A multi-source heterogeneous data fusion atmospheric pollution fingerprint database generation method

Abstract

The present application relates to a kind of multi-source heterogeneous data fusion atmospheric pollution fingerprint database generation method, comprising: S1, the atmospheric pollution data monitored by the collection national control site, meteorological data, geographic information and routine patrol record, form atmospheric pollution prevention and control foundation database;S2, feature matrix design;S3, site high value analysis;S4, K-means clustering and fingerprint generation;S5, by re-tuning to improve clustering effect, form pollution fingerprint database V0;S6, using small-scale real-time tracing platform constructs real-time tracing algorithm framework;S7, repeat S5-S6, form pollution fingerprint atlas with regional characteristics;S8, pollution fingerprint atlas visualization and analysis.The present application breaks through the single dimension limit of traditional manual analysis, provides "pollution dictionary" type knowledge base for pollution prevention and control, can quickly match historical fingerprint library, identify pollutant high value point, is beneficial to realize accurate positioning super-concentration discharge or super-frequency discharge main body, and support fine management and control strategy formulation.

Description

Technical Field

[0001] This invention relates to the fields of air pollution prevention and control and big data analysis technology, and in particular to a method for generating an air pollution fingerprint database by fusing multi-source heterogeneous data. Background Technology

[0002] With the acceleration of industrialization and urbanization, industrial waste gas emissions have become a major source of air pollution, threatening ecology and health. Although a pollutant discharge permit system has been implemented, the traditional regulatory model has obvious shortcomings: on the one hand, the number of law enforcement personnel is limited, making it difficult to achieve all-weather monitoring; on the other hand, the coverage of fixed monitoring stations is insufficient and the timeliness is poor, making it impossible to capture dynamic behaviors such as illegal discharge and excessive discharge. This allows some polluters to evade supervision by exceeding total emission limits, exceeding concentration limits, illegally discharging at night, or tampering with data, and even adjusting processes to create complex pollutant components, increasing the difficulty of tracing the source.

[0003] Against this backdrop, the need to construct a big data pollution fingerprint database for air pollution is urgent. This database integrates enterprise production process parameters, characteristic components of exhaust gases (such as heavy metals and organic pollutant isotopes), spatiotemporal geographic information, and meteorological data. It utilizes machine learning to extract the "unique characteristics" of pollutants, enabling the rapid identification of abnormal emission sources and tracing of pollution pathways from massive amounts of data. Compared to traditional methods, its advantage lies in its ability to drive a shift in regulation from "passive response" to "proactive prevention and control" through the fusion and intelligent analysis of multi-source heterogeneous data. This provides technical support for precise pollution control and scientific decision-making, and is of great significance for the continuous improvement of air quality. Summary of the Invention

[0004] To address the aforementioned shortcomings, this invention provides a method for generating an atmospheric pollution fingerprint database by fusing multi-source heterogeneous data, which aims to improve the above-mentioned problems.

[0005] To achieve the above objectives, this invention provides a method for generating an atmospheric pollution fingerprint database by fusing multi-source heterogeneous data, comprising the following steps:

[0006] S1, collect air pollution data, meteorological data, geographic information, and daily inspection records from national monitoring stations to form a basic database for air pollution prevention and control; the air pollution data includes sulfur dioxide (SO2), nitrogen dioxide (NO2), and inhalable particulate matter (PM2.5). 10 Fine particulate matter (PM) 2.5 The meteorological data includes temperature, humidity, air pressure, wind speed, and wind direction; carbon monoxide (CO) and ozone (O3).

[0007] S2, Feature Matrix Design: Construct a feature matrix with single pollution events or single-day pollution processes as rows and core indicators as columns; where the row content includes event definition standards, time attributes, spatial attributes, and event classification labels, and the column content includes pollution intensity, meteorological conditions, temporal dynamics, spatial impact, and pollution source characteristics;

[0008] S3, Site High Value Analysis: By analyzing the total number of high values ​​at a site, or the annual and monthly high value data, the contribution of pollution to the site is analyzed, and the high value is categorized using the triggering conditions to form a high value AI source tracing library. The contribution of pollution to the site is quantified through multi-dimensional data analysis and model calculation to quantify the contribution ratio of different pollution sources to the site's high values. The high value categorization steps include classifying high value events and extracting feature patterns based on the differences in triggering conditions.

[0009] S4, K-means clustering and fingerprint generation: K-means / PCA is embedded to cluster urban monitoring station data, and meteorological data is combined to trace the pollution transmission path. Convolutional neural network (CNN) is used to analyze high-resolution satellite imagery to identify the distribution of pollution sources in industrial areas and to conduct pollution source identification and tracing analysis.

[0010] S5 improves clustering performance through re-tuning, forming the contaminated fingerprint database V0;

[0011] S6, using a small-scale real-time source tracing platform to build a real-time source tracing algorithm framework: the emission inventory is split into hours by dynamic time allocation coefficient, coupled with AERMOD, and the concentration and contribution rate of each pollution source to the spatial grid and site are calculated in real time to form a new pollution fingerprint database V1;

[0012] S7, repeat S5-S6 to form a pollution fingerprint spectrum with regional characteristics: the pollution fingerprint spectrum integrates a closed-loop learning framework of multi-source feedback, adversarial verification and dynamic evolution of spectrum structure.

[0013] S8, Pollution Fingerprint Visualization and Analysis: Displaying air pollution conditions using heat maps, radar charts, spatial distribution maps, and regional characteristic analysis maps.

[0014] Preferably, the event definition criteria include single pollution events and single-day pollution processes. A single pollution event is marked by the continuous exceedance of pollutant concentration, and a single-day pollution process is based on a calendar day. The time attributes include the start / end time of the event, the total duration, and the time of pollution peak occurrence. The spatial attributes include the geographical scope of impact, the coordinates of the core pollution area, and the pollution transmission path. The event classification label includes classification according to pollution intensity.

[0015] Preferably, the pollution intensity indicators include daily average PM2.5 concentration and PM2.5 concentration. 10Peak concentration, 90th percentile of O3-8h, and SO2 / NO2 ratio; meteorological condition indicators include average relative humidity, minimum wind speed, inversion layer height, and diurnal temperature range; temporal dynamic indicators include the number of days of pollution duration, concentration rise rate, and post-peak residence time; spatial impact indicators include the number of affected cities, cross-regional transport contribution rate, and pollution radius; pollution source characteristic indicators include traffic flow index, industrial emission intensity, biomass combustion labeling, and the proportion of secondary inorganic sources.

