A method and device for predicting a composite atmospheric health risk grading early warning index

Abstract

This application relates to a method and device for predicting an atmospheric complex health risk classification and early warning index. By integrating the health effects of temperature and multiple pollutants, an atmospheric complex health risk classification and early warning index is obtained. This can alleviate the shortcomings of traditional technologies that ignore the independent and interactive health effects of temperature, and can more completely and realistically reflect the complex health risks faced by people in the actual atmospheric environment.

Description

Technical Field

[0001] This application relates to health risk assessment technology, specifically to a method and apparatus for predicting an atmospheric composite health risk classification and early warning index. Background Technology

[0002] The impact of air pollution and meteorological factors on human health has become a major public health concern worldwide. The Air Quality Health Index (AQHI), as an important environmental health risk management tool, has been widely used in many countries and regions. Traditional AQHI and its derivative indices mainly construct a comprehensive index reflecting the potential health risks of short-term air pollution exposure by integrating the concentrations of various common air pollutants, and based on this, issue health risk warnings to guide the public in taking protective measures.

[0003] However, the traditional AQHI system focuses primarily on air pollutants themselves when assessing health risks to the population. Its assessment of health risks has a systematic bias and cannot fully reflect the complex health threats posed by the atmospheric environment, resulting in inaccurate prediction results.

[0004] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0005] This application provides a method and apparatus for predicting atmospheric composite health risk classification and early warning index, which alleviates the technical problem of inaccurate prediction results in the current application.

[0006] To achieve the above objectives, embodiments of this application provide a method for predicting an atmospheric composite health risk classification and early warning index, comprising: Near-ground concentration data of various air pollutants and ambient temperature data that are spatiotemporally matched with the near-ground concentration data are obtained through multi-source data fusion. Based on the ambient temperature data and the corresponding regional population health outcome data, exposure-response relationship models of ambient temperature and each air pollutant with population health outcomes are constructed to determine the independent excess health risk contribution of ambient temperature and each air pollutant. Based on the independent excess health risk contribution of each pollutant, the independent excess health risk contribution of ambient temperature is integrated with the independent excess health risk contribution of each air pollutant using an additive model to determine a composite health risk value for characterizing the temperature-pollutant combined exposure effect. Based on the statistical distribution of the composite health risk values ​​and the risk level thresholds dynamically adjusted according to the differences in environmental temperature distribution characteristics of different geographical and climatic zones, the atmospheric composite health risk classification and early warning index for each grid unit is determined.

[0007] This application also provides an atmospheric composite health risk classification and early warning index prediction device, which includes: The acquisition module is used to acquire near-ground concentration data of various air pollutants and ambient temperature data that is spatiotemporally matched with the near-ground concentration data through multi-source data fusion. The identification module is used to construct exposure-response relationship models between ambient temperature and various air pollutants and population health outcomes based on the ambient temperature data and the corresponding regional population health outcome data, so as to determine the independent excess health risk contribution of ambient temperature and various air pollutants. The determination module is used to integrate the independent excess health risk contribution of the ambient temperature with the independent excess health risk contribution of each air pollutant based on an additive model, according to each independent excess health risk contribution, to determine a composite health risk value for characterizing the temperature-pollutant combined exposure effect. The prediction module is used to determine the atmospheric composite health risk classification and early warning index for each grid unit based on the statistical distribution of the composite health risk value and multiple risk level thresholds dynamically adjusted according to the differences in environmental temperature distribution characteristics of different geographical and climatic zones.

[0008] This application provides a method and apparatus for predicting the atmospheric composite health risk classification and early warning index. It elevates ambient temperature from a confounding variable in traditional models to a core risk variable alongside air pollutants, systematically integrating the health effects of temperature and multiple pollutants. The resulting atmospheric composite health risk classification and early warning index (AtmosHI-S) can fundamentally overcome the shortcomings of traditional AQHI, which ignores the independent and interactive health effects of temperature. It can more completely and realistically reflect the composite health risks faced by people in the actual atmospheric environment.

[0009] Other features and advantages of this application will be described in detail in the following detailed description section. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] To gain a more complete understanding of this application and its beneficial effects, the following description will be provided in conjunction with the accompanying drawings, wherein the same reference numerals in the following description denote the same parts.

[0012] Figure 1This is a flowchart illustrating a method for predicting atmospheric composite health risk grading and early warning index provided in an embodiment of this application. Figure 2 This is a schematic diagram of an atmospheric composite health risk classification and early warning index prediction device provided in an embodiment of this application. Detailed Implementation

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the protection scope of this application.

[0014] In traditional air quality health index assessment systems, ambient temperature is not included as a core risk variable, resulting in a structural flaw in the health risk assessment model. This flaw manifests as the failure to systematically quantify the independent impact of temperature on population health and its interaction with air pollutants, making the risk assessment results unable to accurately reflect the true health burden of multi-factor combined exposure in the actual atmospheric environment.

[0015] For example, during the winter heating season in northern China, the coexistence of low temperatures and high concentrations of PM2.5 creates a typical combined exposure scenario. In this scenario, when the ambient temperature is below 5 degrees Celsius and the PM2.5 concentration exceeds 75 micrograms per cubic meter, the combined effects of low temperature and particulate matter exposure on the human respiratory and cardiovascular systems produce significant synergistic health effects. However, existing air quality health indices only assess risk based on the concentration of pollutants such as PM2.5, failing to incorporate low temperature as an independent risk factor into the calculation framework. Consequently, under these combined exposure conditions, the actual health risks are systematically underestimated, and the risk warning level is lower than the true risk level.

[0016] If the aforementioned technical problems are not resolved, the scientific validity and practicality of the atmospheric health risk assessment system will be fundamentally limited. Bias in the risk assessment results will lead to systematic misjudgments of early warning information in different climatic regions, rendering the early warning system ineffective when extreme temperature events and pollution events overlap. To address this, this application proposes a method for predicting an atmospheric composite health risk classification and early warning index. In this application:

[0017] Multi-source data fusion refers to a method that integrates data from different sources and types, such as satellite remote sensing data, ground monitoring data, and meteorological reanalysis data, and applies specific processing techniques (such as machine learning and statistical models) to generate more comprehensive, accurate, and higher spatiotemporal resolution integrated data products. This approach aims to overcome the limitations of a single data source and improve data quality and coverage.

[0018] Near-ground concentration data refers to the concentration of air pollutants at altitudes close to the Earth's surface (usually the height of the human breathing zone). This data directly reflects the actual level of air pollution exposure to the population and is a key input for assessing health risks.

[0019] Ambient temperature data refers to air temperature data within a specific area and time period. This data not only affects public health as an independent environmental factor, but also interacts in complex ways with air pollutants, collectively constituting a complex health risk.

[0020] Population health outcome data refers to quantifiable population health status indicators related to environmental exposures, such as total non-accidental deaths, cardiovascular disease incidence, and respiratory disease hospitalization rates. This data is used to establish statistical associations between environmental factors and health impacts.

[0021] Exposure-response model: A mathematical model that describes the quantitative relationship between environmental exposure (such as pollutant concentration and temperature) and population health outcomes. This model uses statistical analysis to reveal how changes in exposure levels affect the risk or severity of health outcomes.

[0022] Independent excess health risk contribution: This refers to the additional risk increase in population health outcomes caused by a single environmental factor (such as a particular pollutant or ambient temperature) after controlling for other confounding factors. This contribution quantifies the independent impact of each factor on total health risk.

[0023] Additive model: refers to a mathematical model that simply sums the effects of multiple independent factors to calculate the total effect. In this embodiment, the model is used to integrate the independent excess health risk contributions of ambient temperature and various air pollutants to obtain a composite health risk value.

[0024] Composite health risk value: refers to the total excess risk to human health posed by the combined effects of ambient temperature and multiple air pollutants. This value characterizes the quantitative results of the combined temperature-pollutant exposure effect.

[0025] Risk level threshold: This refers to the critical value used to classify composite health risk values ​​into different risk levels (such as low risk, medium risk, high risk, etc.). This threshold is the basis for determining the early warning index, and its setting directly affects the sensitivity and accuracy of the early warning.

[0026] Atmospheric Composite Health Risk Grading and Early Warning Index: This is a graded early warning index that comprehensively reflects the impact of ambient temperature and various air pollutants on public health. The index aims to provide the public with intuitive health risk information and guide them in taking appropriate protective measures.

[0027] In some embodiments, such as Figure 1As shown, the atmospheric composite health risk classification and early warning index prediction method provided in this application includes the following steps: S101: Obtain near-ground concentration data of various air pollutants and ambient temperature data that is spatiotemporally matched with the near-ground concentration data through multi-source data fusion.

[0028] One approach involves using a single ground-based monitoring station to collect pollutant concentrations and ambient temperature. However, this method may result in limited data coverage and difficulty in acquiring large-area or high-resolution exposure data. Another approach utilizes satellite remote sensing to obtain large-scale pollutant column concentration data, which is then converted to near-surface concentrations using a simple empirical formula, while simultaneously obtaining regional average temperature data from meteorological departments. While this method expands coverage, it may have limitations in spatial resolution and near-surface concentration conversion accuracy, and it fails to adequately consider the heterogeneity of different data sources.

[0029] S102: Based on the ambient temperature data and the corresponding regional population health outcome data, construct exposure-response relationship models between ambient temperature and each air pollutant and population health outcomes, in order to determine the independent excess health risk contribution of ambient temperature and each air pollutant.

[0030] One approach uses a simple linear regression model to analyze the relationship between pollutant concentration and health outcomes, ignoring the nonlinear and lag effects of temperature. This approach may fail to accurately capture the complex impacts of environmental factors on health, leading to biased estimates of the independent excess health risk contribution. Another approach constructs a uniform exposure-response model based solely on regionally averaged data, without considering the differences between different grid cells. This approach may fail to reflect the specificity of local areas, reducing the model's granularity.

[0031] S103: Based on each independent excess health risk contribution, the independent excess health risk contribution of ambient temperature is integrated with the independent excess health risk contribution of each air pollutant using an additive model to determine a composite health risk value used to characterize the temperature-pollutant combined exposure effect.

