Dataset Information

Ensemble-based deep learning for estimating PM_2.5 over California with multisource big data including wildfire smoke.

ABSTRACT:

Introduction

Estimating PM_2.5 concentrations and their prediction uncertainties at a high spatiotemporal resolution is important for air pollution health effect studies. This is particularly challenging for California, which has high variability in natural (e.g, wildfires, dust) and anthropogenic emissions, meteorology, topography (e.g. desert surfaces, mountains, snow cover) and land use.

Methods

Using ensemble-based deep learning with big data fused from multiple sources we developed a PM_2.5 prediction model with uncertainty estimates at a high spatial (1 km × 1 km) and temporal (weekly) resolution for a 10-year time span (2008-2017). We leveraged autoencoder-based full residual deep networks to model complex nonlinear interrelationships among PM_2.5 emission, transport and dispersion factors and other influential features. These included remote sensing data (MAIAC aerosol optical depth (AOD), normalized difference vegetation index, impervious surface), MERRA-2 GMI Replay Simulation (M2GMI) output, wildfire smoke plume dispersion, meteorology, land cover, traffic, elevation, and spatiotemporal trends (geo-coordinates, temporal basis functions, time index). As one of the primary predictors of interest with substantial missing data in California related to bright surfaces, cloud cover and other known interferences, missing MAIAC AOD observations were imputed and adjusted for relative humidity and vertical distribution. Wildfire smoke contribution to PM_2.5 was also calculated through HYSPLIT dispersion modeling of smoke emissions derived from MODIS fire radiative power using the Fire Energetics and Emissions Research version 1.0 model.

Results

Ensemble deep learning to predict PM_2.5 achieved an overall mean training RMSE of 1.54 μg/m³ (R²: 0.94) and test RMSE of 2.29 μg/m³ (R²: 0.87). The top predictors included M2GMI carbon monoxide mixing ratio in the bottom layer, temporal basis functions, spatial location, air temperature, MAIAC AOD, and PM_2.5 sea salt mass concentration. In an independent test using three long-term AQS sites and one short-term non-AQS site, our model achieved a high correlation (>0.8) and a low RMSE (<3 μg/m³). Statewide predictions indicated that our model can capture the spatial distribution and temporal peaks in wildfire-related PM_2.5. The coefficient of variation indicated highest uncertainty over deciduous and mixed forests and open water land covers.

Conclusion

Our method can be generalized to other regions, including those having a mix of major urban areas, deserts, intensive smoke events, snow cover and complex terrains, where PM_2.5 has previously been challenging to predict. Prediction uncertainty estimates can also inform further model development and measurement error evaluations in exposure and health studies.

SUBMITTER: Li L

PROVIDER: S-EPMC7643812 | biostudies-literature |

REPOSITORIES: biostudies-literature

ACCESS DATA

Similar Datasets

Project description:BackgroundSubstantial increases in wildfire activity have been recorded in recent decades. Wildfires influence the chemical composition and concentration of particulate matter ≤2.5 μm in aerodynamic diameter (PM2.5). However, relatively few epidemiologic studies focus on the health impacts of wildfire smoke PM2.5 compared with the number of studies focusing on total PM2.5 exposure.ObjectivesWe estimated the associations between cardiorespiratory acute events and exposure to smoke PM2.5 in Colorado using a novel exposure model to separate smoke PM2.5 from background ambient PM2.5 levels.MethodsWe obtained emergency department visits and hospitalizations for acute cardiorespiratory outcomes from Colorado for May-August 2011-2014, geocoded to a 4 km geographic grid. Combining ground measurements, chemical transport models, and remote sensing data, we estimated smoke PM2.5 and non-smoke PM2.5 on a 1 km spatial grid and aggregated to match the resolution of the health data. Time-stratified, case-crossover models were fit using conditional logistic regression to estimate associations between fire smoke PM2.5 and non-smoke PM2.5 for overall and age-stratified outcomes using 2-day averaging windows for cardiovascular disease and 3-day windows for respiratory disease.ResultsPer 1 μg/m3 increase in fire smoke PM2.5, statistically significant associations were observed for asthma (OR = 1.081 (1.058, 1.105)) and combined respiratory disease (OR = 1.021 (1.012, 1.031)). No significant relationships were evident for cardiovascular diseases and smoke PM2.5. Associations with non-smoke PM2.5 were null for all outcomes. Positive age-specific associations related to smoke PM2.5 were observed for asthma and combined respiratory disease in children, and for asthma, bronchitis, COPD, and combined respiratory disease in adults. No significant associations were found in older adults.DiscussionThis is the first multi-year, high-resolution epidemiologic study to incorporate statistical and chemical transport modeling methods to estimate PM2.5 exposure due to wildfires. Our results allow for a more precise assessment of the population health impact of wildfire-related PM2.5 exposure in a changing climate.