[0016] Preferably, the multi-dimensional data analysis and model calculation includes the receptor model analytical method, the trajectory reverse simulation method, and the dynamic weight evaluation method.

[0017] Preferably, the step of re-optimizing to improve clustering performance includes:

[0018] S51, Initial Clustering and Feature Evaluation: Initial clustering is performed based on multi-source monitoring data to generate preliminary fingerprint categories. Then, the internal consistency and external discriminative power of each cluster are evaluated.

[0019] S52, Introducing a feedback loop for optimization: including feedback based on source tracing results and adaptive learning based on new data; the feedback based on source tracing results includes: applying the fingerprints obtained from the initial clustering to actual source tracing cases, and identifying fingerprint categories with weak discriminative power by comparing the contribution of the model simulation with the actual situation; the adaptive learning based on new data includes: the system continuously inputs new monitoring data, uses the new data to incrementally learn or fine-tune the existing clustering model, paying special attention to data points falling in the fuzzy area of ​​the existing cluster boundary, triggering the re-division or merging of specific clusters;

[0020] S53, Dynamic Parameters and Feature Reweighting: Automatically adjusts the clustering algorithm or feature weights based on feedback results;

[0021] S54, Iterative convergence and V0 database formation: Repeat steps S52 and S53. When the change in the cluster structure is lower than the preset threshold in several consecutive iterations and the source tracing accuracy on the independent test case set reaches a stable optimal level, the algorithm is determined to have converged. The stable and optimized clustering results obtained at this time are then solidified into the contaminated fingerprint database V0.

[0022] Preferably, the step of constructing a real-time tracing algorithm framework using a small-scale real-time tracing platform includes:

[0023] S61, Dynamic Time Allocation Coefficient and Hourly Emission Inventory Generation: Construct a coefficient self-learning module to continuously compare the difference between emissions estimated based on allocation coefficients and online monitoring data of emission outlets, and use machine learning to dynamically correct and optimize the time allocation coefficient library so that the estimated values ​​approximate the actual emission dynamics.

[0024] S62, Real-time Coupled Calculation with the AERMOD Model: The generated hourly dynamic emission inventory is coupled with the synchronously acquired hourly meteorological field to drive the AERMOD model for real-time simulation; a strategy combining source-receptor contribution matrix pre-calculation and real-time interpolation is adopted; during real-time operation, based on the current actual meteorological conditions, fast interpolation or similarity matching is performed in the pre-calculated response matrix library, and then multiplied by the actual dynamic source strength of the current hour to obtain the concentration contribution of each source to the grid and station;

[0025] S63, forming a real-time source tracing algorithm framework and a new fingerprint database V1: The real-time source tracing algorithm framework runs continuously. Every time new hourly meteorological and operating condition data is received, a complete dynamic inventory generation-diffusion simulation-contribution calculation process is triggered, outputting the real-time contribution concentration and contribution rate of all pollution sources to the receptor point.

[0026] 7. The method for generating an atmospheric pollution fingerprint database according to claim 1, characterized in that: in S7, the closed-loop learning framework includes:

[0027] S71, Constructing the framework and dynamic update mechanism of the regional map: Based on the V1 database, pollution source fingerprints, meteorological scenarios, geospatial grids, and receptor point concentrations are linked to form a multi-dimensional, heterogeneous knowledge graph initial framework; then the system continuously imports new source tracing case results, new enterprise emission data, and new monitoring data, and the new data triggers map updates: confirming and strengthening repeatedly verified source-receptor-meteorological association paths; downgrading or removing contradictory or invalid associations; and updating potential association patterns;

[0028] S72, introduce adversarial verification and uncertainty quantification: for each association and prediction result in the graph, an uncertainty quantification index is attached, which integrates data quality, model simulation error and case verification frequency;

[0029] S73, forming a pollution fingerprint map with regional characteristics: through long-term operation and continuous learning, the macro-level regional pollution formation and transport patterns that transcend individual cases are finally extracted. The pollution fingerprint map with regional characteristics is a pollution climatology knowledge base deeply embedded in regional geography, industrial layout, and meteorological and climatic characteristics.

[0030] Based on the above technical solution, the advantages of the present invention are:

[0031] This invention breaks through the single-dimensional limitations of traditional manual analysis, providing a "pollution dictionary"-like knowledge base for pollution prevention and control. It enables high-value pollution events to be rapidly matched against historical fingerprint databases, with a calculation speed reaching the second level. Compared to manual analysis, it shortens the high-value response time from hours to minutes, enabling real-time early warning. Furthermore, it uncovers the hidden characteristics of high-value pollution processes, improving the accuracy of high-value cause diagnosis. Because this invention can quickly match historical fingerprint databases and identify high-value pollutant locations, it facilitates the precise location of entities exceeding concentration or frequency of emissions, supporting the formulation of refined control strategies. Attached Figure Description

[0032] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0033] Figure 1 A schematic diagram of the architecture for generating an atmospheric pollution fingerprint database through the fusion of multi-source heterogeneous data;

[0034] Figure 2 A step-by-step diagram illustrating the method for generating an atmospheric pollution fingerprint database;

[0035] Figure 3 This is a flowchart of dynamic clustering and typing.

[0036] Figure 4 This is a schematic diagram of a knowledge graph architecture;

[0037] Figure 5 This is a schematic diagram of the multi-source data fusion processing flow.

[0038] Figure 6 Example of modeling a convolutional neural network (CNN);

[0039] Figure 7 This is an example of a source tracing decision engine;

[0040] Figure 8 A step-by-step diagram for re-optimizing and improving clustering results;

[0041] Figure 9 A step diagram illustrating the steps involved in constructing a real-time traceability algorithm framework using a small-scale real-time traceability platform;

[0042] Figure 10 This is a flowchart of a closed-loop learning framework. Detailed Implementation

[0043] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0044] This invention provides a method for generating an atmospheric pollution fingerprint database by fusing multi-source heterogeneous data, such as... Figure 1 , Figure 2 As shown, it includes the following steps:

[0045] S1, air pollution data monitored by national monitoring stations (sulfur dioxide SO2, nitrogen dioxide NO2, inhalable particulate matter PM2.5). 10 Fine particulate matter (PM) 2.5 The database will consist of data on atmospheric pollution prevention and control, including carbon monoxide (CO), ozone (O3), meteorological data (temperature, humidity, air pressure, wind speed, wind direction), geographic information, and daily inspection records provided by the project team.

