K ≤ LST < 350K, triggers three UAVs and two mobile monitoring vehicles; Level 3 warning (emergency monitoring), when LST ≥ 350K, triggers six or more UAVs and three or more mobile monitoring vehicles.

[0049] Step S1032: Develop a resource scheduling algorithm to optimize the deployment paths of drones and mobile monitoring vehicles. This algorithm uses an improved Dijkstra algorithm, integrating road traffic conditions (real-time traffic data), drone endurance (return home 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 of ≤ 3 minutes.

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

[0051] In step S1034, edge computing devices are used to summarize the traceability results in real time and standardize the format to generate a real-time traceability report, including the pollution contribution rate matrix and confidence interval of each fire point and emergency prevention and control priority recommendations.

[0052] Furthermore, the aerosol source parsing algorithm for sudden fires in the entire airspace using deep learning also includes the following steps:

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

[0054] In an embodiment of the present invention, it is necessary to establish a fire case library to record the complete observation data and tracing results of each emergency, including: multi-platform monitoring raw data, dynamic source spectrum characteristics, 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 types of combustibles; develop a performance evaluation module to regularly test the model's ability to retrospectively analyze historical emergencies and verify key indicators: source contribution rate error, impact range prediction accuracy, fire feature fingerprint matching, etc.

[0055] In a preferred embodiment of the present invention, the continuous optimization of 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 to record the complete observation data and traceability results of each emergency, including: multi-platform monitoring raw data, 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, focusing on optimizing the ability to identify new combustion products.

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

Claims

1. A deep learning-based aerosol source analysis algorithm for sudden fires in the entire airspace, characterized by: Here are the steps: (1) Build an integrated emergency monitoring network covering the entire airspace, including “sky, air, and ground”; (2) Design a dynamically updated deep learning-PMF hybrid model, analyze the pollutant transmission path in the fire area through the spatiotemporal graph neural network, and extract the dynamic source spectrum characteristics of the fire through the variational autoencoder. The probability matrix of the pollutant transmission path output by the spatiotemporal graph neural network and the dynamic source spectrum characteristics of the fire generated by the variational autoencoder are used as constraints to constrain the PMF model and achieve accurate analysis of the fire pollution source; (3) Establish a full-process rapid response mechanism 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 deep learning-based full-airspace fire aerosol source analysis algorithm according to claim 1 is characterized in that: The specific implementation process of step (1) is as follows: Space-based: Obtain 10-minute fire point temperature data from geostationary satellites. When the thermal anomaly intensity exceeds 300K, an early warning is triggered and the latitude and longitude coordinates of the fire point are output simultaneously. Airborne: Deploy a swarm of long-endurance drones. After an early warning is triggered, the swarm arrives at the fire area to conduct intensified observations, generating a vertical profile of near-ground pollutant concentrations every 10 minutes. Groundwork: Mobile monitoring vehicles are dispatched to collect ground pollutants, including PM2.5, PM10, black carbon, carbon monoxide, volatile organic compounds, and nitrogen oxides. Gas chromatography-mass spectrometry is installed on the mobile monitoring vehicles to determine the content of benzopyrene in the pollutants. Meteorological parameters such as wind speed and direction are also monitored simultaneously. Data fusion and three-dimensional pollution field construction: Edge computing equipment is deployed at the drone ground station and mobile monitoring vehicle terminals to unify multi-source data to a horizontal spatial resolution of 500 meters and a 10-minute interval; It is divided into three layers in the vertical direction: the 0-200-meter ground burning layer, the 200-500-meter canopy burning layer, and the 500-1000-meter smoke diffusion layer. A continuous concentration distribution is generated through Kriging interpolation; ultimately, a continuous three-dimensional pollution field is formed in time and space, which is stored in NetCDF format and supports real-time calls.

3. The deep learning-based full-airspace fire aerosol source analysis algorithm according to claim 1 is characterized in that: The specific implementation process of step (2) is as follows: Construct a spatiotemporal graph neural network: Nodes contain satellite pixels, drone observation points, and ground monitoring data. Each node contains nine feature dimensions: pollutant concentration, observation time, and longitude and latitude coordinates. Edge weights are dynamically calculated based on real-time wind field conditions, and a probability matrix of pollutant transmission paths is output. Training the variational autoencoder: Real-time observation data from geostationary satellites, long-endurance drone swarms, and mobile monitoring vehicles is input, including fire point temperature, vertical profile concentration, and pollutant concentration. The variational autoencoder is trained using the Adam optimizer with a batch size of 64. The training set contains real-time observation data of fire events from the past five years. Training is terminated when the validation set accuracy reaches 92%. The variational autoencoder outputs the dynamic source spectrum characteristics and their uncertainties of the fire, including fire characteristic fingerprints, contribution weights of each altitude layer, and source intensity time series variation curves. The probability matrix of the pollutant transmission path output by the spatiotemporal graph neural network is used as a spatial constraint, and the dynamic source spectrum characteristics of the fire generated by the variational autoencoder are used as component constraints to input into the PMF model, and the solution is obtained using the alternating least squares method. The number of preset factors of the PMF model ranges from 5 to 12. The PMF model is run with different preset factor numbers, and the preset factor contribution matrix and preset factor load matrix output each time are recorded. The variance explanation rate of the factor load matrix is ​​calculated based on the preset factor contribution matrix and the preset factor load matrix, and the minimum preset factor number that meets the factor load matrix variance explanation rate threshold of ≥85% is selected. The number of iterations was set to 200, and the results were output when the convergence threshold 1e-6 was reached, including the spatial contribution distribution and component contribution ratio of each pollutant.

4. The deep learning-based full-airspace fire aerosol source analysis algorithm according to claim 1 is 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, when the satellite thermal anomaly intensity is 330K ≤ LST < 340K, triggers one drone and one mobile monitoring vehicle to conduct basic observations along the preset route; Level 2 warning, when 340K ≤ LST < 350K, triggers three drones and two mobile monitoring vehicles; Level 3 warning, LST ≥ 350K, triggering 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: Using an improved Dijkstra algorithm, we integrate road traffic conditions, drone endurance, and wind direction in the fire zone to generate the optimal deployment path, with a scheduling response time of ≤3 minutes. Build an automated processing pipeline to achieve unmanned operation from data collection in step (1), the probability matrix of pollutant transmission paths output by the spatiotemporal graph neural network in step (2), and the dynamic source spectrum characteristics of the fire generated by the variational autoencoder to the source of the fire pollution, and update the traceability results every 10 minutes; Use edge computing devices to summarize the traceability results in real time and standardize the format to generate a real-time traceability report, including the pollution contribution rate matrix and confidence interval of each fire point, as well as emergency prevention and control priority recommendations.

5. The deep learning-based full-airspace fire aerosol source analysis algorithm according to claim 1 is characterized in that: The specific implementation process of step (4) is as follows: Establish a fire case database to record the complete observation data and traceability results of each emergency, including multi-platform monitoring raw data, 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, focusing on optimizing the ability to identify new types of combustion objects; Develop a performance evaluation module to regularly test the deep learning-PMF hybrid model's ability to retrospectively analyze historical emergencies, and verify key indicators: source contribution rate error, impact range prediction accuracy, and fire feature fingerprint matching.