[0032] One approach involves simply summing the excess health risk contributions of each factor using a basic arithmetic method. While conceptually intuitive, this approach may fail to adequately account for the potential nonlinear interactions between factors, leading to inaccurate estimates of the composite health risk value. Another approach involves assigning fixed weights to different factors based on expert experience and then performing a weighted summation. Although this method introduces weights, its subjectivity may affect the objectivity and universality of the composite health risk value.

[0033] S104: Based on the statistical distribution of the composite health risk value and the risk level thresholds obtained by dynamic adjustment based on the differences in environmental temperature distribution characteristics of different geographical and climatic zones, determine the atmospheric composite health risk classification and early warning index for each grid unit.

[0034] One approach involves using nationally standardized fixed percentiles (e.g., using P25, P50, P75, and P90 of the composite health risk value as thresholds) to classify risk levels. While simple to implement, this method may not adapt to the varying adaptations of populations to temperature and pollutants across different geographical and climatic regions, resulting in insufficient applicability of warnings in localized areas. Another approach involves setting a fixed set of risk level thresholds based on historical data without dynamic adjustments. This method may fail to respond promptly to changes in risk distribution caused by climate change or seasonal characteristics, reducing the flexibility and accuracy of warnings.

[0035] The following example will provide a more detailed explanation of the above technical solution: Suppose that in a specific area (Location A), an early warning of atmospheric complex health risks is required for the next 24 hours. The early warning system for this area first acquires high spatiotemporal resolution near-surface concentration data of air pollutants (such as PM2.5, O3, NO2, and HCHO) through multi-source data fusion, along with spatiotemporally matched ambient temperature data. Specifically, the system integrates satellite remote sensing data, ground monitoring station data, and meteorological reanalysis data. Through advanced fusion algorithms, it generates hourly and per square kilometer grid-based predicted values ​​for PM2.5 concentration (e.g., 50 micrograms per cubic meter at a certain time), predicted O3 concentration (e.g., 120 micrograms per cubic meter at a certain time), and corresponding predicted ambient temperature (e.g., 35 degrees Celsius at a certain time) for Location A over the next 24 hours.

[0036] Next, based on historical ambient temperature data, air pollutant data, and population health outcome data (such as the total number of non-accidental deaths) at location A, the system constructed exposure-response models for the relationship between ambient temperature and health outcomes, and between each air pollutant and health outcomes. For example, through statistical analysis, the system determined that for every 10 micrograms per cubic meter increase in PM2.5 concentration, the risk of non-accidental death in the population increases by X%; and for every 10 micrograms per cubic meter increase in O3 concentration, the risk of non-accidental death in the population increases by Y%. Simultaneously, the system also quantified the independent excess risk contribution of 35 degrees Celsius high temperature to population health outcomes. These models considered the nonlinear and lag effects of temperature, as well as the influence of other confounding factors, thus accurately quantifying the independent excess health risk contribution of each environmental factor.

[0037] Then, based on these independent excess health risk contributions, the system integrates the independent excess health risk contributions of ambient temperature with those of each air pollutant using an additive model. For example, if the independent excess health risk contribution of PM2.5 is 2%, O3 is 3%, HCHO is 1%, NO2 is 0.5%, and the independent excess health risk contribution of 35 degrees Celsius is 4%, then the composite health risk value for location A during that time period is calculated using the additive model as 2% + 3% + 1% + 0.5% + 4% = 10.5%. This composite health risk value comprehensively characterizes the combined exposure effect under the combined influence of temperature and multiple pollutants.

[0038] Finally, the system dynamically adjusts the risk level threshold based on the statistical distribution of the composite health risk value and the environmental temperature distribution characteristics of the geographical and climatic zone to which location A belongs. For example, if location A is located in a hot southern region during the summer, the system will set a risk level threshold applicable to that climatic zone based on the historical summer temperature distribution of the region and the population's adaptability to high temperatures. If 10.5% of the composite health risk value falls into the "high-risk" range set for that zone (e.g., 8%-15% is high risk), the system will determine that the atmospheric composite health risk classification warning index for location A during that period is "high-risk". This warning index can then be published through a public information platform to guide users A and other members of the public to take appropriate protective measures.

[0039] The aforementioned technical solution significantly improves the comprehensiveness, accuracy, and regional adaptability of atmospheric composite health risk early warning by incorporating ambient temperature as a core risk variable into the health risk assessment system and employing techniques such as multi-source data fusion, refined exposure-response relationship modeling, and dynamic grading threshold adjustment. Compared to the existing Air Quality Health Index (AQHI), which only focuses on air pollutants and ignores the independent health effects of ambient temperature, the method in this application can more comprehensively reflect the comprehensive health burden faced by the population in actual complex atmospheric environments. For example, in the above example, if only the risks of PM2.5 and O3 are considered, a conclusion of "moderate risk" may only be drawn, ignoring the significant health risks brought by high temperatures of 35 degrees Celsius. This application integrates the independent excess health risk contribution of temperature with pollutant risk, enabling the composite health risk value to more realistically reflect the actual situation and avoiding the problem of insufficient early warning in traditional AQHI when facing the superposition of extreme temperature events and high pollution events.

[0040] Furthermore, this application introduces a mechanism for dynamically adjusting risk level thresholds based on differences in environmental temperature distribution characteristics across different geographical and climatic zones, addressing the problem of a "one-size-fits-all" approach in traditional early warning indices. In the example above, if location A is in a northern region, 35 degrees Celsius might be considered extreme heat, potentially contributing more to the risk. Alternatively, the population's adaptability to high temperatures may vary, necessitating corresponding adjustments to the risk level threshold. This application can dynamically adjust the risk level threshold based on the specific climatic zone characteristics of location A, thereby ensuring the applicability and accuracy of early warning information across different regions. This avoids false alarms or omissions due to regional differences, enhancing the credibility and guiding value of early warnings.

[0041] In summary, the method of this application provides a more scientific, accurate, and practical composite health risk early warning solution by systematically integrating the health effects of ambient temperature and various air pollutants, and combining it with a dynamic and regionalized risk classification strategy. This makes an important technical contribution to public health management and individual health protection.

[0042] In some embodiments, acquiring near-surface concentration data of various air pollutants and ambient temperature data that spatiotemporally matches the near-surface concentration data through multi-source data fusion includes: collecting heterogeneous data sources, which include satellite remote sensing data, pollutant concentration data measured by ground monitoring stations, meteorological reanalysis data, and digital elevation model data, wherein the satellite remote sensing data includes pollutant column concentration data retrieved from satellite remote sensing; performing time alignment, spatial interpolation, unit standardization, and missing value imputation on the heterogeneous data sources; determining an inversion model from the satellite remote sensing data and the meteorological reanalysis data to the pollutant concentration data based on a machine learning model and according to the nonlinear relationship between the satellite remote sensing data, the meteorological reanalysis data, and the pollutant concentration data; processing the satellite remote sensing data and meteorological reanalysis data covering the entire target area using the inversion model, and outputting high spatiotemporal resolution gridded near-surface concentration data of the target area; and extracting the ambient temperature data corresponding to the grid from the meteorological reanalysis data.

[0043] As a specific implementation method, to obtain near-surface concentration data of various air pollutants such as PM2.5, O3, and NO2, as well as ambient temperature data for a target area (e.g., North X region), data from different sources can be collected first. Pollutant concentration data measured at ground monitoring stations can be obtained from real-time monitoring data released by the National Environmental Monitoring Centre. Meteorological reanalysis data can be obtained from the ERA5 dataset of the European Centre for Medium-Range Weather Forecasts (ECMWF), from which meteorological elements such as boundary layer height, relative humidity, wind speed, and surface temperature can be acquired. Digital elevation model data can be obtained using SRTM (ShuttleRadar Topography Mission) data. During the data preprocessing stage, the temporal resolution of all data can be unified to hours, and the spatial resolution to a 1km x 1km grid. For ground station data, methods such as Kriging interpolation or Geographically Weighted Regression (GWR) can be used to interpolate it to a 1km grid. For missing values, time-series-based imputation methods (such as linear interpolation or spline interpolation) or machine learning-based imputation methods (such as random forest imputation) can be used. When constructing the inversion model, gradient boosting decision tree (GBDT) or deep learning models (such as convolutional neural networks (CNN) combined with long short-term memory networks (LSTM)) can be used as machine learning models. This model uses satellite remote sensing data (such as AOD and NO2 column concentrations), meteorological reanalysis data (such as boundary layer height, relative humidity, wind speed, and surface temperature), and digital elevation model (DEM) data as input features, and near-surface pollutant concentrations interpolated from ground monitoring stations as training labels. Through training, the model learns the nonlinear mapping relationship between these input features and near-surface pollutant concentrations. After the model is trained, it is applied to satellite remote sensing data and meteorological reanalysis data covering the entire North China region, outputting gridded near-surface concentration data for PM2.5, O3, and NO2 at a 1km x 1km resolution and hourly scale covering the entire region. Simultaneously, corresponding 1km x 1km gridded, hourly-scale ambient temperature data is extracted from ERA5 meteorological reanalysis data to ensure accurate spatiotemporal matching with the pollutant concentration data.

[0044] Through the aforementioned technical solutions, this application effectively overcomes the limitations of single data sources in terms of spatial coverage, temporal resolution, and data quality. By acquiring and refining multi-source heterogeneous data, the integrity, consistency, and accuracy of the input data are ensured. In particular, by utilizing machine learning models to capture complex nonlinear relationships, accurate inversion from satellite remote sensing and meteorological reanalysis data to high spatiotemporal resolution near-surface pollutant concentrations is achieved, significantly improving data coverage and precision. Simultaneously, the extraction of spatiotemporally matched ambient temperature from meteorological reanalysis data ensures the synchronicity of exposure factors. This high-quality, comprehensive, and high-resolution near-surface pollutant concentration and ambient temperature data provides a solid and reliable foundation for subsequently constructing accurate exposure-response relationship models and predicting atmospheric composite health risk classification and early warning indices, significantly improving the accuracy and practicality of health risk assessment.