[0046] S2, Feature Matrix Design: Construct a feature matrix with single pollution events or single-day pollution processes as rows and core indicators as columns. For example, each row of data includes: daily average PM2.5 concentration, ... 10 Indicators include peak value, 90th percentile of O3, average relative humidity, minimum wind speed, number of days of pollution, number of affected cities, traffic flow, and emission intensity of industrial enterprises.

[0047] Specifically, each line represents a single pollution event or a pollution process over a calendar day (24 hours), and must include the following:

[0048] Event definition standards: Single pollution event: The start and end point is marked by the continuous exceedance of pollutant concentration. (Example: PM2.5 ≥75μg / m³ for 12 consecutive hours and lasting ≥24 hours). Single pollution process: The unit is the calendar day, even if the pollution does not last for the whole day.

[0049] Time attributes: event start / end time (accurate to the hour), total duration (hours), and the time when the pollution peak occurs.

[0050] Spatial attributes: Geographical extent of impact (e.g., list of cities or grid number), coordinates of the core polluted area (latitude and longitude), and pollution transport path (e.g., regional transport direction).

[0051] Preferably, it may also include an event classification label: classified by pollution intensity (light, moderate, heavy), with reference to PM2.5 peak concentrations: light: 75–150 μg / m³, moderate: 150–250 μg / m³, heavy: >250 μg / m³.

[0052] Furthermore, the indicators are divided into five categories: pollution intensity, meteorological conditions, temporal dynamics, spatial impact, and pollution source characteristics, as detailed below:

[0053] 1. Pollution intensity indicators:

[0054] <![CDATA[PM2.5 daily average concentration]]> 24-hour average, reflecting the basic pollution level. Real-time data from monitoring stations <![CDATA[PM 10 Peak concentration The highest 1-hour average during the event indicates dust or industrial emissions. Peak records at monitoring stations <![CDATA[O3-8h 90th percentile]]> Assessing photochemical pollution by the 90th percentile of the maximum 8-hour moving average of ozone. Continuous monitoring data calculation <![CDATA[SO2 / NO2 ratio]]> >1.5 may indicate a coal-fired source, <0.8 may indicate a motor vehicle source. Pollutant concentration ratio

[0055] 2. Meteorological conditions indicators

[0056] Average relative humidity >70% promotes secondary particulate matter formation Weather station data Minimum wind speed A sustained diffusion velocity of <1.5 m / s indicates deteriorating diffusion conditions. Weather station minute-level data Inversion layer height At altitudes below 500 meters, pollutant accumulation is aggravated. Meteorological sounding or model inversion Diurnal temperature range A diurnal temperature range greater than 10℃ can easily induce the hygroscopic growth of particulate matter. Weather station data calculation

[0057] 3. Time-based dynamic indicators

[0058] Number of days of pollution Total number of days from start to finish Event timeline statistics Concentration rise rate <![CDATA[average increase of PM2.5 from the baseline to the peak]]> Calculation of slope of monitoring data Post-peak dwell time Peak pollution duration in hours reflects removal capacity. Event Process Analysis

[0059] 4. Spatial Impact Indicators

[0060] Number of cities affected Total number of cities with pollutant concentrations exceeding standards Regional monitoring network statistics Cross-regional transport contribution rate The proportion of external transmission to local pollution (e.g., the proportion of transmission from the Beijing-Tianjin-Hebei region). Model simulation (such as CMAQ) Pollution radius The maximum distance (km) from the core region to where the concentration drops to the threshold. GIS spatial interpolation

[0061] 5. Pollution source characteristic indicators

[0062] Traffic flow index Average daily traffic volume on urban main roads (10,000 vehicles / day) Traffic management department data Industrial emission intensity <![CDATA[Daily average SO2 / NOx emissions (tons) in key industries (such as iron and steel, chemical industry)]]> Enterprise Online Monitoring Platform Biomass burning marker <![CDATA[K⁺ / PM2.5 ratio > 0.05 or sudden increase in black carbon concentration]]> Component monitoring or satellite inversion Secondary inorganic source ratio <![CDATA[Mass fractions of sulfate, nitrate and ammonium in PM2.5]]> Chemical composition analysis

[0063] S3, Site High Value Analysis: By analyzing the total number of high values ​​at a site, or the high value data year by year or month by month, the contribution of pollution to the site is analyzed, and the initiation conditions are used to classify high values ​​and form a high value AI source tracing library.

[0064] The contribution of pollution to the monitoring stations is quantified through multi-dimensional data analysis and model calculation, which quantifies the contribution ratio of different pollution sources to the high values ​​of the monitoring stations. Preferably, the multi-dimensional data analysis and model calculation includes receptor model analysis, trajectory reverse simulation, and dynamic weight evaluation.

[0065] 1) Receptor model analysis method

[0066] Principle: Based on the law of conservation of mass, the pollutant concentration monitored by the station is decomposed into a linear combination of different pollution sources. The calculation formula is as follows:

[0067] ;

[0068] Among them The contribution concentration of pollution source i. Its contributing factor, C 站点 This refers to the total concentration of pollutants measured at the monitoring station.

[0069] Implementation steps:

[0070] Component monitoring: Obtaining characteristic components in PM2.5 (such as NO3). ⁻ Indicating vehicle source, SO4 ²⁻ K⁺ indicates coal combustion source; K⁺ indicates biomass combustion.

[0071] Factor decomposition: The positive matrix factorization (PMF) model is used to decompose the pollutant concentration matrix at the site into: source component spectrum matrix (the proportion of each pollution source component) and contribution concentration matrix (the contribution value of each pollution source to each high value event).

[0072] The contribution rate is calculated using the following formula: .

[0073] 2) Trajectory Reverse Simulation Method

[0074] Core technology: based on meteorological flow field reconstruction and air mass trajectory tracking.

[0075] Forward diffusion: Simulate the diffusion path of pollutant emissions from the source to predict the concentration at the site.

[0076] Reverse source tracing: Tracing the origin of air masses from the site location to pinpoint upstream pollution sources.

