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. 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 (1km × 1km) 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. 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. 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.

High-resolution spatiotemporal wind speed mapping is useful for atmospheric environmental monitoring, air quality evaluation and wind power siting. Although modern reanalysis techniques can obtain reliable interpolated surfaces of meteorology at a high temporal resolution, their spatial resolutions are coarse. Local variability of wind speed is difficult to capture due to its volatility. Here, a two-stage approach was developed for robust spatiotemporal estimations of wind speed at a high resolution. The proposed approach consists of geographically weighted ensemble machine learning (Stage 1) and downscaling based on meteorological reanalysis data (Stage 2). The geographically weighted machine learning method is based on three base learners, which are an autoencoder-based deep residual network, XGBoost and random forest, and it incorporates spatial autocorrelation and heterogeneity to boost the ensemble predictions. With reanalysis data, downscaling was introduced in Stage 2 to reduce bias and spatial abrupt (non-natural) variation in the predictions inferred from Stage 1. The autoencoder-based residual network was used in Stage 2 to adjust the difference between the averages of the fine-resolution predicted values and the coarse-resolution reanalysis data to ensure consistency. Using mainland China as a case study, the geographically weighted regression (GWR) ensemble predictions were shown to perform better than individual learners’ predictions (with an approximately 12–16% improvement in R² and a decrease of 0.14–0.19 m/s in root mean square error). Downscaling further improved the predictions by reducing inconsistency and obtaining better spatial variation (smoothing). The proposed approach can also be applied for the high-resolution spatiotemporal estimation of other meteorological parameters or surface variables involving remote sensing images (i.e. reliable coarsely resolved data), ground monitoring data and other relevant factors.

Accurate estimation of fine particulate matter with diameter ≤2.5 μm (PM_2.5) at a high spatiotemporal resolution is crucial for the evaluation of its health effects. Previous studies face multiple challenges including limited ground measurements and availability of spatiotemporal covariates. Although the multiangle implementation of atmospheric correction (MAIAC) retrieves satellite aerosol optical depth (AOD) at a high spatiotemporal resolution, massive non-random missingness considerably limits its application in PM_2.5 estimation. Here, a deep learning approach, i.e., bootstrap aggregating (bagging) of autoencoder-based residual deep networks, was developed to make robust imputation of MAIAC AOD and further estimate PM_2.5 at a high spatial (1 km) and temporal (daily) resolution. The base model consisted of autoencoder-based residual networks where residual connections were introduced to improve learning performance. Bagging of residual networks was used to generate ensemble predictions for better accuracy and uncertainty estimates. As a case study, the proposed approach was applied to impute daily satellite AOD and subsequently estimate daily PM_2.5 in the Jing-Jin-Ji metropolitan region of China in 2015. The presented approach achieved competitive performance in AOD imputation (mean test R²: 0.96; mean test RMSE: 0.06) and PM_2.5 estimation (test R²: 0.90; test RMSE: 22.3 μg/m³). In the additional independent tests using ground AERONET AOD and PM_2.5 measurements at the monitoring station of the U.S. Embassy in Beijing, this approach achieved high R² (0.82–0.97). Compared with the state-of-the-art machine learning method, XGBoost, the proposed approach generated more reasonable spatial variation for predicted PM_2.5 surfaces. Publically available covariates used included meteorology, MERRA2 PBLH and AOD, coordinates, and elevation. Other covariates such as cloud fractions or land-use were not used due to unavailability. The results of validation and independent testing demonstrate the usefulness of the proposed approach in exposure assessment of PM_2.5 using satellite AOD having massive missing values.

Spatial sampling is widely used in environmental and social research. In this paper we consider the situation where instead of a single global estimate of the mean of an attribute for an area, estimates are required for each of many geographically defined reporting units (such as counties or grid cells) because their means cannot be assumed to be the same as the global figure. Not only may survey costs greatly increase if sample size has to be a function of the number of reporting units, estimator sampling error tends to be large if the population attribute of each reporting unit can be estimated by using only those samples actually lying inside the unit itself. This study proposes a computationally simple approach to multi-unit reporting by using analysis of variance and incorporating ‘twice-stratified’ statistics. We assume that, although the area is heterogeneous (the mean varies across the area), it can be zoned (or stratified) into homogeneous subareas (the mean is constant within each subarea) and, in addition, that it is possible to acquire prior knowledge about this partition. This zoning of the study area is independent of the reporting units. The zone estimates are transferred to the reporting units. We call the methodology sandwich estimation and we report two contrasting empirical studies to demonstrate the application of the methodology and to compare its performance against some other existing methods for tackling this problem. Our study shows that sandwich estimation performs well against two other frequently used, probabilistic, model-based approaches to multi-unit reporting on stratified heterogeneous surfaces whilst having the advantage of computational simplicity. We suggest those situations where sandwich estimation might be expected to do well.

Existing methods for fine-scale air quality assessment have significant gaps in their reliability. Purely data-driven methods lack any physically-based mechanisms to simulate the interactive process of air pollution, potentially leading to physically inconsistent or implausible results. Here, we report a hybrid multilevel graph neural network that encodes fluid physics to capture spatial and temporal dynamic characteristics of air pollutants. On a multi-air pollutant test in China, our method consistently improved extrapolation accuracy by an average of 11–22% compared to several baseline machine learning methods, and generated physically consistent spatiotemporal trends of air pollutants at fine spatial and temporal scales.