[0045] In some embodiments, the air pollutants include formaldehyde. When acquiring near-surface concentration data of multiple air pollutants and ambient temperature data that spatiotemporally matches the near-surface concentration data through multi-source data fusion, the method further includes the following steps: extracting the vertical column concentration of formaldehyde in the troposphere from the satellite remote sensing data; simulating the concentration distribution difference of formaldehyde in the near-surface layer and the upper atmosphere using a chemical transport model; determining the average proportion of near-surface formaldehyde concentration to the total tropospheric column concentration based on the concentration distribution difference, as a conversion factor; and obtaining near-surface formaldehyde exposure concentration data for health risk assessment based on the vertical column concentration of formaldehyde and the conversion factor.

[0046] Extracting the vertical column concentration of formaldehyde in the troposphere from the satellite remote sensing data involves using specific sensors onboard the satellite to observe the atmosphere and then converting the spectral signals received by the sensors into the total formaldehyde content in the atmosphere using an inversion algorithm. This extraction method provides large-scale, high-frequency formaldehyde distribution information, providing fundamental data for subsequent near-surface concentration estimation. Simulating the difference in formaldehyde concentration distribution between the near-surface and upper atmosphere using chemical transport models involves inputting data such as meteorological conditions, emission source information, and chemical reaction mechanisms, and then using numerical simulation tools to describe the generation, transport, transformation, and removal processes of chemical substances in the atmosphere. This simulates the concentration distribution of formaldehyde at different altitudes, revealing the difference in formaldehyde concentration between the near-surface and upper atmosphere. For example, three-dimensional chemical transport models (such as GEOS-Chem and WRF-Chem) can be used to simulate the vertical profile of formaldehyde, or one-dimensional boundary layer models can be used to finely characterize the vertical gradient of the near-surface layer. Based on the aforementioned concentration distribution differences, the average proportion of near-surface formaldehyde concentration to the total tropospheric column concentration is determined as a conversion factor. This refers to calculating the average ratio of near-surface concentration to vertical column concentration output by the model within a specific region and time period through statistical analysis of the chemical transport model simulation results. Alternatively, this proportional relationship can be established by combining ground observation data and vertical sensing data using statistical regression methods. Obtaining near-surface formaldehyde exposure concentration data for health risk assessment based on the vertical column concentration of formaldehyde and the conversion factor involves multiplying the formaldehyde vertical column concentration extracted by satellite remote sensing by the determined conversion factor, thereby transforming the column concentration information into a more representative near-surface exposure concentration.

[0047] As a specific implementation method, when acquiring near-surface formaldehyde exposure concentration data, the Tropospheric-Tropospheric Monitor (TROPOMI) aboard the ESA's Sentinel-5P satellite can be used to obtain vertical column concentration data of formaldehyde. This TROPOMI sensor can provide high spatial resolution global formaldehyde column concentration observations. Simultaneously, a global three-dimensional chemical transport model, such as the WRF-Chem model developed by the National Center for Atmospheric Research (NCAR), can be used, combined with actual meteorological data and emission inventories, to simulate the concentration distribution of formaldehyde at different altitudes (e.g., 0-100 meters for the near-surface layer, and above 100 meters for the upper atmosphere) within the target area. By analyzing the simulation results of the WRF-Chem model, the average ratio of near-surface formaldehyde concentration to the overall tropospheric column concentration can be calculated under different seasons and geographical regions, serving as a conversion factor. For example, in urban areas during summer, this conversion factor might be 0.6-0.8, while in rural areas during winter, it might be 0.4-0.6. Finally, by multiplying the formaldehyde vertical column concentration observed by TROPOMI by the corresponding conversion factor, the near-ground exposure concentration data of formaldehyde for health risk assessment can be obtained.

[0048] By employing the aforementioned technical solution, this application effectively addresses the inherent limitations of satellite remote sensing data in terms of vertical resolution, particularly in acquiring near-surface concentration data for pollutants with complex vertical distribution characteristics, such as formaldehyde. This solution, through the introduction of chemical transport model simulation and conversion factor calculation, transforms satellite-observed tropospheric column concentration data into near-surface concentration data that more closely reflects actual human exposure, significantly improving the accuracy and reliability of near-surface formaldehyde exposure concentration data. This enables subsequent models based on this data—constructing exposure-response relationships between environmental temperature and various air pollutants and human health outcomes, as well as determining composite health risk values—to more accurately reflect the actual impact of formaldehyde on human health, thereby enhancing the overall accuracy and practicality of the entire atmospheric composite health risk classification and early warning index prediction method.

[0049] In some implementations, constructing an exposure-response model of ambient temperature and various air pollutants with population health outcomes includes: for each grid cell, constructing a temperature-hysteresis response surface using cross-basis functions to simultaneously characterize the nonlinear immediate effect of ambient temperature and its sustained effect over a preset lag period; incorporating the temperature-hysteresis response surface as the core independent variable into a generalized linear model, controlling for other confounding factors that vary over time, and fitting the correlation parameters between ambient temperature and health outcomes; determining the estimated cumulative effect of ambient temperature on health outcomes over the preset lag period based on the correlation parameters; and merging the estimated cumulative effects of each grid cell through meta-analysis to obtain a unified curve and parameters describing the exposure-response relationship between ambient temperature and population health outcomes.

[0050] The use of cross-basis functions to construct a temperature-hysteresis response surface that simultaneously characterizes the nonlinear immediate effect and the sustained effect over a predetermined lag period of ambient temperature refers to the ability of cross-basis functions, as a statistical modeling tool, to capture both the nonlinear effect of ambient temperature and its time-lag effect. By cross-combining temperature and lag time variables, a two-dimensional basis function matrix can be generated, thus mathematically constructing the temperature-hysteresis response surface. This surface visually demonstrates the strength and direction of the influence of ambient temperature on health outcomes at different temperatures and at different lag time points. For example, natural cubic splines can be used to parameterize the nonlinear effects of temperature and lag time separately, and then the two spline basis functions can be multiplied and combined to form the cross-basis function. Alternatively, polynomial splines or B-splines can be used to construct the basis functions, and the cross-modeling of temperature and lag time can also be achieved through multiplication. Nonlinear immediate effect refers to the nonlinear impact of ambient temperature on health outcomes at the current moment, while persistent effect within a preset lag period refers to the impact of ambient temperature on health outcomes at the current moment that will persist for a period of time in the future (i.e., the preset lag period).

[0051] Using the temperature-lag response surface as the core independent variable and incorporating it into a generalized linear model (GLM) to control for other time-varying confounding factors, the association parameters between ambient temperature and health outcomes are obtained. This is because a GLM is a flexible statistical model capable of handling situations where the dependent variable follows a non-normal distribution; for example, population health outcome data are typically count data, which can be handled using Poisson regression or negative binomial regression. The basis function matrix corresponding to the temperature-lag response surface constructed from cross-basis functions is considered the core independent variable, containing all information about ambient temperature and its lag effects. Time-varying confounding factors refer to those factors that change over time and simultaneously affect both ambient temperature and population health outcomes. For example, a time trend term can be included in the GLM, using natural cube spline functions to smoothly control for long-term time trends and seasonal variations, or the day-of-the-week effect can be included to control for weekly periodic variations in health outcomes. The association parameters are coefficients obtained after fitting the GLM; they quantify the strength and direction of the statistical association between the core independent variable and the health outcome.

[0052] The cumulative effect estimate of environmental temperature on health outcomes within a preset lag period, determined based on the aforementioned correlation parameters, refers to the total impact of environmental temperature on health outcomes over the entire preset lag period. Since the cross-basis function model can provide the effect of temperature at each lag time point, the cumulative effect estimate can be obtained by integrating or summing these effects over the preset lag period. For example, the cumulative excess risk of health outcomes relative to a reference temperature within a lag period of 0 to 21 days can be calculated at a specific temperature.

[0053] Meta-analysis, by combining the cumulative effect estimates from various grid cells, yields a unified curve and parameters describing the exposure-response relationship between ambient temperature and population health outcomes. Meta-analysis is a statistical method used to systematically combine results from multiple independent studies or datasets to obtain a more accurate and robust overall estimate. Since each grid cell may independently calculate the cumulative effect estimate of ambient temperature on health outcomes, meta-analysis combines these estimates from different grid cells to obtain a representative and unified exposure-response relationship over a larger geographical area. For example, fixed-effects models or random-effects models can be used for merging. After meta-analysis, a curve representing the relationship between ambient temperature and population health outcomes across the entire target area, along with statistical parameters describing the characteristics of this curve, is obtained.

[0054] The following is a concrete example. As a specific implementation method, when constructing an exposure-response model of the relationship between ambient temperature and population health outcomes, a temperature-hysteresis response surface can first be built for each grid cell using cross-basis functions. For example, natural cubic spline functions can be used to parameterize the nonlinear effects of ambient temperature (e.g., daily average temperature) and hysteresis time (e.g., 0-21 days) respectively. Specifically, the temperature dimension can be set with 3 internal nodes and 2 boundary nodes, and the hysteresis dimension can be set with 3 internal nodes. Multiplying and combining these two spline basis functions forms a cross-basis function, which is used to simultaneously characterize the nonlinear immediate effect of temperature and its sustained effect over a preset hysteresis period. Subsequently, the basis matrix generated by this cross-basis function is used as the core independent variable and incorporated into the generalized linear model. For example, when the health outcome data is the daily number of hospital admissions due to cardiovascular disease, a Poisson regression model can be used. In the model, in addition to the core independent variables, other confounding factors with temporal variations can be controlled for, such as by including dummy variables representing the days of the week and using natural cube spline functions to smoothly control for long-term time trends and seasonal variations. By fitting this generalized linear model, the association parameters between ambient temperature and health outcomes can be obtained. Based on these association parameters, the cumulative effect estimate of ambient temperature on the risk of hospitalization for cardiovascular disease can be calculated within a pre-defined 0-21 day lag period. Finally, to obtain a uniform exposure-response relationship, a meta-analysis can be performed on the cumulative effect estimates from different grid cells. For example, a random-effects model can be used to weighted average the estimates from each grid cell, thereby obtaining a uniform exposure-response curve describing the relationship between ambient temperature and the risk of hospitalization for cardiovascular disease, along with the corresponding statistical parameters.