[0077] Output results: Display a spatial heat map of contribution rate (showing the intensity of contribution of surrounding areas to the high value of the site); ranking of key sources (list of pollution source enterprises with contribution rate >90%).

[0078] 3) Dynamic weighting evaluation method

[0079] For high-value events with multiple sources and mixed characteristics, time weights and spatial weights are introduced:

[0080] The formula for calculating time weights is as follows:

[0081] ;

[0082] Wt represents the time weight, indicating the degree to which the pollution source contributes to the peak concentration at the site over time. A larger value indicates a higher degree of temporal matching.

[0083] The overlap between the emission period of the pollution source and the peak concentration at the monitoring station: refers to the length (or proportion) of overlap between the actual time interval of pollutant emission from the pollution source and the time interval during which the pollutant concentration at the monitoring station reaches its highest value.

[0084] Total high value duration: refers to the total duration during which the pollutant concentration observed at the monitoring station remains at a high level (exceeding a specific threshold).

[0085] The formula for calculating spatial weights is as follows:

[0086] ;

[0087] Ws: Spatial weight, which reflects the proximity of pollution sources to the site in terms of spatial distance. The closer the distance, the greater the weight, reflecting the dominant role of short-distance transmission.

[0088] Distance from pollution source to monitoring station²: This refers to the square of the straight-line distance between the source of the pollutant and the monitoring station. Using the squared form is typically used to amplify the weight of nearby sources (i.e., the weight decreases sharply as the distance increases slightly).

[0089] Overall contribution rate:

[0090] ;

[0091] C 综合 (Comprehensive Contribution): This refers to the final quantitative contribution of a specific pollution source to the concentration monitored at the site, after simultaneously considering both temporal and spatial factors.

[0092] C 源 (Source contribution concentration): refers to the baseline pollutant concentration generated by the pollution source alone, or the basic value of the source contribution calculated in advance by the source apportionment model.

[0093] Example: During a high PM2.5 event at a certain station, a steel plant 10km away continuously emitted PM2.5 during the concentration rise phase. Its contribution rate is calculated as follows:

[0094] C 钢铁厂 =38ug / m 3 ×2h3h×1102=25.3%.

[0095] Furthermore, the high-value classification technology implementation path includes classifying high-value events and extracting feature patterns based on differences in triggering conditions (meteorology, emissions, transmission), specifically including the following methods:

[0096] 1) Analytical Dimensions

[0097] Weather-triggered Static stability index (wind speed < 1.5 m / s + inversion layer height < 500 m) Local accumulation type, regional transmission type Emission-driven Industrial emission intensity surge rate, traffic flow change rate Industrial source burst type, motor vehicle peak type Compound pollution type <![CDATA[O3 / PM2.5 synergistic increase rate, secondary ion generation rate]]> Photochemical-particulate coupled

[0098] 2) Dynamic clustering and classification process, such as Figure 3 :

[0099] 1. High-value event data: Integrating PM2.5, NO... x Online monitoring of VOCs, satellite remote sensing, meteorological data, and emission inventory data are used to screen for periods / areas with excessive or abnormally high values.

[0100] 2. Feature Extraction: Extract key indicators (such as concentration gradient, wind direction and speed, temperature and humidity, and traffic / industrial activity intensity) from the spatiotemporal sequence and construct a multidimensional feature vector.

[0101] 3. Dynamic clustering decision: Adaptive clustering algorithms (such as DBSCAN, K-means++ or deep learning embedding) are used to dynamically divide the pollution type clusters according to the real-time data distribution, avoiding the limitations of static thresholds.

[0102] 4. Six types of emissions: industrial emissions-dominated, traffic exhaust accumulation, regional transmission mixed, dust pollution, biomass combustion, and combined pollution.

[0103] 5. Generate subtype labels: Automatically label the dominant cause type for each high-value event to support targeted control and source tracing analysis.

[0104] This process achieves a closed loop of "data → features → clustering → labels", improving the efficiency of pollution cause analysis and the accuracy of governance.

[0105] 3) Construction of the typing feature library

[0106] Each high-value subtype contains three types of features:

[0107] Core triggering factors: such as "industrial source outbreak type" should include: SO2 hourly concentration surge rate >200%, wind speed <1m / s;

[0108] Typical pollution fingerprint: Component characteristics: Zn / Cu ratio > 3 (indicating industrial smelting), black carbon concentration > 15%;

[0109] Contribution source structure: Preset contribution rate ranking: industrial sources (50±15%) > motor vehicles (20±8%) > dust (10±5%).

[0110] Furthermore, the construction and application of the high-value AI traceability library in this invention includes the following:

[0111] 1) High-value AI traceability library architecture design, such as Figure 4 :

[0112] Input layer: historical high-value events (PM2.5 / O3, etc.) + multi-source heterogeneous data (meteorology, emissions, remote sensing, online monitoring) → feature matrix (spatiotemporal alignment + dimensionality reduction).

[0113] AI Faction Engine: Unsupervised / Semi-supervised Clustering → Three Typical Contamination Patterns:

[0114] Weather stable type (low boundary layer, low wind speed);

[0115] Industrial explosion type (sudden increase in power source, local dominance);

[0116] Regional transmission type (wind direction transport, cross-regional contribution);

[0117] Knowledge building layer: Each type of event is mapped to "triggering condition threshold + contribution source structure weight" → building a structured knowledge base.

[0118] Knowledge graph layer: Entities (pollution sources / meteorological factors / geographical units) + Relationships (causality / contribution / spatiotemporal correlation) → Support reasoning and querying.

[0119] Application layer: Real-time matching of current observation data → dynamic identification of types, quantification of source contribution, and support for precise pollution control decisions.

[0120] 2) Dynamic updates

[0121] Incremental learning: When the classification error of a new event is greater than 20%, the model is retrained and the source component spectral database is updated every quarter.

[0122] Weight decay: The weight of events from 5 years ago is reduced to 0.3 to ensure the timeliness of the library;

[0123] Mixed contamination analysis: Introducing a convolutional neural network to separate mixed signals;

[0124] Small-scale source tracing accuracy: LiDAR scanning + 3D flow field reconstruction (1km×1km resolution).

[0125] Furthermore, S4, K-means clustering and fingerprint generation: K-means / PCA is embedded to cluster urban monitoring station data, and meteorological data is combined to trace the pollution transmission path. Convolutional neural networks (CNN) are used to analyze high-resolution satellite images (such as Sentinel-5 and fused with radar images (Sentinel-1)) to identify the distribution of pollution sources in industrial areas, and to conduct pollution source identification and tracing analysis.

