Spatiotemporal Prediction of Fine Particulate Matter During the 2008 Northern California Wildfires Using Machine Learning

Reid, Colleen E.; Jerrett, Michael; Petersen, Maya; Pfister, Gabriele; Morefield, Philip E.; Tager, Ira B.; Raffuse, S. M.; Balmes, John R.

doi:10.1021/es505846r

Cited by 219 publications

(134 citation statements)

References 73 publications

(118 reference statements)

Supporting

Mentioning

131

Contrasting

Order By: Relevance

“…Machine learning algorithms are attractive methods for analyzing large data sets due to their computational speed and easy implementation for massive data, partly driven by the recent availability of highly optimized computing software. In this review paper, we have chosen Random Forest, Support Vector Regression and Neural Network for comparison, because these methods have already been used for exposure modeling (Hu et al, 2017;Liu et al, 2017;Reid et al, 2015) and software within R is readily available.…”

Section: Discussionmentioning

confidence: 99%

A comparison of statistical and machine learning methods for creating national daily maps of ambient PM2.5 concentration

Berrocal

Guan

Muyskens

et al. 2020

Atmospheric Environment

View full text Add to dashboard Cite

A typical problem in air pollution epidemiology is exposure assessment for individuals for which health data are available. Due to the sparsity of monitoring sites and the limited temporal frequency with which measurements of air pollutants concentrations are collected (for most pollutants, once every 3 or 6 days), epidemiologists have been moving away from characterizing ambient air pollution exposure solely using measurements. In the last few years, substantial research efforts have been placed in developing statistical methods or machine learning techniques to generate estimates of air pollution at finer spatial and temporal scales (daily, usually) with complete coverage. Some of these methods include: geostatistical techniques, such as kriging; spatial statistical models that use the information contained in air quality model outputs (statistical downscaling models); linear regression modeling approaches that leverage the information in GIS covariates (land use regression); or machine learning methods that mine the information contained in relevant variables (neural network and deep learning approaches). Although some of these exposure modeling approaches have been used in several air pollution epidemiological studies, it is not clear how much the predicted exposures generated by these methods differ, and which method generates more reliable estimates. In this paper, we aim to address this gap by evaluating a variety of exposure modeling approaches, comparing their predictive performance and computational difficulty. Using PM 2.5 in year 2011 over the continental U.S. as case study, we examine the methods' performances across seasons, rural vs urban settings, and levels of PM 2.5 concentrations (low, medium, high).

show abstract

Section: Discussionmentioning

confidence: 99%

A comparison of statistical and machine learning methods for creating national daily maps of ambient PM2.5 concentration

Berrocal

Guan

Muyskens

et al. 2020

Atmospheric Environment

View full text Add to dashboard Cite

show abstract

“…Wildland fire smoke can affect air quality locally and regionally, but it can be difficult to quantify for purposes of studying health impacts [66]. Three main methods have been used in health studies to characterize exposure to wildfire emissions: 1) atmospheric chemical transport modeling, 2) air quality monitoring, and 3) satellite measures of pollutant concentration or density in the atmosphere, data often combined with in-situ monitoring or other models.…”

Section: Spatiotemporal Smoke Exposure Approachesmentioning

confidence: 99%

Mapping Modeled Exposure of Wildland Fire Smoke for Human Health Studies in California

et al. 2019

View full text Add to dashboard Cite

Wildland fire smoke exposure affects a broad proportion of the U.S. population and is increasing due to climate change, settlement patterns and fire seclusion. Significant public health questions surrounding its effects remain, including the impact on cardiovascular disease and maternal health. Using atmospheric chemical transport modeling, we examined general air quality with and without wildland fire smoke PM 2.5 . The 24-h average concentration of PM 2.5 from all sources in 12-km gridded output from all sources in California (2007-2013) was 4.91 µg/m 3 . The average concentration of fire-PM 2.5 in California by year was 1.22 µg/m 3 (~25% of total PM 2.5 ). The fire-PM 2.5 daily mean was estimated at 4.40 µg/m 3 in a high fire year (2008). Based on the model-derived fire-PM 2.5 data, 97.4% of California's population lived in a county that experienced at least one episode of high smoke exposure ("smokewave") from 2007-2013. Photochemical model predictions of wildfire impacts on daily average PM 2.5 carbon (organic and elemental) compared to rural monitors in California compared well for most years but tended to over-estimate wildfire impacts for 2008 (2.0 µg/m 3 bias) and 2013 (1.6 µg/m 3 bias) while underestimating for 2009 (−2.1 µg/m 3 bias). The modeling system isolated wildfire and PM 2.5 from other sources at monitored and unmonitored locations, which is important for understanding population exposure in health studies. Further work is needed to refine model predictions of wildland fire impacts on air quality in order to increase confidence in the model for future assessments. Atmospheric modeling can be a useful tool to assess broad geographic scale exposure for epidemiologic studies and to examine scenario-based health impacts.

show abstract

“…On the other hand, as other studies have depicted before, RF has become one of the most important machine learning methods based on ensemble learning [2,8,[32][33][34]. It is developed as the extension of decision trees [35].…”

Section: Methodsmentioning

confidence: 99%

“…Therefore, a detailed exploration of the predictors, such as the relative importance and correlations with fire ignition, should be included in the modeling process. Moreover, many studies used integrated approaches such as geographic information systems (GIS), remote sensing (RS), and geostatistical methods for mapping fire occurrence [7][8][9][10]. Furthermore, machine learning (ML) and other regression techniques such as ordinary least squares (OLS), geographically and temporally weighted regression (GTWR), and geographically weighted regression (GWR) [6,9,11,12] have been employed widely in environmental and ecological fields because of their advantages.…”

Section: Introductionmentioning

confidence: 99%

A Comparison between Spatial Econometric Models and Random Forest for Modeling Fire Occurrence

et al. 2017

View full text Add to dashboard Cite

Abstract:Fire occurrence, which is examined in terms of fire density (number of fire/km 2 ) in this paper, has a close correlation with multiple spatiotemporal factors that include environmental, physical, and other socioeconomic predictors. Spatial autocorrelation exists widely and should be considered seriously for modeling the occurrence of fire in urban areas. Therefore, spatial econometric models (SE) were employed for modeling fire occurrence accordingly. Moreover, Random Forest (RF), which can manage the nonlinear correlation between predictors and shows steady predictive ability, was adopted. The performance of RF and SE models is discussed. Based on historical fire records of Hefei City as a case study in China, the results indicate that SE models have better predictive ability and among which the spatial autocorrelation model (SAC) is the best. Road density influences fire occurrence the most for SAC, while network distance to fire stations is the most important predictor for RF; they are selected in both models. Semivariograms are employed to explore their abilities to explain the spatial structure of fire occurrence, and the result shows that SAC works much better than RF. We give a further explanation for the generation of residuals between fire density and the common predictors in both models. Therefore, decision makers can make use of our conclusions to manage fire safety at the city scale.

show abstract

Spatiotemporal Prediction of Fine Particulate Matter During the 2008 Northern California Wildfires Using Machine Learning

Cited by 219 publications

References 73 publications

A comparison of statistical and machine learning methods for creating national daily maps of ambient PM2.5 concentration

A comparison of statistical and machine learning methods for creating national daily maps of ambient PM2.5 concentration

Mapping Modeled Exposure of Wildland Fire Smoke for Human Health Studies in California

A Comparison between Spatial Econometric Models and Random Forest for Modeling Fire Occurrence

Contact Info

Product

Resources

About