[0055] Through the aforementioned technical solutions, this application can comprehensively and meticulously capture the complex impact patterns of environmental temperature on human health. By constructing a temperature-hysteresis response surface using cross-basis functions, it can simultaneously characterize the nonlinear immediate effect of temperature and its sustained effect within a preset hysteresis period. This avoids the complex time delays and nonlinear relationships that traditional models may overlook, making the assessment of temperature effects more accurate. Incorporating this surface into a generalized linear model and controlling for other confounding factors of time variation further improves the accuracy of model fitting, ensuring that the obtained correlation parameters truly reflect the independent contribution of environmental temperature. Furthermore, by merging the cumulative effect estimates of each grid cell through meta-analysis, it effectively integrates data from different regions, overcoming the limitations that may exist with data from a single region, thus obtaining a more representative and robust unified exposure-response relationship curve and parameters. This not only significantly improves the accuracy and reliability of the exposure-response relationship model between environmental temperature and human health outcomes but also provides a more solid and refined foundation for the subsequent calculation of composite health risk values, thereby improving the prediction accuracy and practicality of the atmospheric composite health risk classification and early warning index.

[0056] In some implementations, after determining the atmospheric composite health risk classification and early warning index for each grid unit, the method further includes: dividing the target area into multiple climate zones with different temperature benchmarks and ranges of variation based on long-term climate statistical characteristics; for each climate zone, determining the temperature value or temperature range corresponding to the lowest health risk within the climate zone based on historical temperature data and population health data; establishing a set of zoning thresholds that maps the composite health risk value to different risk levels based on the temperature risk benchmark and overall temperature distribution of each climate zone; and classifying the composite health risk value according to the climate zone to which each grid unit belongs by calling the set of zoning thresholds of the climate zone.

[0057] Based on long-term climate statistics, the target area is divided into multiple climate zones with different temperature baselines and ranges of variation. This aims to identify the heterogeneity of climate patterns within the geographical region, particularly temperature-related characteristics, providing a more refined regional division for subsequent risk assessment. This can be achieved by using long-term temperature observation data from historical meteorological stations, combined with the spatial analysis capabilities of a Geographic Information System (GIS), and delineating climate zones through clustering algorithms (such as K-means and hierarchical clustering) or expert experience. For example, divisions can be based on indicators such as annual average temperature, seasonal temperature variation, and frequency of extreme temperature events. Alternatively, long-term climate simulation data output from global or regional climate models can be used, combined with geographical factors such as topography and latitude, to classify the target area using machine learning methods (such as support vector machines and decision trees), forming sub-regions with similar climate characteristics.

[0058] For each climate zone, based on historical temperature data and population health data, the temperature value or temperature range corresponding to the lowest health risk within that climate zone is determined. Its purpose is to identify the "comfortable" temperature conditions with the lowest population health risk within each specific climate zone, serving as a benchmark for risk assessment in that region. This can be achieved by constructing an exposure-response curve of temperature versus population health outcomes (e.g., mortality, morbidity) within each climate zone, identifying the temperature value or temperature range corresponding to the lowest health risk point on the curve. This typically involves using statistical methods such as distributed lag nonlinear models (DLNM) to capture the nonlinearity and lag effects of temperature. Alternatively, it can be achieved by analyzing the statistical distribution of population health outcomes under different temperature conditions in historical data, identifying temperature ranges with significantly lower-than-average incidence rates of health outcomes, and defining these as the temperature range corresponding to the lowest health risk.

[0059] Based on the temperature risk benchmarks and overall temperature distribution of each climate zone, a set of zoned thresholds is established to map the composite health risk value to different risk levels. This aims to customize the risk level classification criteria according to the specific temperature risk benchmarks and temperature distribution characteristics of each climate zone, ensuring the regional adaptability of risk assessment. This can be done by extending from the lowest health risk temperature value of each climate zone towards high and low temperatures, and combining the quantiles (e.g., 25%, 50%, 75%, 90%) of the zone's historical temperature data to set the threshold for the composite health risk value. For example, a composite health risk value higher than a certain percentile can be defined as high risk. Alternatively, an expert system or decision tree model can be used to dynamically generate a set of risk level thresholds applicable to each climate zone, using factors such as the temperature risk benchmark, temperature variability, and population temperature sensitivity as inputs, to classify the composite health risk value into different levels such as low, medium, high, and very high.

[0060] Based on the climate zone to which each grid cell belongs, the system uses the zoning threshold set of that climate zone to classify the composite health risk value. This compares the composite health risk value of each grid cell with the customized thresholds of its respective climate zone, thus obtaining a more region-specific risk level. The system can first determine the geographic coordinates of each grid cell, then query a pre-defined climate zone layer to determine which climate zone the grid cell belongs to. Next, it extracts the threshold set for the corresponding climate zone from the stored zoning threshold set database and uses this set to determine the level of the grid cell's composite health risk value. Alternatively, a lookup table or conditional logic can be implemented programmatically. After calculating the composite health risk value of a grid cell, the program automatically matches it to its corresponding climate zone based on the grid cell's geographic location information and applies the zone's unique risk level threshold set to output the final risk level.

[0061] The following is a concrete example. Assume the target region is North China, a region with complex climate characteristics and significant seasonal and regional differences. First, based on long-term meteorological observation data and geographical information of North China, the region can be divided into several climate zones, such as the "Northern Mountain Climate Zone," the "Central Plain Climate Zone," and the "Southern Transitional Climate Zone." For example, the northern mountainous area may have a lower average annual temperature and longer winters; the central plains have hot summers and cold winters; and the southern transitional zone falls somewhere in between. For the "Central Plain Climate Zone," by analyzing its historical temperature data and respiratory disease incidence data, it can be determined that the temperature range corresponding to the lowest health risk within this zone is likely between 18°C ​​and 25°C. Subsequently, based on this temperature risk benchmark for the "Central Plain Climate Zone" and the historical distribution of composite health risk values ​​within this zone, a set of risk level thresholds specifically for this zone can be established. For example, a composite health risk value above a certain percentile can be defined as "moderate risk," and above a higher percentile can be defined as "high risk." When a grid cell located in Shijiazhuang (belonging to the Central Plains Climate Zone) calculates its composite health risk value X, the system will call the threshold set of the "Central Plains Climate Zone" to determine the risk level corresponding to X, instead of using a uniform threshold applicable to the entire North China region or even the whole country.

[0062] Through the aforementioned technical solution, the method of this application can dynamically adjust the risk level threshold according to the specific characteristics of different geographical and climatic zones, thereby making the determined atmospheric composite health risk classification and early warning index more regionally specific and accurate. This avoids potential misjudgments or insufficient early warnings that may result from using a uniform threshold, enabling early warning information to more accurately reflect the actual health risks in local areas. This, in turn, provides local residents with more targeted health protection recommendations, effectively improving the precision of health risk assessment and the practical value of early warnings.

[0063] In some implementations, before determining the atmospheric composite health risk classification and early warning index for each grid unit, the method further includes: forming multiple candidate classification schemes that divide risks into the same number of levels but with different thresholds by setting different combinations of composite health risk value thresholds; statistically analyzing the frequency of occurrence and spatial coverage of each risk level under each candidate classification scheme; determining the statistical correlation strength between the increase in risk level and the increase in actual health damage to the population under each candidate classification scheme; and determining, based on the frequency of occurrence and spatial coverage of each risk level, as well as the statistical correlation strength, a candidate classification scheme that achieves a balance between the rationality of risk distribution and the sensitivity of health indicators, and setting it as the risk level threshold.

[0064] This involves setting different combinations of composite health risk thresholds to create multiple candidate grading schemes that classify risks into the same number of levels but with different thresholds. This means constructing a set of thresholds to discretize continuous composite health risk values ​​into different risk levels based on historical distribution characteristics or expert experience. These threshold sets can be set based on different quantiles (e.g., 25%, 50%, 75% quantiles) or specific health effect thresholds, thus creating multiple potential risk level classification methods. The frequency and spatial coverage of each risk level under each candidate grading scheme are statistically analyzed to evaluate the performance of different threshold schemes in both time and space. Frequency statistics quantify the number of times or duration of each risk level within a given time period; spatial coverage measures the geographical distribution range or area occupied by each risk level. Determining the statistical strength of the association between increased risk level and increased actual health damage in the population under each candidate grading scheme verifies whether the set risk levels have actual health significance. This typically involves using epidemiological methods to correlate different risk levels with observed population health outcome data (such as disease incidence and number of outpatient visits) to quantify the incremental impact of increased risk levels on health damage. Based on the frequency and spatial coverage of each risk level, as well as the strength of the statistical correlation, candidate grading schemes that balance the rationality of risk distribution and the sensitivity of health indicators are identified and determined as risk level thresholds. This involves comprehensively considering the above assessment results and selecting the optimal threshold scheme. Rationality of risk distribution requires that each risk level be representative and interpretable in both time and space, avoiding situations where a certain level is too frequent or too rare; sensitivity of health indicators requires that an increase in risk level significantly and stably corresponds to an increase in population health damage. By balancing these two factors, the final risk level threshold for atmospheric composite health risk grading and early warning is determined.