[0126] Preferably, the multi-source data fusion processing flow is as follows: Figure 5 As shown, the specific steps include:

[0127] 1) Divide the city into 1km×1km grids, and deploy micro monitoring stations at the center of each grid;

[0128] 2) Construct a pollutant concentration-meteorological joint matrix: (WS: wind speed, WD: wind direction, n is the number of timestamps).

[0129] 3) K-means clustering: The number of clusters k is dynamically set according to meteorological fluctuations (k=3-5);

[0130] 4) Output: Probability map of dominant contamination transmission paths;

[0131] Example: Cluster center 1 (northeast wind, wind speed 3-5m / s) corresponds to path A, contributing 42%, and the accuracy of the path is verified by concentration gradient.

[0132] 5) Satellite imagery pollution source identification:

[0133] 6) Data preprocessing: Download NO2 / SO2 column concentration data from Sentinel-5P (resolution 2×2km²); overlay high-resolution optical images (such as GF-2, 0.8m resolution) to locate industrial facilities.

[0134] 7) Convolutional Neural Network (CNN) Modeling:

[0135] like Figure 6As shown, in air pollution control, CNN extracts spatial-temporal features through multi-source data fusion (such as satellite remote sensing, ground monitoring stations, meteorological data, etc.) to achieve pollutant concentration prediction or pollution source identification. Its convolutional layers automatically learn local patterns (such as PM2.5 diffusion gradient), pooling layers reduce dimensionality and retain key information, and fully connected layers integrate multi-source heterogeneous data, ultimately outputting classification / regression results, improving prediction accuracy and generalization ability.

[0136] This is a pseudocode for a convolutional neural network (CNN) built using the Keras Sequential API, designed for a three-class classification task (coal / chemical / metallurgical), with an input image size of 256×256×3.

[0137] Output shapes for each layer:

[0138] 1. First convolutional layer (32 3×3 filters) → Output 254×254×32;

[0139] 2. Max pooling layer (2×2) → Output 127×127×32;

[0140] 3. Second convolutional layer (64 3×3 filters) → Output 125×125×64;

[0141] 4. Flatten the layer → transform it into a 1,000,000-dimensional vector;

[0142] 5. Fully connected layer (128 neurons) → Output 128 dimensions;

[0143] 6. Output layer (3 neurons, softmax) → Output 3-dimensional (three types of probabilities);

[0144] Total number of parameters:

[0145] (1) First convolutional layer: 896;

[0146] (2) Second convolutional layer: 18,496;

[0147] (3) First fully connected layer: 128,000,128;

[0148] (4) Output layer: 387;

[0149] (5) Pooling and flattening have no parameters.

[0150] 8) Generation of a contaminated fingerprint database, as shown in the following example:

[0151] Coal-fired power plants <![CDATA[High SO₂ + Rectangular Cooling Tower Image]]> FP-01 Chemical Industrial Park <![CDATA[NO2 / SO2 mixture + dense pipe network image]]> FP-02 Iron and steel smelting High CO + Chimney Cluster Images FP-03

[0152] Source tracing decision engine: multi-evidence fusion logic

[0153] like Figure 7 As shown, this process is a multi-evidence fusion logic with an adaptive feedback mechanism, which aims to improve the credibility of pollution source identification and assessment results.

[0154] Its core process is as follows:

[0155] 1. Initialization and Characteristic Evaluation: The system accesses and evaluates various monitoring data sources (such as ground stations, remote sensing satellites, meteorological data, emission inventories, etc.), establishes inter-source relationships, and assigns them initial confidence levels.

[0156] 2. Stress Loop Tuning: This is a crucial adaptive step. It simulates a conflict detection mechanism. When data from different sources contradict each other (e.g., significant differences between satellite inversion and ground monitoring values), this loop triggers initial parameter adjustments to enhance the algorithm's robustness to anomalous or interfering data.

[0157] 3. Dynamic Reweighting and Fusion: Based on the optimization results, the weights of each data source are dynamically updated to ensure that evidence with high consistency and quality dominates the fusion process. Subsequently, multi-source features are mathematically fused using weighted synthesis rules.

[0158] 4. Iterative convergence and output: The above process often requires multiple iterations until the fusion result tends to stabilize, and finally outputs a highly credible conclusion.

[0159] S5 improves clustering performance through re-tuning, resulting in the contaminated fingerprint database V0.

[0160] In step S5, to achieve "improving clustering performance through re-optimization to form a contaminated fingerprint database V0," the core innovation lies in constructing a dynamic iterative optimization framework, rather than a one-time static clustering. Here, "re-optimization" specifically refers to introducing feedback mechanisms and adaptive learning to optimize the clustering process multiple times until a stable and highly discriminative contaminated fingerprint database is formed. For example... Figure 8 As shown, the specific implementation path is as follows:

[0161] S51, Initial Clustering and Feature Assessment: First, initial clustering is performed based on multi-source monitoring data (such as online components of enterprise emissions, meteorological factors, and receptor point exceedance factors) to generate preliminary fingerprint categories. Subsequently, the internal consistency (such as silhouette coefficient) and external discriminative power (such as the ability to distinguish between different pollution events) of each cluster are assessed.

[0162] S52, introduces a feedback loop for optimization:

[0163] Feedback based on the source tracing results: The fingerprints obtained from the initial clustering are applied to actual source tracing cases. By comparing the degree of agreement between the contribution simulated by the model and the actual situation, fingerprint categories with weak discriminative power (i.e., leading to misjudgment or ambiguous source tracing) are identified.

[0164] Adaptive learning based on new data: The system continuously incorporates new monitoring data. It uses the new data to incrementally learn or fine-tune the existing clustering model, paying particular attention to data points that fall within the ambiguous regions of the existing cluster boundaries, triggering the re-division or merging of specific clusters.

[0165] S53, Dynamic Parameters and Feature Reweighting: Based on feedback results, automatically adjust clustering algorithms (such as the neighborhood parameters of DBSCAN, the K value of K-means) or feature weights. For example, if a chemical component is proven to be a key differentiating factor in multiple source tracings, its weight in the feature space will be increased.