Project description:IntroductionDespite the London Underground (LU) handling on average 2.8 million passenger journeys per day, the characteristics and potential health effects of the elevated concentrations of metal-rich PM2.5 found in this subway system are not well understood.MethodsSpatial monitoring campaigns were carried out to characterise the health-relevant chemical and physical properties of PM2.5 across the LU network, including diurnal and day-to-day variability and spatial distribution (above ground, depth below ground and subway line). Population-weighted station PM2.5 rankings were produced to understand the relative importance of concentrations at different stations and on different lines.ResultsThe PM2.5 mass in the LU (mean 88??g?m-3, median 28??g?m-3) was greater than at ambient background locations (mean 19??g?m-3, median 14??g?m-3) and roadside environments in central London (mean 22??g?m-3, median 14??g?m-3). Concentrations varied between lines and locations, with the deepest and shallowest submerged lines being the District (median 4??g?m-3) and Victoria (median 361??g?m-3 but up to 885??g?m-3). Broadly in agreement with other subway systems around the world, sampled LU PM2.5 comprised 47% iron oxide, 7% elemental carbon, 11% organic carbon, and 14% metallic and mineral oxides. Although a relationship between line depth and air quality inside the tube trains was evident, there were clear influences relating to the distance from cleaner outside air and the exchange with cabin air when the doors open. The passenger population-weighted exposure analysis demonstrated a method to identify stations that should be prioritised for remediation to improve air quality.ConclusionPM2.5 concentrations in the LU are many times higher than in other London transport Environments. Failure to include this environment in epidemiological studies of the relationship between PM2.5 and health in London is therefore likely to lead to a large exposure misclassification error. Given the significant contribution of underground PM2.5 to daily exposure, and the differences in composition compared to urban PM2.5, there is a clear need for well-designed studies to better understand the health effects of underground exposure.

Project description:Health advances are contingent on continuous development of new methods and approaches to foster data-driven discovery in the biomedical and clinical sciences. Open-science and team-based scientific discovery offer hope for tackling some of the difficult challenges associated with managing, modeling, and interpreting of large, complex, and multisource data. Translating raw observations into useful information and actionable knowledge depends on effective domain-independent reproducibility, area-specific replicability, data curation, analysis protocols, organization, management and sharing of health-related digital objects. This study expands the functionality and utility of an ensemble semi-supervised machine learning technique called Compressive Big Data Analytics (CBDA). Applied to high-dimensional data, CBDA (1) identifies salient features and key biomarkers enabling reliable and reproducible forecasting of binary, multinomial and continuous outcomes (i.e., feature mining); and (2) suggests the most accurate algorithms/models for predictive analytics of the observed data (i.e., model mining). The method relies on iterative subsampling, combines function optimization and statistical inference, and generates ensemble predictions for observed univariate outcomes. The novelty of this study is highlighted by a new and expanded set of CBDA features including (1) efficiently handling extremely large datasets (>100,000 cases and >1,000 features); (2) generalizing the internal and external validation steps; (3) expanding the set of base-learners for joint ensemble prediction; (4) introducing an automated selection of CBDA specifications; and (5) providing mechanisms to assess CBDA convergence, evaluate the prediction accuracy, and measure result consistency. To ground the mathematical model and the corresponding computational algorithm, CBDA 2.0 validation utilizes synthetic datasets as well as a population-wide census-like study. Specifically, an empirical validation of the CBDA technique is based on a translational health research using a large-scale clinical study (UK Biobank), which includes imaging, cognitive, and clinical assessment data. The UK Biobank archive presents several difficult challenges related to the aggregation, harmonization, modeling, and interrogation of the information. These problems are related to the complex longitudinal structure, variable heterogeneity, feature multicollinearity, incongruency, and missingness, as well as violations of classical parametric assumptions. Our results show the scalability, efficiency, and usability of CBDA to interrogate complex data into structural information leading to derived knowledge and translational action. Applying CBDA 2.0 to the UK Biobank case-study allows predicting various outcomes of interest, e.g., mood disorders and irritability, and suggests new and exciting avenues of evidence-based research in the context of identifying, tracking, and treating mental health and aging-related diseases. Following open-science principles, we share the entire end-to-end protocol, source-code, and results. This facilitates independent validation, result reproducibility, and team-based collaborative discovery.

Project description:Elevated levels of ambient air pollution has been implicated as a major risk factor for morbidities and premature mortality in India, with particularly high concentrations of particulate matter in the Indo-Gangetic plain. High resolution spatiotemporal estimates of such exposures are critical to assess health effects at an individual level. This article retrospectively assesses daily average PM2.5 exposure at 1 km × 1 km grids in Delhi, India from 2010-2016, using multiple data sources and ensemble averaging approaches. We used a multi-stage modeling exercise involving satellite data, land use variables, reanalysis based meteorological variables and population density. A calibration regression was used to model PM2.5: PM10 to counter the sparsity of ground monitoring data. The relationship between PM2.5 and its spatiotemporal predictors was modeled using six learners; generalized additive models, elastic net, support vector regressions, random forests, neural networks and extreme gradient boosting. Subsequently, these predictions were combined under a generalized additive model framework using a tensor product based spatial smoothing. Overall cross-validated prediction accuracy of the model was 80% over the study period with high spatial model accuracy and predicted annual average concentrations ranging from 87 to 138 μg/m3. Annual average root mean squared errors for the ensemble averaged predictions were in the range 39.7-62.7 μg/m3 with prediction bias ranging between 4.6-11.2 μg/m3. In addition, tree based learners such as random forests and extreme gradient boosting outperformed other algorithms. Our findings indicate important seasonal and geographical differences in particulate matter concentrations within Delhi over a significant period of time, with meteorological and land use features that discriminate most and least polluted regions. This exposure assessment can be used to estimate dose response relationships more accurately over a wide range of particulate matter concentrations.

Dataset Information

Ensemble-based deep learning for estimating PM_2.5 over California with multisource big data including wildfire smoke.

Introduction

Methods

Results

Conclusion

Similar Datasets

OmicsDI is part of the ELIXIR infrastructure

Tweets

Dataset Information

Ensemble-based deep learning for estimating PM2.5 over California with multisource big data including wildfire smoke.

Introduction

Methods

Results

Conclusion

Similar Datasets

OmicsDI is part of the ELIXIR infrastructure

Tweets

Ensemble-based deep learning for estimating PM_2.5 over California with multisource big data including wildfire smoke.