[0065] The following is a specific example to illustrate this. Before determining the atmospheric composite health risk classification and early warning index for each grid unit, several candidate classification schemes can be set first. For example, Scheme 1 can be set to divide the composite health risk value into three levels: low, medium, and high, with thresholds set at the 30th and 70th percentiles of the historical composite health risk value distribution, respectively; Scheme 2 can divide the risk into four levels: low, medium, high, and extremely high, with thresholds set at the 20th, 50th, and 80th percentiles, respectively; Scheme 3 may be based on specific pollutant concentration or temperature thresholds, and the corresponding thresholds for composite health risk values ​​can be derived by back-calculating from a model. Next, historical data back-analysis is performed on each scheme to statistically analyze the frequency of each risk level in the past year (e.g., low risk level occurred 60% of the time, medium risk level occurred 30% of the time, and high risk level occurred 10% of the time) and its average spatial coverage ratio in the target area. Simultaneously, using historical population health outcome data (such as the daily number of outpatient visits for respiratory diseases), a generalized linear regression model was constructed for each candidate risk level scheme, with risk level as the independent variable and the number of outpatient visits as the dependent variable, controlling for confounding factors such as time trends and day-of-week effects. The health indicator sensitivity was assessed by comparing the percentage increase in outpatient visits corresponding to each risk level increase under different schemes and their statistical significance. For example, if Scheme 2 shows an average increase of 15% in outpatient visits when the risk level rises from "medium" to "high," with high statistical significance, while Scheme 1 shows a smaller or insignificant increase, then Scheme 2 performs better in terms of health indicator sensitivity. Finally, considering the reasonableness of the risk distribution of each scheme (e.g., avoiding a certain level being too rare or too common) and the health indicator sensitivity, an optimal scheme was selected as the final risk level threshold.

[0066] Through the aforementioned technical solution, the risk level thresholds of the atmospheric composite health risk grading and early warning index are no longer empirical or single-dimensional, but rather derived through a systematic, data-driven optimization process. This makes the risk level classification more scientific and reliable, and can more accurately reflect the actual health risk situation. Simultaneously, because it fully considers the statistical correlation strength between risk levels and actual health damage to the population, this early warning index can provide more targeted and effective information when guiding the public to take protective measures, thereby enhancing the practical value and decision support capabilities of the early warning system.

[0067] In some implementations, determining the statistical association strength between an increase in risk level and an increase in actual health damage to the population under each candidate grading scheme includes: constructing a generalized linear regression model that controls for time trends and day-of-week effects, using the daily number of population health outcomes as the dependent variable and the risk level determined by the candidate grading scheme as the core independent variable; extracting the effect estimate corresponding to the risk level variable from the generalized linear regression model, wherein the effect estimate is used to quantify the percentage increase in the risk of health outcomes associated with each increase in risk level; and comparing the magnitude and statistical significance of the effect estimates corresponding to different candidate grading schemes to screen for the statistical association strength.

[0068] Specifically, using the daily number of population health outcomes as the dependent variable means using the statistical data of actual daily population health events (e.g., the number of people seeking medical attention for respiratory diseases, the number of people hospitalized for cardiovascular diseases, etc.) as the variable to be explained in the model. This dependent variable is usually count data, reflecting the frequency of occurrence of specific health impairments. Using the risk level determined by the candidate grading scheme as the core independent variable means using the risk level (e.g., low risk, medium risk, high risk, very high risk, etc.) divided by each candidate grading scheme as the key predictor in the model to explain changes in the dependent variable. This risk level can be an ordinal variable, representing an increase in risk level. Constructing a generalized linear regression model that controls for time trends and day-of-the-week effects means using a statistical model that can handle non-normally distributed dependent variables (such as count data), such as Poisson regression or negative binomial regression models. In this model, in addition to the core independent variable (risk level), time trends and day-of-the-week effects also need to be included as covariates for control. Time trends can be captured by introducing time variables (such as continuous dates or time spline functions) to eliminate their confounding effects on health outcomes. The day-of-the-week effect can be controlled by introducing categorical variables (such as dummy variables) representing different dates of the week to eliminate fluctuations in health outcomes caused by differences in weekday and weekend lifestyles. Extracting the effect estimate corresponding to the risk level variable from the generalized linear regression model refers to obtaining the regression coefficients of the risk level variable after model fitting. For models such as Poisson regression, these coefficients are usually converted to relative risk or incidence rate ratio through indexation, i.e., the effect estimate. The effect estimate quantifies the percentage increase in the risk of health outcomes associated with each increase in risk level. This means that the estimate intuitively represents the expected change in the number of health outcomes (such as morbidity or consultation rate) when the risk level rises from one level to the next. Comparing the magnitude and statistical significance of the effect estimates corresponding to different candidate grading schemes to screen for the strength of the statistical association involves performing the aforementioned modeling and effect estimation on multiple candidate grading schemes and then comparing their results. Comparing the magnitude of the effect estimates can determine which scheme has a stronger association between its risk level and health outcome, i.e., which scheme's changes in risk level more sensitively reflect changes in health damage. Simultaneously, comparing statistical significance (usually assessed through p-values ​​or confidence intervals) can determine whether this association is statistically significant, thereby ensuring the reliability of the selected scheme.

[0069] The following is a concrete example to illustrate this. Suppose we need to evaluate two candidate risk classification schemes, A and B, to determine which scheme's risk classification is more strongly associated with the number of daily respiratory emergency room visits in a certain region. First, we collect data on the number of daily respiratory emergency room visits in the region over the past year (as the dependent variable), and daily atmospheric composite health risk level data obtained according to scheme A and scheme B respectively (as core independent variables). Then, for scheme A, we construct a generalized linear regression model, such as a Poisson regression model, where the dependent variable is the number of daily emergency room visits, and the core independent variable is the risk level of scheme A. Simultaneously, we include a natural cube spline function representing the date to control for time trends, and a dummy variable representing the day of the week to control for the day-of-week effect. After model fitting, we extract the effect estimate corresponding to the risk level variable of scheme A. For example, if each increase in risk level leads to a 5% increase in the incidence of emergency room visits (relative risk of 1.05), and this result is highly statistically significant (p-value much less than 0.05). Next, the same modeling process was repeated for Scheme B to obtain the effect estimate corresponding to the risk level variable. For example, if the incidence of emergency room visits increases by 3% for each level of risk increase (relative risk of 1.03), its statistical significance is relatively low (p-value close to 0.05). By comparing the effect estimates and statistical significance of the two schemes, it can be found that the association between the risk level of Scheme A and the number of emergency room visits for respiratory diseases is stronger and more reliable. Therefore, Scheme A can be selected as the better risk grading scheme, and its risk level threshold can be determined as the final warning index threshold.

[0070] Through the aforementioned technical solution, this application provides an objective and quantitative method to assess the health indicative sensitivity of different risk grading schemes. This method, by constructing a generalized linear regression model and controlling for confounding factors such as time trends and day-of-week effects, can accurately quantify the statistical association strength between increased risk levels and increased actual health damage in the population. This allows for the selection of the scheme with the strongest association and highest statistical significance with population health outcomes from multiple candidate grading schemes, thereby ensuring that the determined risk level thresholds of the atmospheric composite health risk grading and early warning index are more scientific and reliable. Ultimately, this helps improve the accuracy and effectiveness of the early warning index, enabling it to more sensitively reflect the impact of environmental risks on population health and provide more precise guidance for public health protection.

[0071] In some implementations, after determining the atmospheric composite health risk classification and early warning index for each grid cell, the process further includes: integrating historical and real-time satellite remote sensing data, high-resolution meteorological forecast field data, and geographic auxiliary data to form a multi-source input feature dataset with spatiotemporal dimensions; constructing a deep learning model with an attention mechanism at its core, which can simultaneously capture the long-range spatial dependencies and dynamic evolution patterns of input features in time; training the deep learning model using ground monitoring data as supervisory labels; based on the trained deep learning model, outputting gridded pollutant concentration forecast values ​​for a predetermined future time period according to the near-surface concentration data of the various air pollutants in the future time period and the ambient temperature data that spatiotemporally matches the near-surface concentration data; and determining the atmospheric composite health risk classification and early warning index for the future time period based on the pollutant concentration forecast values ​​and the corresponding meteorological temperature forecast data.

[0072] This process integrates historical and real-time satellite remote sensing data, high-resolution meteorological forecast field data, and geographic auxiliary data to form a multi-source input feature dataset with spatiotemporal dimensions. The aim is to provide comprehensive, multi-dimensional, and spatiotemporally continuous input information for prediction models, capturing various complex factors affecting pollutant concentrations and ambient temperature. This step involves acquiring data from different data source interfaces (e.g., APIs, databases) through a data acquisition module, performing preliminary cleaning, format conversion, and storage to ensure data consistency and usability. Furthermore, data fusion techniques can be used to register and interpolate data of different resolutions and types, forming a unified spatiotemporal grid data. For example, satellite remote sensing imagery can be fused with ground station data to compensate for the limitations of a single data source.

[0073] The goal of constructing a deep learning model centered on an attention mechanism is to enable the model to handle complex nonlinear relationships and long-sequence data, thereby more accurately capturing the spatiotemporal dynamics of pollutant concentration and ambient temperature. The attention mechanism allows the model to dynamically focus on the most important parts when processing input data, improving its ability to extract key features. This model can employ a combination of convolutional neural networks (CNNs) and recurrent neural networks (RNNs), introducing an attention layer in its sequence processing part, such as using ConvLSTM combined with an attention mechanism to simultaneously handle spatial and temporal features. Alternatively, a graph neural network (GNN) combined with an attention mechanism can be used, treating the geographic grid as a graph structure and capturing spatial dependencies through a graph attention network (GAT).

[0074] The deep learning model described here can simultaneously capture both long-range spatial dependencies and dynamic evolution patterns of input features over time. This is a specific description of the capabilities of deep learning models, emphasizing their advantages in processing spatiotemporal data. Long-range spatial dependencies refer to the potential mutual influences between different geographical locations on pollutant concentrations or meteorological conditions, while dynamic evolution patterns over time refer to the regularity of these variables' changes over time. The model can learn and integrate information from distant spatial locations across different receptive fields using multi-scale convolutional kernels and pooling operations, combined with spatial attention mechanisms, thereby capturing long-range spatial dependencies. Simultaneously, the model can effectively model long-term dependencies and short-term fluctuations in time-series data using sequence models such as Gated Recurrent Units (GRUs) or Long Short-Term Memory Networks (LSTMs), combined with temporal attention mechanisms, thereby capturing dynamic evolution patterns.