[0166] S54, Iterative Convergence and V0 Database Formation: Repeat steps S52 and S53 to form a closed loop of "clustering-application-evaluation-tuning". When the change in the cluster structure is less than a preset threshold in several consecutive iterations, and its source tracing accuracy on the independent test case set reaches a stable optimal level, the algorithm is considered to have converged. The stable and optimized clustering results obtained at this time are solidified into the contaminated fingerprint database V0. Each fingerprint in this database is associated with a proven, highly discriminative feature combination and dynamic pattern.

[0167] S6 utilizes a small-scale real-time source tracing platform to construct a real-time source tracing algorithm framework: the emission inventory is split into hours by dynamic time allocation coefficients, coupled with AERMOD, to calculate the concentration and contribution rate of each pollution source to the spatial grid and sites in real time, forming a new pollution fingerprint database V1.

[0168] The core of step S6 lies in constructing a real-time, dynamically coupled computational framework for emission inventories and diffusion models, achieving a leap from static inventories to hourly dynamic source apportionment, thereby generating a more accurate and timely pollution fingerprint database V1. For example... Figure 9 As shown, the specific implementation process is as follows:

[0169] S61, Dynamic Time Allocation Factors and Hourly Emission Inventory Generation:

[0170] Traditional emission inventories are typically measured on an annual or monthly scale. This method uses multi-dimensional, dynamic time-allocation coefficients to achieve hourly breakdowns. These coefficients are not only based on typical industry production cycles (such as the "load-emission" curve for electricity and the batch production rhythm of chemicals), but also integrate real-time enterprise operating condition monitoring data (such as production equipment current, valve opening, and DCS system flow) and specific process parameters. For example, by monitoring boiler blast volume and furnace temperature in real time, combined with fuel sulfur content, the actual SO2 emission rate for that hour is dynamically calculated, rather than using a fixed average coefficient.

[0171] This invention constructs a coefficient self-learning module. The system continuously compares the differences between emissions estimated based on allocation coefficients and emission outlet online monitoring (CEMS) data, and uses machine learning (such as online regression models) to dynamically correct and optimize the time allocation coefficient library for each enterprise and each process step, so that the estimated values ​​infinitely approach the actual emission dynamics.

[0172] S62, real-time coupled calculation with the AERMOD model:

[0173] The generated hourly dynamic emission inventory is coupled with the synchronously acquired hourly meteorological field (WRF model output or actual assimilation data) to drive the AERMOD model for real-time simulation. A strategy combining pre-calculation of the "source-receptor" contribution matrix and real-time interpolation is employed. To address the massive demands of real-time computation, the unit source strength contribution transfer coefficient (i.e., response matrix) of each pollution source to each spatial grid and monitoring station under typical meteorological conditions is pre-calculated. During real-time execution, based on the current actual meteorological conditions, rapid interpolation or similarity matching is performed in the pre-calculated response matrix library, and then multiplied by the actual dynamic source strength of the current hour, thus obtaining the concentration contribution of each source to the grid and station in near real-time. This solves the computational bottleneck of traditional hourly execution of the complete AERMOD model.

[0174] S63, forming a real-time tracing algorithm framework and a new fingerprint database V1:

[0175] The framework runs continuously. Every time it receives new hourly meteorological and operating data, it triggers a complete "dynamic inventory generation - diffusion simulation - contribution calculation" process, outputting the real-time contribution concentration and contribution rate of all pollution sources to the receptor point.

[0176] These real-time contribution data are stored in a structured manner and correlated with pollution events. When a pollution exceedance event occurs, the system can immediately trace back the evolution of the contribution of each pollution source over the past few hours, forming a dynamic "contribution spectrum" fingerprint that is strongly correlated with meteorological conditions. The V1 database not only contains the static component characteristics of pollution sources (from V0), but more importantly, it adds their dynamic diffusion characteristics and contribution potential patterns under different meteorological scenarios. For example, the fingerprint of a chemical plant in V1 is associated with the rule pattern of "a rapid increase in the contribution rate to the station 5 kilometers downwind under northerly winds of level 2-3 and a stable atmospheric stratification." This upgrades the fingerprint from a "composition label" to a "behavioral model" with predictive capabilities.

[0177] S7, repeating S5-S6, forms a contamination fingerprint map with regional characteristics. This contamination fingerprint map integrates a closed-loop learning framework that combines multi-source feedback, adversarial verification, and dynamic evolution of the map structure. The iteration termination in S7 is determined by a dynamic convergence mechanism. When the fingerprint vector drift rate is <1-2%, the core source tracing F1 score is >85-90%, the unknown event rate is <5%, and the model has covered the entire seasonal cycle, convergence is determined, and the iteration terminates. Simultaneously, a maximum number of iterations is used as a safety circuit breaker to generate a regional contamination fingerprint map that combines mathematical stability with business practicality.

[0178] The core of step S7 lies in constructing a regional contamination fingerprint knowledge system with continuous evolution capabilities. Its implementation relies on a closed-loop learning framework that integrates multi-source feedback, adversarial verification, and dynamic evolution of the graph structure, such as... Figure 10 As shown, the specific implementation process is as follows:

[0179] S71, Constructing the "skeleton" of regional maps and the dynamic update mechanism:

[0180] Based on the V1 database, pollution source fingerprints, meteorological scenarios, geospatial grids, and receptor point concentrations are linked to form the initial "skeleton" of a multidimensional, heterogeneous knowledge graph. The nodes of this graph include pollution sources, receptor points, meteorological conditions, and chemical compositions, while edges represent contribution relationships, spatial correlations, or conditional probabilities. The system continuously imports new source tracing case results, new enterprise emission data, and new monitoring data. This new data triggers graph updates: confirming and strengthening repeatedly validated "source-receptor-meteorological" association paths (increasing edge weights); reducing or removing contradictory or invalid associations (pruning); and discovering new potential association patterns (adding nodes or edges). This transforms the graph from a static snapshot into a dynamically growing organism.

[0181] S72 introduces adversarial verification and uncertainty quantification:

[0182] To enhance the robustness and reliability of the graph, the system proactively generates "adversarial" test scenarios. For example, it intentionally inputs meteorological data or emission combinations that slightly deviate from historical patterns to test the stability of the graph's recommended source tracing results. Alternatively, it utilizes Generative Adversarial Networks (GANs) to generate indistinguishable simulated pollution events to challenge the existing graph's discrimination capabilities. For each association and prediction result in the graph, an uncertainty quantification index is attached. This index integrates data quality, model simulation error, and case validation frequency.

[0183] When users use graphs for tracing or issuing warnings, they can not only obtain conclusions but also know the confidence level of those conclusions, thus supporting more prudent decision-making.

[0184] S73 forms a regionally specific "pollution climatology" map:

[0185] Through long-term operation and continuous learning, the system will eventually be able to extract macroscopic patterns of regional pollution formation and transport that go beyond individual cases. The map will be able to answer questions such as, "Under stable winter weather conditions, what are the main contributing sources and their spatial distribution characteristics in this region?" or "When prevailing southeasterly winds blow, which receptor points are most susceptible to cross-border transport?" The final map will be a "pollution climatology" knowledge base deeply embedded in regional geography, industrial layout, and meteorological and climatic characteristics. It is not only a source tracing tool but also a decision-making sandbox capable of simulating environmental benefits under different control scenarios, providing dynamic and intelligent core knowledge support for precise prevention and control of regional air pollution at its source.

[0186] S8, Pollution fingerprinting visualization and analysis utilizes heat maps, radar charts, spatial distribution maps, and regional characteristic analysis maps to display air pollution conditions.

[0187] To further illustrate the role of the atmospheric pollution fingerprint database generated using the method of this invention, this invention takes the application of the atmospheric pollution fingerprint database for VOCs prevention and management in an industrial park in Zhenjiang City as an example. Its core lies in upgrading the static database into a dynamic, intelligent control system deeply coupled with the actual situation of the industrial park. The specific implementation path is as follows:

[0188] Constructing a VOCs fingerprint map specific to the industrial park: Integrating online monitoring data, leak detection and repair records, process material balance data, and VOCs component data from all VOCs-related enterprises within the park, as well as VOCs component data from park boundary stations, micro-stations, and mobile monitoring, a VOCs pollution fingerprint database is formed through multi-source data fusion and deep learning clustering. This database can accurately characterize the emission characteristics of each enterprise and each process step. This database not only includes chemical composition profiles but also correlates emission time-series patterns with operating conditions, forming a dynamic "behavioral model."

[0189] Real-time source tracing and rapid early warning: Utilizing a small-scale real-time source tracing platform, coupled with a dynamic emission inventory and atmospheric diffusion model, the system calculates in real-time the contribution rate of each pollution source to VOCs concentrations at sensitive points and grids within the park under different meteorological conditions. Once VOCs concentrations exceed standards or experience abnormal fluctuations, the system can immediately identify suspected pollution sources and their contribution share, achieving source tracing within minutes. The tracing results are then fed back to the database as new cases, continuously optimizing the fingerprint spectrum's discrimination capabilities.

[0190] Implement precise prevention and tiered management: Based on fingerprint mapping, scientifically classify VOCs emission sources in the industrial park. Install operational monitoring and intelligent control equipment for key controlled enterprises (such as those with large emissions, high reactivity, and obvious fingerprint characteristics), linking their emission dynamics to the fingerprint database in real time, realizing the transformation from "end-of-pipe monitoring" to "process prevention." Through continuous learning and improvement, a VOCs pollution fingerprint map with characteristics specific to Zhenjiang Industrial Park will be formed. This map can reveal the dominant sources, transmission paths, and accumulation areas of VOCs in the park under specific meteorological conditions, providing a scientific basis for formulating differentiated and precise staggered production and shutdown / restriction plans.

[0191] Assessing Governance Effectiveness and Continuous Optimization: The fingerprint database is applied to assess the effectiveness of governance. By comparing changes in the intensity of enterprise emission fingerprints and their contribution to the park's environment before and after the implementation of control measures, the actual environmental benefits of governance measures for individual enterprises or regions are quantitatively evaluated. Simultaneously, all monitoring, tracing, and assessment data are continuously input into the system, driving the entire fingerprint map and control strategy to adaptively learn and iteratively optimize, forming a closed-loop intelligent management system of "monitoring-tracing-control-assessment-optimization".

[0192] This invention breaks through the single-dimensional limitations of traditional manual analysis, providing a "pollution dictionary"-like knowledge base for pollution prevention and control. It enables high-value pollution events to be rapidly matched against historical fingerprint databases, with a calculation speed reaching the second level. Compared to manual analysis, it shortens the high-value response time from hours to minutes, enabling real-time early warning. Furthermore, it uncovers the hidden characteristics of high-value pollution processes, improving the accuracy of high-value cause diagnosis. Because this invention can quickly match historical fingerprint databases and identify high-value pollutant locations, it facilitates the precise location of entities exceeding concentration or frequency of emissions, supporting the formulation of refined control strategies.

[0193] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them; although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications can still be made to the specific implementation of the present invention or equivalent substitutions can be made to some technical features without departing from the spirit of the technical solutions of the present invention, and all such modifications and substitutions should be covered within the scope of the technical solutions claimed in the present invention.

Claims

1. A method for generating an atmospheric pollution fingerprint database by fusing multi-source heterogeneous data, characterized in that: Includes the following steps: S1, collect air pollution data, meteorological data, geographic information, and daily inspection records from national monitoring stations to form a basic database for air pollution prevention and control; the air pollution data includes sulfur dioxide (SO2), nitrogen dioxide (NO2), and inhalable particulate matter (PM2.5). 10 Fine particulate matter (PM) 2.5 The meteorological data includes temperature, humidity, air pressure, wind speed, and wind direction; carbon monoxide (CO) and ozone (O3). S2, Feature Matrix Design: Construct a feature matrix with single pollution events or single-day pollution processes as rows and core indicators as columns; where the row content includes event definition standards, time attributes, spatial attributes, and event classification labels, and the column content includes pollution intensity, meteorological conditions, temporal dynamics, spatial impact, and pollution source characteristics; S3, Site High Value Analysis: By analyzing the total number of high values ​​at a site, or the annual and monthly high value data, the contribution of pollution to the site is analyzed, and the high value is categorized using the triggering conditions to form a high value AI source tracing library. The contribution of pollution to the site is quantified through multi-dimensional data analysis and model calculation to quantify the contribution ratio of different pollution sources to the site's high values. The high value categorization steps include classifying high value events and extracting feature patterns based on the differences in triggering conditions. S4, K-means clustering and fingerprint generation: K-means / PCA is embedded to cluster urban monitoring station data, and meteorological data is combined to trace the pollution transmission path. Convolutional neural network (CNN) is used to analyze high-resolution satellite imagery to identify the distribution of pollution sources in industrial areas and to conduct pollution source identification and tracing analysis. S5 improves clustering performance through re-tuning, forming the contaminated fingerprint database V0; S6, using a small-scale real-time source tracing platform to build a real-time source tracing algorithm framework: the emission inventory is split into hours by dynamic time allocation coefficient, coupled with AERMOD, and the concentration and contribution rate of each pollution source to the spatial grid and site are calculated in real time to form a new pollution fingerprint database V1; S7, repeat S5-S6 to form a pollution fingerprint spectrum with regional characteristics: the pollution fingerprint spectrum integrates a closed-loop learning framework of multi-source feedback, adversarial verification and dynamic evolution of spectrum structure. S8, Pollution Fingerprint Visualization and Analysis: Displaying air pollution conditions using heat maps, radar charts, spatial distribution maps, and regional characteristic analysis maps.