[0075] Using ground monitoring data as supervised labels to train the deep learning model is a crucial step in enabling the model to learn the mapping relationship from input to output. Ground monitoring data is generally considered to be high-precision true values, and using it as supervised labels ensures that the model's learned predictions are as consistent as possible with the actual situation. The training process can use mean squared error (MSE) or mean absolute error (MAE) as the loss function, and iteratively adjust the model parameters through backpropagation algorithms and optimizers (e.g., Adam, SGD) until the model's performance on the validation set reaches expectations. Furthermore, the training process may also include data preprocessing (e.g., normalization), dataset partitioning (training set, validation set, test set), hyperparameter tuning (e.g., learning rate, batch size), and early stopping strategies to prevent overfitting and improve the model's generalization ability.

[0076] Based on the trained deep learning model, and using near-surface concentration data of various air pollutants for future periods, along with ambient temperature data that spatiotemporally matches the near-surface concentration data, the model outputs gridded pollutant concentration forecasts for predetermined future times. This is a crucial step in putting the model into practical application, utilizing learned knowledge to predict future pollutant concentrations. Specifically, future satellite remote sensing data, high-resolution meteorological forecast field data, and geographic auxiliary data can be used as inputs. The trained deep learning model performs forward propagation calculations to directly obtain gridded pollutant concentration predictions for different time steps. Alternatively, a rolling prediction approach can be used, predicting a short time window at a time and then using the prediction results as partial input for the next time window, progressively extrapolating to longer-term predictions.

[0077] Based on the predicted pollutant concentrations and corresponding meteorological temperature forecasts, the atmospheric composite health risk classification and early warning index for future periods is determined. This step combines the predicted pollutant concentrations and meteorological temperatures, and uses an established risk assessment model to calculate and determine the future composite health risk index. The predicted pollutant concentrations and meteorological temperatures can be used as inputs, substituted into the exposure-response relationship model and additive model constructed in claim 1, to calculate the composite health risk value for future periods, and then classified according to risk level thresholds. Alternatively, the prediction results can be directly input into a pre-trained risk classification model, which directly outputs the risk level for future periods, thereby achieving rapid risk early warning.

[0078] The following is a concrete example. When integrating multi-source input feature datasets, it can include NO2, SO2, and HCHO column concentration data from the Sentinel-5P satellite, temperature, humidity, wind speed, and boundary layer height data from the ERA5 reanalysis dataset provided by the European Centre for Medium-Range Weather Forecasts (ECMWF), and digital elevation model (DEM) data. When constructing a deep learning model centered on an attention mechanism, the Vision Transformer (ViT) architecture can be specifically adopted, treating the multi-channel spatiotemporal input feature dataset as an image sequence and capturing spatiotemporal dependencies through a multi-head self-attention mechanism. When training the deep learning model, measured concentration data of pollutants such as PM2.5, O3, and NO2 from national environmental monitoring stations can be used as supervisory labels. The Adam optimizer is employed with a learning rate of 0.001, a batch size of 32, and a training duration of 200 epochs. Based on a trained deep learning model, the system can output hourly gridded near-surface concentration forecasts for various air pollutants, including PM2.5, O3, NO2, SO2, CO, and HCHO, for the next 24, 48, or even 72 hours, with a resolution of 1 km x 1 km. Finally, these pollutant concentration forecasts are combined with future temperature forecasts provided by ECMWF and input into pre-established exposure-response relationship models and additive models to calculate the composite health risk value for the future period. Based on preset risk level thresholds (e.g., low, medium, high, and extremely high), a future risk warning map is generated.

[0079] This embodiment integrates multi-source spatiotemporal data and constructs a deep learning model centered on an attention mechanism. This application can effectively predict pollutant concentrations and ambient temperatures in future periods, thereby determining the future atmospheric composite health risk classification and early warning index. This overcomes the limitation of existing methods that can only assess current or historical risks, enabling risk warnings to be issued earlier and providing the public with more timely protective guidance. This achieves proactive management and early intervention of health risks, significantly improving the practicality and social benefits of the early warning system.

[0080] In some implementations, a deep learning model based on an attention mechanism is constructed, including: treating the multi-channel spatiotemporal input feature dataset as an image sequence, dividing it into an image patch sequence, and linearly projecting each image patch into an embedding vector; adding learnable temporal location encoding and two-dimensional geospatial location encoding to the embedding vector sequence; inputting the encoded embedding vector sequence into a stack of multi-layer Transformer encoders, with each layer extracting deep features through a multi-head self-attention mechanism and a feedforward neural network; and mapping the final context feature sequence back to a pollutant concentration grid map for each forecast time period with the same spatial resolution as the input.

[0081] Specifically, treating the multi-channel spatiotemporal input feature dataset as an image sequence and dividing it into image patch sequences is a way to structure the raw, high-dimensional spatiotemporal data, making it effective for the Transformer model. The multi-channel input feature dataset can contain different types of environmental data, such as the concentration of various air pollutants, ambient temperature, wind speed, and humidity, with each channel representing a feature. Treating it as an image sequence means that the data at each time step can be viewed as a multi-channel image. Dividing it into image patch sequences further subdivides these "images" into smaller, localized patches, which helps reduce computational complexity and allows the model to extract features locally. Each image patch is then transformed into an embedding vector through linear projection. This process maps the high-dimensional image patch data to a low-dimensional, semantically rich vector space, providing a unified representation for subsequent attention mechanisms. For example, linear projection can be implemented using a fully connected layer or a 1x1 convolutional layer.

[0082] The addition of learnable temporal and 2D geospatial location codes to the embedding vector sequence compensates for the Transformer model's inherent inability to process the positional information of elements within the sequence. Since the Transformer model processes all elements in the sequence in parallel, it cannot directly perceive the relative or absolute positions of data points in time and space. By introducing learnable location codes, the model can acquire the position of the original data corresponding to each embedding vector within the time series and its position within the geospatial grid. These location codes are trained along with the model parameters, thus better adapting to specific data distributions and patterns. For example, the temporal location code can be a vector associated with the time step, while the 2D geospatial location code can be a vector associated with the grid coordinates; these are added to the corresponding embedding vectors.

[0083] The encoded embedding vector sequence is input into a stack of multi-layer Transformer encoders. Each layer extracts deep features through a multi-head self-attention mechanism and a feedforward neural network, forming the core processing part of this deep learning model. The Transformer encoder stack performs multi-level abstraction and feature extraction on the input embedding vector sequence. The multi-head self-attention mechanism allows the model to simultaneously focus on all other elements in the sequence while processing each element, assigning different weights based on their correlations. This mechanism enables the model to capture long-range dependencies in the data, whether spatial (associations between different geographical locations) or temporal (evolution between different points in time). For example, the multi-head self-attention mechanism can decompose the attention calculation process into multiple independent "heads," each focusing on a different feature subspace, thus more comprehensively capturing complex relationships between data. The feedforward neural network then performs a non-linear transformation on the extracted features after the self-attention mechanism, further enhancing the model's expressive power.

[0084] The output stage of the model involves mapping the final contextual feature sequence back to a pollutant concentration grid map for each forecast time period, consistent with the input spatial resolution, using a lightweight decoder. After processing by multiple Transformer encoders, the model outputs a feature sequence containing rich contextual information. The decoder's role is to transform these abstract features back into concrete, interpretable prediction results. Since the encoder has already performed complex feature extraction, the decoder is designed to be "lightweight," meaning it primarily handles reconstructing the feature sequence into a spatial grid map, rather than performing additional complex feature learning. For example, the decoder can consist of a series of upsampling layers (such as transposed convolutional layers) and convolutional layers to progressively restore the spatial resolution and ultimately output a pollutant concentration grid map for future forecast times, consistent with the spatial resolution of the original input data.

[0085] The following is a concrete example to illustrate this. Suppose we need to predict the PM2.5 concentration grid map of a certain urban area for the next 24 hours. First, historical and real-time satellite remote sensing data, high-resolution meteorological forecast field data, and geographic auxiliary data (e.g., topography, land use type, etc.) are integrated into a multi-channel spatiotemporal input feature dataset. For example, this dataset could be a sequence containing multiple channels such as PM2.5, temperature, humidity, wind speed, and wind direction, covering 256x256 grid cells of the target area, and including data from the past 48 hours and forecast data for the next 24 hours. The 256x256 grid data at each time step is treated as a multi-channel image and divided into a sequence of 16x16 image patches. Each 16x16 image patch (containing data from all channels) is linearly projected through a fully connected layer into a 768-dimensional embedding vector. Next, learnable temporal location encodings (e.g., encodings for 48+24=72 time steps) and two-dimensional geospatial location encodings (e.g., encodings for 16x16=256 spatial locations) are added to each element in the embedding vector sequence. These encodings are added to the embedding vectors to provide the model with spatiotemporal context. The encoded embedding vector sequence is then fed into a stack of 12 Transformer encoder layers. Each encoder layer includes an 8-head multi-head self-attention mechanism to capture complex correlations between different time steps and spatial locations, and a feedforward neural network for feature transformation. Finally, a lightweight decoding head, for example consisting of two transposed convolutional layers and a 1x1 convolutional layer, reconstructs the contextual feature sequence output by the Transformer encoder into a PM2.5 concentration grid map for the next 24 forecast times, consistent with the original 256x256 spatial resolution.

[0086] Through the aforementioned technical solution, this application enables the construction of an efficient and accurate deep learning model for predicting pollutant concentrations in future periods. This model transforms spatiotemporal data into image sequences and image patches, and combines this with learnable positional encoding, allowing the Transformer encoder to fully utilize its powerful multi-head self-attention mechanism to effectively capture the complex long-range dependencies and dynamic evolution patterns of pollutant concentrations in time and space. This significantly improves the accuracy and robustness of the model's predictions of future pollutant concentrations, thus providing more reliable basic data for the subsequent determination of the atmospheric composite health risk classification and early warning index. Furthermore, the lightweight decoder head design ensures that the overall computational efficiency of the model is maintained while preserving high spatial resolution output.

[0087] In some implementations, after determining the atmospheric composite health risk classification and early warning index for each grid cell, the method further includes: based on the risk level data of the gridded atmospheric composite health risk classification and early warning index obtained through real-time calculation or forecast, rendering a risk spatial distribution map on a geographic base map using a preset color coding scheme; for any location in the risk spatial distribution map, analyzing the contribution ratio of each component in its composite health risk value, and identifying one or more components with the largest contribution as the dominant risk factor at that location; and publishing the risk spatial distribution map, the dominant factor, and the risk level through a public information platform.

[0088] The risk level data of the gridded atmospheric composite health risk classification and early warning index obtained through real-time calculation or forecasting refers to the level information of the atmospheric composite health risk classification and early warning index organized in the form of a geographic grid, obtained through real-time data processing or based on future prediction models. Real-time calculation can be achieved by continuously receiving and immediately processing the latest environmental data through sensor networks or data interfaces. For example, using current pollutant concentration and ambient temperature data, combined with a pre-built exposure-response model, the current risk level can be quickly calculated. Forecasting involves using prediction models to predict pollutant concentration and ambient temperature over a future period, thereby estimating the risk level for that period. This data exists in a gridded form, meaning the entire target area is divided into a series of regular grid units, each corresponding to a specific risk level value, facilitating spatial analysis and visualization.

[0089] Rendering a risk spatial distribution map using a preset color coding scheme on a geographic base map refers to presenting the aforementioned gridded risk level data on a geographic background map through visualization technology. The geographic base map can be a satellite image, administrative division map, or topographic map, providing geographic reference information. The preset color coding scheme assigns specific colors to different risk levels (e.g., low risk, medium risk, high risk); for example, green represents low risk, yellow represents medium risk, and red represents high risk. The process of rendering the risk spatial distribution map involves mapping the risk level of each grid cell to a corresponding color and filling or marking the corresponding grid area on the geographic base map with that color, thereby visually displaying the spatial distribution and regional differences of risk.

[0090] For any location on the risk spatial distribution map, analyzing the contribution ratio of each component in the composite health risk value and identifying the one or more components with the largest contribution as the dominant factors of risk at that location means that when calculating the composite health risk value, the value is integrated from the independent excess health risk contributions of ambient temperature and multiple air pollutants. Analyzing the contribution ratio quantifies the relative influence of these components on the final composite risk value. For example, the contribution ratio can be determined by calculating the percentage of each independent excess health risk contribution to the total composite health risk value. The dominant factor refers to one or more environmental factors (such as a specific pollutant or ambient temperature) that have the greatest impact on the composite health risk value at a specific location. Identifying the dominant factor can provide a clear direction for subsequent risk management and intervention measures. For example, if the risk in a certain area is mainly contributed by PM2.5, then the focus should be on reducing PM2.5 emissions.

[0091] Publishing the aforementioned risk spatial distribution map, dominant factors, and risk levels through a public information platform means disseminating the above visualization and analysis results through public information channels. Public information platforms can include various forms such as government official websites, mobile applications, social media, electronic displays, or radio and television broadcasts. "Publishing" means presenting this information to the target audience in an easily understandable and accessible manner, for example, providing interactive risk maps on websites, pushing regional risk warnings on apps, or publishing risk reports through news media. This aims to raise public awareness of atmospheric complex health risks and prompt relevant departments to take effective risk management and intervention measures.

[0092] As a specific implementation method, after determining the atmospheric composite health risk classification and early warning index for a certain urban area for the next 24 hours, the system can first load these gridded risk level data into a Geographic Information System (GIS) platform. This platform pre-loads the city's administrative division map and major road network map as a geographic base map. The system can set a color coding scheme; for example, risk level 1 (low risk) is coded as green, level 2 (low to medium risk) as light yellow, level 3 (medium risk) as orange, level 4 (medium to high risk) as red, and level 5 (high risk) as dark red. Subsequently, the system renders the corresponding color area on the geographic base map according to the risk level of each grid unit, thereby generating an intuitive spatial distribution map of the city's atmospheric composite health risk. For example, the city center may be displayed as red, while suburbs may be displayed as green or yellow. For red areas with higher risk, the system can further analyze which components (such as PM2.5, O3, or ambient temperature) contribute the most to their composite health risk value. For example, if it is found that PM2.5 contributes 60% to the combined risk in a certain area, O3 contributes 30%, and ambient temperature contributes 10%, then PM2.5 is identified as the dominant risk factor in that area. Finally, these spatial distribution maps of risk, information on dominant factors, and specific risk levels can be published in real time through the official website of the city's environmental monitoring center, mobile apps (such as the "City Air Quality Bulletin" app), and electronic displays at major urban transportation hubs. The public can use these platforms to view the risk level and dominant factors of their area; for example, clicking on a specific area on the app will display "This area is currently at high risk, mainly affected by PM2.5."

[0093] Through the aforementioned technical solution, this application can transform the abstract atmospheric composite health risk classification and early warning index into an intuitive and easy-to-understand risk spatial distribution map, greatly improving the transparency and comprehensibility of risk information. By analyzing the contribution ratio of each component in the composite health risk value, the dominant factors of risk in specific areas can be accurately identified, making risk management and public protection measures more targeted. Furthermore, the release of information through a public information platform ensures the timely and widespread dissemination of risk information, effectively improving public risk awareness and self-protection capabilities, and providing strong support for government departments to formulate refined environmental health management strategies, thereby significantly enhancing the practicality and social benefits of the entire early warning system.

[0094] In some implementations... Figure 1The method further includes: establishing a health behavior recommendation knowledge base for the general population and sensitive populations, encompassing multiple risk levels and different combinations of dominant factors; matching corresponding scenarios from the contextualized recommendation knowledge base based on the current or predicted risk level of the atmospheric composite health risk classification and warning index and its dominant factors, and generating text recommendations containing specific protective measures and behavioral guidelines; when the ambient temperature is identified as the dominant factor or an important contributing factor, strengthening the protective guidance related to temperature stress in the generated text recommendations, including heatstroke prevention and cooling measures for high temperatures or warmth protection measures for low temperatures.

[0095] As a specific implementation method, a rule-based health recommendation system can be constructed. This system maintains a health behavior recommendation knowledge base, which can be stored in the form of structured data (e.g., JSON or XML files) containing health recommendation texts corresponding to different risk levels, combinations of dominant factors, and population types (general population, sensitive population). For example, the knowledge base can define the following rule: when the risk level is "high" and the dominant factor includes "PM2.5," the recommendation for "sensitive population" is "strongly recommend avoiding outdoor activities, staying indoors, closing doors and windows, turning on a high-efficiency air purifier, and seeking medical attention if necessary." When the system receives a warning index for a region's atmospheric composite health risk level as "high risk," and the dominant factors are identified as "PM2.5" and "ambient temperature (high temperature)," the system will first match the basic recommendations corresponding to "high risk" and "PM2.5" from the knowledge base. Simultaneously, since ambient temperature (high temperature) is also identified as a dominant factor, the system triggers a temperature stress enhancement module. This module extracts or inserts heat stress-related protective guidelines from the knowledge base or into existing suggestions, such as "avoid going out during high-temperature periods, wear light and breathable clothing, use air conditioning or fans, replenish fluids and electrolytes promptly, and pay attention to your physical condition." Finally, the system combines these matched suggestions into a complete text and pushes it to users through public information platforms or mobile applications. In the pushed text, heat stress-related protective guidelines can be highlighted using bolding, highlighting, or pinning to the top to ensure users notice this crucial information immediately.

[0096] Through the aforementioned technical solution, this application can transform the abstract atmospheric composite health risk classification and early warning index into specific and actionable health protection measures and behavioral guidelines. By establishing a health behavior advice knowledge base and matching it according to risk levels and dominant factors, personalized and contextualized health advice can be provided to the public, effectively compensating for the shortcomings of simply issuing risk information without providing users with information on how to respond. Especially when ambient temperature is identified as a dominant or significant contributing factor, the enhanced protective guidance related to temperature stress ensures that users receive more targeted advice on heatstroke prevention and cooling or keeping warm under extreme temperature conditions, thereby effectively reducing the health risks associated with temperature stress. This significantly improves the practicality and guidance of the early warning information, enabling the public to take timely and effective protective actions based on their environmental risks and personal health conditions, thus more effectively protecting public health, especially for sensitive populations.

[0097] In some implementations... Figure 1 The method further includes: obtaining the user's real-time geographical location through a mobile terminal application and receiving the user's self-inputted or selected personal health attribute information; matching the user's real-time geographical location with the risk spatial distribution map and the dominant factors to obtain personalized environmental risk data of the user's location; interpreting and correcting the personalized environmental risk data based on the user's personal health attribute information to generate a customized risk warning and health guidance information that combines the user's personal health status, and pushing it to the user through the mobile terminal application.

[0098] As one specific implementation method, users can install the health risk warning mobile application provided in this application on their smartphones. When a user opens the application and authorizes access to their geographic location information, the application uses the phone's GPS module to obtain the user's current latitude and longitude coordinates. Simultaneously, users can manually enter their age, history of chronic respiratory diseases (such as bronchitis), and whether they are pregnant, among other personal health attributes, in the "My Health" module within the application. After receiving the user's real-time geographic location, the system backend compares it with a pre-calculated and stored national high-resolution atmospheric composite health risk spatial distribution map. For example, by querying the grid cell where the user's location is located, the system obtains the current risk level (e.g., "moderate risk") and dominant factors (e.g., "PM2.5 and high temperature") for that grid cell. Subsequently, based on the user's input of a "history of chronic respiratory diseases" health attribute, the system corrects the "moderate risk" to interpret it as a "higher risk" for the user. Next, the system will match a set of combined recommendations from a pre-set health behavior advice knowledge base for "higher risk," "PM2.5 and high temperatures," and "patients with chronic respiratory diseases." For example, "Today's PM2.5 concentration is high and the temperature is hot. It is recommended that patients with chronic respiratory diseases reduce outdoor activities, wear N95 masks when going out, and pay attention to replenishing water to avoid heatstroke." Finally, this customized risk warning and health guidance information will be pushed to the user's smartphone in real time via in-app notification or pop-up window.