2. The method for generating an atmospheric pollution fingerprint database according to claim 1, characterized in that: The event definition criteria include single pollution events and single-day pollution processes. A single pollution event is marked by the continuous exceedance of pollutant concentration, and a single-day pollution process is based on a calendar day. The time attributes include the start / end time of the event, the total duration, and the time of the pollution peak. The spatial attributes include the geographical scope of the impact, the coordinates of the core pollution area, and the pollution transmission path. The event classification label includes classification according to pollution intensity.

3. The method for generating an atmospheric pollution fingerprint database according to claim 2, characterized in that: Pollution intensity indicators include daily average PM2.5 concentration, ... 10 Peak concentration, 90th percentile of O3-8h and SO2 / NO2 ratio; meteorological condition indicators including average relative humidity, minimum wind speed, inversion layer height and diurnal temperature range; temporal dynamic indicators including number of days of pollution, rate of increase of concentration and residence time after peak; spatial impact indicators including number of affected cities, cross-regional transport contribution rate and pollution radius; pollution source characteristic indicators including traffic flow index, industrial emission intensity, biomass combustion labeling and proportion of secondary inorganic sources.

4. The method for generating an atmospheric pollution fingerprint database according to claim 1, characterized in that: The multi-dimensional data analysis and model calculation include the receptor model analytical method, the trajectory reverse simulation method, and the dynamic weight evaluation method.

5. The method for generating an atmospheric pollution fingerprint database according to claim 1, characterized in that: The steps for re-optimizing and improving clustering performance include: S51, Initial Clustering and Feature Evaluation: Initial clustering is performed based on multi-source monitoring data to generate preliminary fingerprint categories. Then, the internal consistency and external discriminative power of each cluster are evaluated. S52, Introducing a feedback loop for optimization: including feedback based on source tracing results and adaptive learning based on new data; the feedback based on source tracing results includes: applying the fingerprints obtained from the initial clustering to actual source tracing cases, and identifying fingerprint categories with weak discriminative power by comparing the contribution of the model simulation with the actual situation; the adaptive learning based on new data includes: the system continuously inputs new monitoring data, uses the new data to incrementally learn or fine-tune the existing clustering model, focuses on data points falling in the fuzzy area of ​​the existing cluster boundary, and triggers the re-division or merging of specific clusters; S53, Dynamic Parameters and Feature Reweighting: Automatically adjusts the clustering algorithm or feature weights based on feedback results; S54, Iterative convergence and V0 database formation: Repeat steps S52 and S53. When the change in the cluster structure is lower than the preset threshold in several consecutive iterations and the source tracing accuracy on the independent test case set reaches a stable optimal level, the algorithm is determined to have converged. The stable and optimized clustering results obtained at this time are then solidified into the contaminated fingerprint database V0.

6. The method for generating an atmospheric pollution fingerprint database according to claim 1, characterized in that: The steps for constructing a real-time source tracing algorithm framework using a small-scale real-time source tracing platform include: S61, Dynamic Time Allocation Coefficient and Hourly Emission Inventory Generation: Construct a coefficient self-learning module to continuously compare the difference between emissions estimated based on allocation coefficients and online monitoring data of emission outlets, and use machine learning to dynamically correct and optimize the time allocation coefficient library so that the estimated values ​​approximate the actual emission dynamics. S62, Real-time Coupled Calculation with the AERMOD Model: The generated hourly dynamic emission inventory is coupled with the synchronously acquired hourly meteorological field to drive the AERMOD model for real-time simulation; a strategy combining source-receptor contribution matrix pre-calculation and real-time interpolation is adopted; during real-time operation, based on the current actual meteorological conditions, fast interpolation or similarity matching is performed in the pre-calculated response matrix library, and then multiplied by the actual dynamic source strength of the current hour to obtain the concentration contribution of each source to the grid and station; S63, forming a real-time source tracing algorithm framework and a new fingerprint database V1: The real-time source tracing algorithm framework runs continuously. Every time new hourly meteorological and operating condition data is received, a complete dynamic inventory generation-diffusion simulation-contribution calculation process is triggered, outputting the real-time contribution concentration and contribution rate of all pollution sources to the receptor point.

7. The method for generating an atmospheric pollution fingerprint database according to claim 1, characterized in that: In S7, the closed-loop learning framework includes: S71, Constructing the framework and dynamic update mechanism of the regional map: Based on the V1 database, pollution source fingerprints, meteorological scenarios, geospatial grids, and receptor point concentrations are linked to form a multi-dimensional, heterogeneous knowledge graph initial framework; then the system continuously imports new source tracing case results, new enterprise emission data, and new monitoring data, and the new data triggers map updates: confirming and strengthening repeatedly verified source-receptor-meteorological association paths; downgrading or removing contradictory or invalid associations; and updating potential association patterns; S72, introduce adversarial verification and uncertainty quantification: for each association and prediction result in the graph, an uncertainty quantification index is attached, which integrates data quality, model simulation error and case verification frequency; S73, forming a pollution fingerprint map with regional characteristics: through long-term operation and continuous learning, the macro-level regional pollution formation and transport patterns that transcend individual cases are finally extracted. The pollution fingerprint map with regional characteristics is a pollution climatology knowledge base deeply embedded in regional geography, industrial layout, and meteorological and climatic characteristics.

Images