[0099] Through the aforementioned technical solution, this application can transform generalized regional risk warning information into customized health guidance for individual users. By acquiring users' real-time geographical location and personal health attributes, the system can accurately assess the actual risks faced by users and provide protective recommendations highly matched to their health conditions. This significantly improves the accuracy and effectiveness of risk warnings, enabling users to take more scientific and reasonable protective measures based on their own circumstances, thereby minimizing the health risks from environmental exposure and effectively enhancing the level of refined management of public health protection.

[0100] To facilitate better implementation of the methods in the embodiments of this application, such as Figure 2 As shown in the figure, this application embodiment also provides an atmospheric composite health risk classification and early warning index prediction device, the device comprising: The acquisition module 201 is used to acquire near-ground concentration data of various air pollutants and ambient temperature data that is spatiotemporally matched with the near-ground concentration data through multi-source data fusion. The identification module 202 is used to construct exposure-response relationship models between ambient temperature and various air pollutants and population health outcomes based on the ambient temperature data and the population health outcome data of the corresponding area, so as to determine the independent excess health risk contribution of the ambient temperature and various air pollutants. The determination module 203 is used to integrate the independent excess health risk contribution of the ambient temperature with the independent excess health risk contribution of each air pollutant based on an additive model, according to each independent excess health risk contribution, to determine a composite health risk value for characterizing the temperature-pollutant combined exposure effect. The prediction module 204 is used to determine the atmospheric composite health risk classification and early warning index of each grid unit based on the statistical distribution of the composite health risk value and multiple risk level thresholds dynamically adjusted based on the differences in environmental temperature distribution characteristics of different geographical and climatic zones.

[0101] All of the above technical solutions can be combined in any way to form optional embodiments of this application, and will not be described in detail here.

[0102] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0103] The embodiments, implementation methods, and related technical features of this application can be combined and substituted for each other without conflict.

[0104] The above are merely preferred embodiments of this application and are not intended to limit this application in any way. Although the descriptions of each embodiment in this application have different focuses, and parts not described in detail in a certain embodiment can be referred to the relevant descriptions of other embodiments, any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of this application without departing from the content of the technical solution of this application shall still fall within the scope of the technical solution of this application.

Claims

1. A method for predicting an atmospheric composite health risk classification and early warning index, characterized in that, include: Near-ground concentration data of various air pollutants and ambient temperature data that are spatiotemporally matched with the near-ground concentration data are obtained through multi-source data fusion. Based on the ambient temperature data and the corresponding regional population health outcome data, exposure-response relationship models of ambient temperature and each air pollutant with population health outcomes are constructed to determine the independent excess health risk contribution of ambient temperature and each air pollutant. Based on the independent excess health risk contribution of each pollutant, the independent excess health risk contribution of ambient temperature is integrated with the independent excess health risk contribution of each air pollutant using an additive model to determine a composite health risk value for characterizing the temperature-pollutant combined exposure effect. Based on the statistical distribution of the composite health risk values ​​and the risk level thresholds dynamically adjusted according to the differences in environmental temperature distribution characteristics of different geographical and climatic zones, the atmospheric composite health risk classification and early warning index for each grid unit is determined.

2. The method according to claim 1, characterized in that, The method of acquiring near-ground concentration data of various air pollutants and ambient temperature data that spatiotemporally matches the near-ground concentration data through multi-source data fusion includes: Collect heterogeneous data sources, including satellite remote sensing data, pollutant concentration data measured by ground monitoring stations, meteorological reanalysis data, and digital elevation model data. The satellite remote sensing data includes pollutant column concentration data retrieved from satellite remote sensing. The heterogeneous data sources are subjected to time alignment, spatial interpolation, unit standardization, and missing value imputation. Based on a machine learning model, and according to the nonlinear relationship between the satellite remote sensing data, the meteorological reanalysis data, and the pollutant concentration data, an inversion model from the satellite remote sensing data and the meteorological reanalysis data to the pollutant concentration data is determined. The inversion model is used to process satellite remote sensing data and meteorological reanalysis data covering the entire target area, and outputs high spatiotemporal resolution gridded near-ground concentration data of the target area. The ambient temperature data corresponding to the grid is extracted from the meteorological reanalysis data.

3. The method according to claim 2, characterized in that, The air pollutants include formaldehyde. The method of obtaining near-ground concentration data of multiple air pollutants through multi-source data fusion, and the ambient temperature data that spatiotemporally matches the near-ground concentration data, further includes: Extract the vertical column concentration of formaldehyde in the troposphere from the satellite remote sensing data; A chemical transport model was used to simulate the difference in formaldehyde concentration distribution between the near-surface layer and the upper atmosphere; Based on the aforementioned concentration distribution differences, the average proportion of near-surface formaldehyde concentration to the total tropospheric column concentration is determined as a conversion factor. The formaldehyde near-ground exposure concentration data for health risk assessment are obtained based on the vertical column concentration of formaldehyde and the conversion factor.

4. The method according to claim 1, characterized in that, The constructed exposure-response model for environmental temperature, various air pollutants, and population health outcomes includes: For each grid cell, a temperature-hysteresis response surface is constructed using cross basis functions to simultaneously characterize the nonlinear instantaneous effect of ambient temperature and the sustained effect within a preset hysteresis period. The temperature-hysteresis response surface was used as the core independent variable and incorporated into the generalized linear model. Other confounding factors that vary over time were controlled, and the correlation parameters between ambient temperature and health outcomes were obtained by fitting the model. The cumulative effect of ambient temperature on health outcomes within a preset lag period is estimated based on the correlation parameters. By combining the cumulative effect estimates from each grid cell through meta-analysis, a unified curve and parameters describing the exposure-response relationship between ambient temperature and population health outcomes are obtained.

5. The method according to claim 1, characterized in that, Before determining the atmospheric composite health risk classification and early warning index for each grid unit based on the statistical distribution of the composite health risk value and multiple risk level thresholds dynamically adjusted according to the differences in environmental temperature distribution characteristics of different geographical and climatic zones, the method further includes: Based on long-term climate statistics, the target area is divided into multiple climate zones with different temperature baselines and ranges of variation. For each climate zone, based on historical temperature data and population health data, determine the temperature value or temperature range within the climate zone that corresponds to the lowest health risk; Based on the temperature risk benchmark and overall temperature distribution of each climate zone, a set of zoned thresholds is established to map the composite health risk value to different risk levels. Based on the climate zone to which each grid cell belongs, the composite health risk value is classified into levels by calling the set of zoning thresholds for the climate zone.

6. The method according to claim 1, characterized in that, After determining the atmospheric composite health risk classification and early warning index for each grid cell, the following is also included: By setting different combinations of the composite health risk value thresholds, multiple candidate classification schemes are formed that divide risks into the same number of levels but with different thresholds. The frequency of occurrence and the breadth of spatial coverage of each risk level under each candidate grading scheme are statistically analyzed. Determine the statistical association strength between the increased risk level and the actual increase in health damage to the population under each of the candidate grading schemes; Based on the frequency of occurrence and spatial coverage of each risk level, as well as the strength of the statistical correlation, a candidate classification scheme that achieves a balance between the rationality of risk distribution and the sensitivity of health indicators is determined and set as the risk level threshold.

7. The method according to claim 6, characterized in that, Determining the statistical association strength between increased risk level and increased actual health damage in the population under each candidate grading scheme includes: Using the number of daily population health outcomes as the dependent variable and the risk level determined by the candidate grading scheme as the core independent variable, a generalized linear regression model that controls for time trend and day of the week effect was constructed. From the generalized linear regression model, the effect estimate corresponding to the risk level variable is extracted. The effect estimate is used to quantify the percentage increase in the risk of health outcome associated with each increase in risk level. The magnitude and statistical significance of the effect estimates corresponding to different candidate grading schemes are compared to screen out the statistical association strength.

8. The method according to claim 1, characterized in that, After determining the atmospheric composite health risk classification and early warning index for each grid cell, the following is also included: By integrating historical and real-time satellite remote sensing data, high-resolution meteorological forecast field data, and geographic auxiliary data, a multi-source input feature dataset with spatiotemporal dimensions is constructed. Construct a deep learning model with attention mechanism at its core, which can simultaneously capture the long-range spatial dependencies and dynamic evolution patterns of input features over time. The deep learning model is trained using ground monitoring data as supervised labels. Based on the trained deep learning model, according to the near-ground concentration data of various air pollutants in the future time period and the ambient temperature data that is spatiotemporally matched with the near-ground concentration data, the gridded pollutant concentration forecast value for the future time period is output. Based on the pollutant concentration forecast values ​​and the corresponding meteorological temperature forecast data, the atmospheric composite health risk classification and early warning index for the future period is determined.

9. The method according to any one of claims 1 to 8, characterized in that, After determining the atmospheric composite health risk classification and early warning index for each grid cell, the following is also included: Based on the risk level data of the gridded atmospheric composite health risk classification and early warning index obtained in real time or by forecast, a risk spatial distribution map is generated on the geographic base map using a preset color coding scheme. For any location in the risk spatial distribution map, analyze the contribution ratio of each component in the composite health risk value, and identify one or more components with the largest contribution as the dominant factors of the risk at that location. The risk spatial distribution map, the dominant factors, and the risk levels will be published through a public information platform.

10. A device for predicting atmospheric composite health risk classification and early warning index, characterized in that, include: The acquisition module is used to acquire near-ground concentration data of various air pollutants and ambient temperature data that is spatiotemporally matched with the near-ground concentration data through multi-source data fusion. The identification module is used to construct exposure-response relationship models between ambient temperature and various air pollutants and population health outcomes based on the ambient temperature data and the corresponding regional population health outcome data, so as to determine the independent excess health risk contribution of ambient temperature and various air pollutants. The determination module is used to integrate the independent excess health risk contribution of the ambient temperature with the independent excess health risk contribution of each air pollutant based on an additive model, according to each independent excess health risk contribution, to determine a composite health risk value for characterizing the temperature-pollutant combined exposure effect. The prediction module is used to determine the atmospheric composite health risk classification and early warning index for each grid unit based on the statistical distribution of the composite health risk value and multiple risk level thresholds dynamically adjusted according to the differences in environmental temperature distribution characteristics of different geographical and climatic zones.

Images